Lenny's Newsletter

This Week #5: Overcoming impostor syndrome, introducing growth to an org, and how to partner with your data scientist

Impostor syndrome in PM roles is usually a sign of success — you only feel it when stakes are real. Executive coach Kate Hosie frames it on a self-compassion-to-growth axis: first accept it's normal, then diagnose the cause. Burnout is often the driver — over-caring erodes perspective, lack of peer appreciation hurts as much as managerial neglect, and operating below your strengths kills confidence. Her Four Factor Model for complex roles: maintain perspective-taking capacity (tower view, not coal face), build a mindfulness habit, keep shared purpose visible, and stay in open dialogue.

To introduce cross-functional growth into a sales-and-marketing org: bring respected growth leaders for fireside chats, secure CEO sponsorship, or run a six-month experiment with a small pod and show measurable return.

Partnering with a data scientist: share the "why," not just the query; maintain a joint priority list; ask "what will we do with this answer?" before commissioning it — many questions dissolve. Give them public credit. Be data-informed, not data-driven.

This Week #9: Breaking into growth, leading with influence, and (not) stepping on toes 🦶

Breaking into growth requires understanding how the specific company defines the role — cross-functional product team, marketing unit, or expansion team. Reforge founder Brian Balfour advises mapping the product's growth loops (not funnels), demonstrating foundational tools (data analysis, user psychology, experiment methodology) rather than tactical ideas, and building a portfolio of applied work over credentials.

When bureaucracy stalls a request, no decision-maker believes it's a priority. Identify who actually decides, ask "what would it take to convince you?", set a deadline, and if small fixes routinely take weeks, surface a collected set of examples to leadership as a mirror.

On ambiguous ownership, communicate intent and a response deadline, then proceed without waiting. Airbnb's deeper fix: organize teams around outcomes rather than product surfaces, eliminating ownership disputes at the source.

This Week #10: Keeping designers and engineers excited about metrics + Transitioning from DS to PM 🕺

Designers and engineers disengage from metrics when handed goals rather than co-creating them. The fix is four shared rituals: define quarterly goals together, tracing company objectives to team levers; review dashboards regularly so the team understands what they're accountable for; start prioritization sessions with goal-tracking and require everyone—PMs included—to argue impact; and reflect on releases against pre-defined success criteria. On DS-to-PM transitions: internal transfer is the easiest path, a one-level demotion is normal, and a DS background gives strong execution skills but gaps in influence and product taste—identify and fill them.

Leading your company through a pandemic - Issue 20

Pandemic response splits on which bucket a company falls into: growing, collapsing, or holding steady. Kevin Barry's B2B SaaS data: 20% in growth markets should land-grab while CAC is cheap; 20% in hard-hit sectors should pause acquisition and protect retention; 60% in the middle should pivot messaging toward remote work. Dan Hockenmaier's argument: before you know your scenario, preserve optionality — model six months of zero revenue, extend runway, avoid knee-jerk pricing changes. Facebook CPMs fell 45% by late March 2020, but on-site conversion dropped 53%, pushing net CAC up 28%; shift spend toward prospecting and fill the free-user "lake" now (Loom halved its paid price to build habit). Assign a single DRI with authority to scrap OKRs and move daily. On customers: offer free months over discounts, give 30-day contract outs, personally engage lapsing subscribers. Recovery shape turns on two questions: did customers churn or go dormant, and has underlying behavior permanently changed? Travel faces an L-curve; food delivery and online education face permanent step-ups.

Strategy and tactics for increasing conversion

Conversion wins come from three levers — maintaining focus, maintaining motivation, and reducing friction — applied inside the funnel, plus a re-engagement strategy for users who drop off. At Airbnb, opening listings in a new tab was one of the biggest wins (keeping options visible); Instant Book drove the largest friction reduction; scarcity and highlights sustained motivation. For re-engagement, reminders, price changes, and availability updates outperformed novelty; email and retargeting were the primary channels. Experiment volume beats betting on a few big ideas.

Prioritizing conversion opportunities

Motivation beats focus beats friction for conversion upside — but friction absorbs the most effort because motivation has few levers (mostly copy) and focus even fewer (removing links). Jeff Chang (Pinterest) maps this to intent/ease: high-intent visitors (direct URL) need ease work; low-intent arrivals (generic search) need motivation. Isaac Silverman (ex-Uber Rider Growth) scores candidates by impact × success probability ÷ engineering days. Lex Roman's spreadsheet "Impact Calculators" quantify each funnel step. Ideate across all three levers, then double down where ROI is highest.

How today's fastest growing B2B businesses found their first ten customers

Every B2B company studied found its first ten customers through exactly three levers — personal network, going where customers already congregate, and press — and almost all used the first two in tandem.

Personal network dominates: Slack begged friends at other companies; Salesforce had every employee contact anyone they knew, landing their second customer in a grocery store line; Workday's earliest buyers "bought the friendship, not the software." How far your network reaches is the key variable. A strong investor base or YC cohort extends it — Carta's early customers came from angel investors and sister-portfolio founders; Gusto, Stripe, and Amplitude all got first customers from batch peers.

Going where customers are is the complement: Shopify and New Relic both emerged from the Ruby on Rails community; Segment and Airtable launched on Hacker News; Square walked door-to-door to local merchants, who became genuine daily users where network contacts had not.

Press works but rarely starts the engine — Twilio via TechCrunch, Canva via investor coverage generating a 50k waitlist. The underlying reason personal network matters most in B2B: early customers need a reason to trust you before buying.

Winning at SEO

Beating an SEO incumbent comes down to one move: generate many high-quality pages programmatically using data your business already owns. Blog posts and keyword stuffing don't scale; templated pages do. Brian Ta (Airbnb, AngelList, Strava) grew SpotAngels and Upsolve from under 1,000 to over 250,000 organic visitors monthly this way.

Three steps: identify unique data you hold — Airbnb has listings, Strava has routes — and map it to long-tail formulas like "{flower type} for sale in {city}" rather than head terms. Build one template that fills the data in like Mad Libs; block thin pages with noindex to protect quality. Then get fundamentals right: unique title tags, aggressive internal linking (the most overlooked lever), server-side HTML for crawlers, and a tight robots.txt to guard crawl budget.

Traffic that doesn't convert is vanity. Start niche, then move up the funnel as authority builds.

How today's fastest-growing B2B startups turned their early users into paying customers

Early B2B monetization collapses into three strategies: bottom-up self-service, bottom-up with inside sales, and founder-led outbound. Across 25 companies studied, every one eventually built a sales team regardless of which path they started on.

Seven companies — Segment, Figma, Airtable among them — launched with no paid plan, using free adoption to learn what customers would actually pay before introducing pricing. Figma didn't charge until 2017, a year after general availability; customers asked to pay so the product would survive. Segment ran Van Westendorp price-sensitivity surveys before setting rates. Airtable found users entering credit cards into a hidden billing flow before the team had built a real payment system.

Sales-led companies ran on founder hustle: Okta's Todd McKinnon and Frederic Kerrest personally sold every early customer. Intercom's Des Traynor emailed prospects by hand. Salesforce had every employee work the network — a product manager signed a customer while waiting in a grocery-store line.

Strategy pivots are common. Zoom started outbound and became bottom-up; Box did the reverse. Shopify's original 3.75%-of-sales model nearly killed the company; switching to flat monthly SaaS pricing in 2007–08 saved it. Carta converted 2,000+ customers from per-certificate fees to subscription.

Pricing has four levers: flat monthly, per-seat, usage-based, or transaction fee. Most companies land on per-seat, sometimes layered with a flat base.

Flywheels, flywheels, flywheels

Flywheels are self-reinforcing business loops where improving any element accelerates the whole. Amazon's canonical loop—lower prices → more customer visits → more third-party sellers → lower fulfillment costs → lower prices still—comes from Brad Stone's *The Everything Store*. Uber's version: more drivers → more coverage → faster pickups → more riders. Build one by listing assets, user actions, needs, outputs, and optimizations, then finding items that directly drive each other. Jim Collins (*Turning the Flywheel*) caps components at 4–6 identified from past successes and failures; beyond six, as both Netflix diagrams and team experience demonstrate, it stops being useful.

What it feels like when you've found product-market fit

PMF always arrives with a recognizable signal, but the shape depends on whether the product fits a broad or narrow initial market. First-hand accounts from 25 iconic founders show roughly half hit PMF immediately; the other half needed months or years — Netflix 18 months, Airbnb 2 years, Superhuman 3 years, Amplitude 4 years.

Signals split into three types. Sudden unmissable pull: Dropbox's demo video going viral on Digg; Uber growing "like a weed" with zero marketing budget; Robinhood hitting 600 concurrent visitors on an unannounced site and 10,000 sign-ups day one; GitHub's private-beta users asking "Can we pay for this?" Steady compounding pull: Segment flipping from pushing ideas on reluctant users to customers pulling toward the next problem; Nextdoor users calling in a panic within ten minutes of an outage; Gusto seeing NPS above 80 and low churn before founders even noticed. Milestones that prove the concept: Airbnb's co-founder knowing when his mom booked a stay; Canva spotting Guy Kawasaki publishing designs made with their tool.

Pull intensity equals fit quality times initial market breadth. A 10x product in a large market (Dropbox, Tinder) produces an immediate flood; the same quality in a narrow beachhead (Instacart, Superhuman) produces what Mullen called "a calm breeze" that builds into "a powerful wind." Both are real PMF — either form eventually becomes unmistakable.

Top 5 most interesting things about Booking.com's early growth strategy – Issue 46

Booking.com built an ~$80B business almost entirely on Google AdWords managed by two people — a competitive banker and a data-engineer coder — who ran $100M+/year in spend mostly by hand. Performance marketing drove supply decisions: if they were losing a keyword, the supply team investigated why. Shut out of chain hotels by Expedia, they dominated secondary destinations with geo-targeted landing pages and machine-translated ad copy that beat human copywriters in A/B tests.

The Transition: Layering sales onto a bottom-up self-serve product

Waiting too long to add direct sales to a self-serve product carries a real opportunity cost — Dropbox lost enterprise segments to competitors who adopted sales-assisted motions while staying self-serve-only. Pete Kazanjy (author of *Founding Sales*, founder of Atrium) argues that for bottom-up B2B the question is not whether to add sales, but when and how.

Self-serve works when the product is simple enough for independent "aha" moments (Zoom, Stripe for new developers), when the offering is genuinely new (Calendly), when it can coexist with an incumbent (Slack alongside GChat), or when targeting small orgs not yet locked into a legacy provider. Complex all-or-nothing replacements — HRIS, enterprise email platforms — don't self-serve well.

Two reasons justify adding salespeople: consolidating scattered usage inside a large account (the Slack/Zoom "scoop up the pods" motion), and improving conversion of high-value users who abandoned before activating. Economics must pencil out at roughly 4× the rep's fully loaded cost in incremental revenue. At 40 opportunities/month, 25% win rate, $5k average contract, and $10k monthly rep cost, the math works.

Timing signals: inbound "Contact Sales" requests arriving once or twice a week, or enterprise logos appearing in signup logs without activating. Add title and company fields — Snowflake does; New Relic notably didn't. Enrich via Clearbit; pipe activation data (feature breadth, page views, login recency) into a CRM and set alerting thresholds so salespeople know when a high-value account has stalled.

Founders should run the first few dozen sales conversations themselves — product fluency and founder credibility outperform any early hire. Four pitfalls: delegating sales too early, avoiding it out of fear, pursuing top-down enterprise motions before the product has organizational penetration, and failing to allocate engineering for SSO and SOC 2 compliance that enterprise contracts require.

When NOT to run an experiment – Issue 54

Default to running experiments — but skip them in three cases. First, when sample sizes make the wait prohibitive: detecting a 5% lift on a 10% conversion step requires 60,000+ users per variant. Second, when downside risk is low and experimentation setup costs are high — the real question is whether skipping an experiment frees time to build a better experimentation framework. Third, when launching something genuinely new with no control to compare against; set independent success criteria instead.

Generating buzz

Nobody cares about your new product — attention must be earned by doing something worth remarking about. Ten strategies with 49 examples: remarkable video (Dollar Shave Club), demo (Dropbox's MVP), or value prop (Gmail's 1GB, Robinhood's $0 commissions); offline stunt (WePay dumping 600 lbs of ice at a PayPal conference); controversy; giveaway; viral mechanic (Harry's captured 100k emails pre-launch via tiered referral prizes); influencer seeding with artificial scarcity (Superhuman's 275k waitlist); pre-launch tease; and being visibly everywhere at once.

How to run SEO experiments

SEO experiments require page-level bucketing, not user-level — Google is the only "user" you're designing for and must see a consistent experience on every crawl. No commercial platform supports this, so you must build your own framework by hashing canonical URLs to assign treatment/control per page.

The only valid metric is organic traffic. Keyword rankings fluctuate daily, Google Search Console data is heavily sampled, and conversions are a counter-metric only. Run for 2–4 weeks so Google discovers changes, then use difference-in-difference analysis to control for unequal traffic between buckets. Only on-page changes are testable; internal-linking effects require separate manual analysis.

Prerequisites are steep: 5,000+ organic visitors per day and thousands of programmatic pages. For 95% of startups still in the build phase, the framework isn't worth constructing.

Results from those who built it: title tag experiments drove 15–20% traffic lift at Airbnb; adding routes to Strava's "Where to run in {city}" pages drove 20%; meta description optimization added 6% at Airbnb via improved CTR.

Positioning

Positioning fires assumptions in buyers' minds before you say another word. Call your product a "CRM" and buyers assume Salesforce is the benchmark, head of sales is the buyer, price is lower. Good positioning fires true assumptions; bad positioning fires false ones your team must undo.

April Dunford's five-component sequence — competitive alternatives → differentiated attributes → value created → customers who care most → market category making the value obvious — must run in that order. Value only exists relative to alternatives; category frames it for the right buyer.

Two repositioning wins: a "Microsoft Access killer" with six real users became an "embeddable database for mobile devices" and was acquired for hundreds of millions. A startup fighting Siebel shifted to "CRM for investment banks" after finding its relationship-modeling feature decisive for relationship-dependent firms; revenue grew from under $2M to $80M before Siebel acquired it for $1.3B.

Three traps: treating status quo — spreadsheets, doing nothing — as a competitor rather than inertia; listing phantom competitors you never lose deals to; over-valuing category creation. Salesforce and Gainsight built niches in existing categories before redrawing boundaries. Category creators routinely lose to later entrants — 90% of tech IPOs in the prior five years were in existing categories.

Content-driven growth

Content-driven growth maps onto a 2x2: SEO vs. virality on one axis, user-generated vs. employee-generated on the other. That yields five strategies — user-generated SEO (Quora, Reddit), data-generated SEO (Zapier, Thumbtack), user-generated viral (TikTok), editorially viral (Spotify Wrapped), and editorially generated SEO-optimized (EGSO). The focus here is EGSO, where employees create content to rank for keywords.

HubSpot started publishing before they had a product; co-founder Dharmesh Shah's OnStartups blog proved that valuable content builds audience faster than interruption advertising. Early on, 30%+ of HubSpot customers came from the blog, and their organic traffic now rivals TechCrunch. Ahrefs grew their blog gradually from 2015 to 1.5M visits/month across 300+ posts. Slidebean maxed out paid search for "pitch deck," moved to SEO articles (~150 in one year), then pivoted to YouTube — 18 months of bad videos before finding a format that worked, now $125K/year in production. Intercom's founders wrote the first 100 posts themselves; the content team later split into Audiences, Enablement, and Channels pillars.

Cross-company patterns: start cheap and iterate until something works; separate SEO content (keyword-research-driven, predictable) from viral content (creativity-driven, unpredictable); model the traffic TAM before committing to a topic area; expect years before payoff. Webflow adds a sixth use case — video to improve activation, not just acquisition: users who complete Webflow University courses convert to paid at significantly higher rates.

How people discover new products

Product discovery falls into seven channels: friends/colleagues (cheapest, but only works when sharing is natural—Slack, Snapchat); organic online (social chatter, press, SEO—drove Clubhouse, e.l.f. Cosmetics); paid online ads (Casper, Calm—best when customers aren't actively searching); organic in-person (shelf placement, seeing a Tesla); out-of-home ads (billboards—Brex targeted SF founders this way); in-home promotions (TV, podcast ads, direct mail—suits DTC and high-AOV products); and outbound sales (Oracle, Workday—required for enterprise). To choose, map your early adopters, where they spend time, and what's cheapest to reach them first.

Finding your distribution advantage

Poor distribution — not product — is the primary cause of startup failure. Rising CPMs, crowded SEO, and privacy changes make distribution advantages decisive. At scale, winners dominate one channel: Wish (performance marketing), Pinduoduo (virality), Wayfair (SEO). Early startups have seven routes: founder audience (Tesla, Kylie Cosmetics); viral loop (Faire, Dropbox); first on emerging platform (Zynga/Facebook); remarkable story earning free PR (Airbnb); pre-existing buyer relationships (Carta); strategic partnerships (Netflix/DVD, PayPal/eBay); and hustle — Tony Xu personally delivered DoorDash food. Stacking two or three multiplies effect.

GTM motions of 30 B2B SaaS companies

Every product-led B2B SaaS company eventually adds a sales team — 19 of 30 studied started product-led, and 100% layered in sales; four switched to fully sales-led. The reverse never happened. All 30 moved upmarket over time: Amplitude (SMB→Enterprise), Slack (SMB→Mid-market+Enterprise), Gusto (VSB→SMB). Starting with VSBs or SMBs is the default because they move faster and require less compliance infrastructure; only Workday, Snowflake, and Databricks opened with Enterprise, justified because their problems (data scaling, workforce management) are Enterprise-specific. Targeting stays narrow — one to three personas maximum, e.g. Amplitude targeting mobile PMs, Looker targeting VP of Eng. Sales-led incumbents like HubSpot and Salesforce later added self-serve products, but primarily as lead generation, not revenue.

How behavioral science can boost your conversion rates

Small environmental changes, not feature overhauls, drive large conversion lifts. Irrational Labs co-founder Kristen Berman documents four cases: at Steady (gig-worker income app), replacing a skippable prompt with a forced accept/decline choice lifted bank-account linkage from 7.1% to 11.6%; a progress-completion frame pushed it to 15.9%. At Livongo (diabetes management), rewording an email from "Join the program" to "Claim your welcome kit" produced a 120% registration lift via the endowment effect. At EarnUp, prompting borrowers to round mortgage payments to a natural number boosted overpayment opt-ins 40%. At a Latino Community Credit Union pilot, embedding "round up to savings" as a day-one opt-out on loan forms hit 36% opt-in versus 14% via later email — day-one momentum is the highest-leverage moment. The underlying process: literature review first, one specific target behavior, a granular behavioral map, barriers/benefits at each step via the 3Bs framework, then single-variable experiments.

Picking a wedge

Win a large market by first capturing a narrow slice — a specific problem within it, or a large share of a small adjacent one. Not always necessary: Zoom, Slack, and Datadog attacked broad markets head-on and won. Most valuable against entrenched or crowded markets, where focus accelerates iteration, builds social proof faster, and reduces capital needs.

A good wedge solves one problem for one group extremely well, builds momentum, and extends naturally upward. Pick the narrowest, most painful problem; then pick the segment with the most acute pain or fastest sales cycle — a product 10x better on average may be 50x better for one group. Carta wedged on cap table management; PayPal on eBay auction payments (10K to 5M users in months); Twilio on Ruby on Rails developers; Uber on black cars for price-insensitive SF riders.

Breaking into growth

Moving from generalist PM to growth PM takes three steps: Ask, Learn, Do. Telling your manager explicitly — and repeatedly — was the primary reason the author landed Airbnb's supply-growth role six months later. Learning means working through seven domains: A/B testing and statistical significance, growth strategy (loops vs. funnels, Reforge's racecar framework), SEO, paid acquisition (CAC, ad mechanics), virality types, conversion optimization, and SQL/stats. Doing means running actual experiments, auditing your company's SEO, and shipping a real ad before you hold the title.

Six rules of hiring for growth

Before hiring anyone for growth, founders must define a growth model — the combination of levers (acquisition, retention, monetization) and motions (product-led, sales-led, marketing-led). SurveyMonkey won via product-led prosumer loops; Qualtrics entered the same market with a sales-led university motion — nearly identical products, opposite models.

The founding team must draft that first model themselves. A new Head of Growth will paste in patterns from their last company, lacking the months needed to understand product DNA. Once the model exists, hire a Builder (growth PM, growth marketer, lifecycle manager) — not an Innovator or Optimizer. The growth function is barely a decade old, so external talent is scarce, expensive, and often unwilling to repeat the build; promote from analytics, product, or engineering instead. Verna, Casey Winters, and Bangaly Kaba all came from other departments.

Build data infrastructure before headcount: acquisition, retention, and monetization metrics predict revenue and let the first hire run immediately. Place the growth hire inside the team that owns the relevant motion; start embedded rather than stand-alone.

A founder's guide to community

Building an audience means helping people; building a community means helping people help each other. That shift is how Duolingo runs 2,600 events a month with a three-person team and how 83% of Salesforce customer questions get answered by other customers.

David Spinks (co-founder of CMX, VP Community at Bevy) organizes the investment decision around the SPACES model — six objectives: Support, Product, Acquisition, Contribution, Engagement, Success. Stage determines priority: pre-PMF companies should focus on Product (fast feedback from a beta cohort); growth-stage on Acquisition and Engagement; mature on Support and Success. Platform businesses — marketplaces, open-source, wikis — need Contribution from day one.

Members join for benefits, not belonging; belonging comes after relationships form. Before launching, interview at least 10 potential members live and confirm their motivations align with the chosen objective. Misalignment kills technically well-built communities.

Strategy runs at three levels. Business goals sit at the top, tracked through correlation — do community members buy, renew, and refer at higher rates? Community health sits in the middle: monthly active users (CMX's Slack runs 14% MAU, its Facebook group 18%), NPS, and belonging surveys (1–10 ratings on safety, voice, relationships, inclusion). Tactical programs sit at the bottom, each assigned its own metric.

For launch, start with 10–50 founding members and a single channel. They seed the cultural mold before scaling; going too big too fast is the most common failure. Grow at 2x every two weeks. For events, prove community-market fit yourself first — repeat attendance and high post-event NPS — then codify a host playbook and open volunteer applications starting with three to five hosts. Finimize runs 200+ events annually with a three-person team this way.

The top criterion for a community hire is genuine curiosity about the subject, not prior expertise in it.

What is a good payback period?

CAC payback benchmarks differ sharply by segment: B2C targets are under 1 month (great), 6 months (good), 12 months (acceptable); B2B SMB thresholds are 6/12/18 months; enterprise 12/18/24 months. Exceptional cases recoup on the first transaction or via prepayment. The common calculation error is using revenue not gross profit — a $100 CAC at $10/mo with 80% margins is 12.5 months, not 10.

Higher periods are defensible when LTV is predictable (sticky SaaS, multi-year contracts), when a mature business has cohort data, or when fueling a growth loop. Early-stage companies should stay short because LTV assumptions are unvalidated. For paid-heavy growth, track by channel not blended — organic masks paid inefficiency. Reduction levers: annual plans collect cash upfront, PLG self-serve cuts sales cost, usage-based pricing escalates naturally. Target depends on model: low-loyalty/low-frequency products need near-immediate payback; high-switching-cost products can justify 1–2 years.

Freemium vs. trial

Free is an acquisition strategy, not a pricing strategy. Two modes: business-model disruption (give away the core, monetize elsewhere — Robinhood via payment-for-order-flow, Chime via interchange, Square via transaction cuts) or lead gen via freemium/trial. Among ~50 SaaS products surveyed, 90% of freemium companies also offer a 7–30-day paid trial. Go trial-only when onboarding requires hand-holding (Okta, HubSpot); go freemium without trial when the upgrade is self-evident (Figma, Miro). Patrick Campbell's rule: don't launch freemium until you understand conversion — it's a scalpel, not a sledgehammer.

April Dunford on product positioning, segmentation, and optimizing your sales process

Weak positioning damages the entire pipeline — sluggish marketing, confused sales calls, and churn when the product isn't what customers expected. The diagnostic: if a prospect asks to "start over," calls the product "just like Salesforce," or says "I don't get why anyone would pay for that," positioning is broken.

Positioning defines how a product is best at delivering value a specific set of buyers cares about. Five pieces in sequence: (1) Competitive alternatives — what must you beat? Include status quo (spreadsheets, interns), which drives 40% of "no decision" losses. (2) Differentiated capabilities against those alternatives. (3) Map capabilities to value; two or three themes emerge that wouldn't appear if you asked "why does everyone love our stuff?" (4) Target segment: what characteristics make buyers care most about that value? (5) Market category: context that makes value obvious to those buyers. Never start with category — the first four are what you evaluate it against.

Help Scout illustrates the chain. Alternatives: Zendesk and shared email. Value: customer service as a growth driver, not a cost center. Target: direct-to-consumer e-commerce brands where service drives loyalty. Their narrative opens with that proposition, shows rivals treat service as overhead, and positions Help Scout for companies that disagree.

Early-stage companies should keep positioning loose until the market reveals who converts and why — lock it down once the pattern is clear.

In B2B, five to seven people influence a purchase, but only one persona matters: the champion who builds the short list and drives consensus. Arm them to sell IT, legal, and the economic buyer — if positioning doesn't land with the champion, no other stakeholder is reached. Misalignment between founder, sales, marketing, and product is the most common root cause; fix it in a cross-functional workshop ending with a storyboarded sales narrative.

Elena Verna on how B2B growth is changing, product-led growth, product-led sales, why you should go freemium not trial, what features to make free, and much more

Growth teams amplify strong product-market fit but cannot manufacture it — most B2B growth tactics either don't work or actively hurt.

On hiring: bring in a growth team only after founders have achieved solid PMF, good retention, and enough data to run experiments. For declining businesses, a growth team lifts results 10–15% at best; it cannot fix a product or market problem. Stop the decline first.

On wasted tactics: homepage redesigns always produce a step-back taking three to six months to recover — never promise acquisition lift at launch. Copying competitors fails 95% of the time; you're usually seeing an untested orphan page. Your problem is almost never unique — find who already solved it. Growth teams that over-index on SEO and paid search rent distribution from algorithms that can revoke access anytime. The highest-leverage investment is earned channels nobody else can compete in: virality, user-generated content, referrals. At Dropbox, sharing loops drive over 50% of acquisition. If sample size takes more than a month, do pre/post rather than A/B testing. Color tests, OAuth additions, one-off emails, and "simplify onboarding" as a line item never move metrics.

Growth models have S-curves. Dedicate 20–25% of annual team capacity to introducing a new loop before the current one plateaus; each needs roughly 18 months before contributing real revenue. Three frameworks Verna returns to: growth loops (Brian Balfour/Andrew Chen at Reforge), Lenny and Dan Hockenmaier's race-car framework (engines, fuel, turbo boosts, optimizations), and Bangaly Kaba's adjacent-user theory for expanding reach without changing the core product.

Career contrarian take: full-time employment is one monetization package for your skills, not the default. Optionality — choosing what fits your life — is a better north star than title progression. Advisors are the fastest learning shortcut; vet them with a paid workshop before committing to a retainer.

How to kickstart and scale a consumer business—Step 2: Identify your super-specific who

Early-stage consumer products fail because the audience is wrong, not the product. Pinterest went nowhere with tech employees; Ben Silbermann found traction when 30-something female bloggers discovered it at a design conference. The product hadn't changed — the audience had — and within a year Pinterest hit a million users.

The target must be "almost comically narrow." Andy Johns at Wealthfront defined their early adopter as a 25–35-year-old engineer at a pre-IPO tech company who prefers to delegate money management. Andy Rachleff frames it as a beachhead: own a niche, grow via references.

Discord started with Final Fantasy XIV gamers; Netflix recruited from DVD-enthusiast forums; Instagram hand-picked designers with large Twitter followings to set platform tone.

To find your own: name three people who'd be exceptionally excited, identify what they share, pin three specific attributes — age, role, context. Typeform's broad "data collection for everyone" stalled at year three; narrowing to marketers using forms for growth unlocked the next phase.

How to kickstart and scale a consumer business—Step 3: Craft your pitch

Your pitch — not your product — often determines early traction. People's time is already allocated; good isn't enough. Almost no one gets the hook right first try — Netflix took 18 months before 'no late fees plus subscription' immediately worked; DoorDash's Tony Xu pivoted from 'mobile technology' to 'revenue, risk-free' after weeks of failure. Four routes: find your unique differentiator vs. the incumbent (Robinhood's $0 commission, Netflix's no late fees); listen to how obsessed users describe it (Dropbox = 'throw away your USB drive'); frame the job being done (iPod's '1,000 songs in your pocket'); or lead with something bold and specific (Zillow's Zestimate, Domino's 30-minutes-or-free). If nothing lands: bad pitch, wrong audience, or product nobody wants.

How to sell your ideas and rise within your company | Casey Winters, Eventbrite

Selling ideas inside a company is less about idea quality and more about communicating at the right altitude. Casey Winters, CPO at Eventbrite and former growth lead at Pinterest and GrubHub, argues that most PMs under-communicate upward — and when they do, start at "chapter six," skipping the strategy and metric context executives need. The fix: find the last point obvious to your audience, start there, build forward.

Preparation happens before the room. Role-play each stakeholder — the CFO will ask about unit economics, the CEO about strategic fit. Pre-meetings defuse surprises. At Pinterest, Jack (head of product) opened every review with diagnostic data questions; failing those poisoned the rest.

On product complexity: progressive disclosure, segmentation, and unbundling all broke down for Eventbrite because creators span the full sophistication range and shift tiers over time. The answer is "perceived simplicity" — advanced features discoverable when sought, invisible when not. WhatsApp is the benchmark.

Non-sexy investments (performance, stability, developer velocity) get chronically underfunded because they resist clean metrics. The lever is horizontal buy-in: if your engineering manager and design lead are convinced, building a custom metric or proof-of-concept becomes a coalition effort. Product-market fit also erodes silently — user expectations and competitive bars rise continuously, and neglecting maintenance can drift a product out of fit over years.

"Kindle strategies" are non-scalable hacks whose only job is to unlock "fire strategies" — viral loops, content loops, paid acquisition. Hire a dedicated growth person only once a fire strategy is proven. The underrated lever: data network effects — proprietary usage data that improves targeting builds a first-party edge as Facebook and Apple restrict third-party signals.

The great filter for senior PM careers is strategy. Execution carries you to senior PM; writing a strategy document without prompting is what opens director and CPO levels.

How to scrappily hire for, measure, and unlock growth | Crystal Widjaja, Gojek and Kumu

The best growth data comes from testing the actual experience — even with 30 users, the directional signal matches what you'd get at scale; only precision changes. Crystal Widjaja built Gojek's growth function (170M users, more food deliveries than GrubHub + Uber Eats + DoorDash combined) and now leads product at Kumu. Her team validated a subscription feature through a WhatsApp group of 100 drivers selling in-ride, with interns issuing vouchers in the backend — no engineering. New onboarding screens were tested as designer mockup overlays sent as in-app messages.

Growth strategy starts with "physics" — what the market, product, model, and channels actually allow — then finds underused levers before changing anything else. Gojek's biggest GoPay unlock came from drivers: when matched to a customer who had never topped up, the driver got a real-time incentive to pitch GoPay during the ride, which became 60% of GoPay acquisition. A parallel advisory case: adding a "pause" button to a subscription app eliminated the top cancellation reason without any reacquisition campaign.

The core analytics failure is tracking events without context properties. Knowing a user landed on the map tells you nothing; knowing they saw two drivers, in a specific city, with surge pricing active lets you ask why they didn't book. Measurements are observations; insights answer the why. GoFood users whose friends had ordered from a merchant were twice as likely to try that merchant — solving for trust upstream of conversion.

Retention benchmarks: free products need 60% week-one retention and the curve should flatten near that level; friends-and-family cohorts should hold 80%. On hiring: prioritize statistical intuition over experience — someone who spots selection bias and designs a controlled experiment. Give a four-hour take-home case study and reward candidates who admit they had to look things up.

How to launch and grow your product | Ryan Hoover of Product Hunt and Weekend Fund

Product Hunt started as a newsletter experiment, not a startup — incorporated four to five months after launch. That framing is Hoover's core advice: call it an experiment where the goal is learning, not success, and don't raise until you can see yourself working on the idea for a decade.

On fundraising: raise only when you know what it's for. Hoover raised because he couldn't pay people otherwise. His regret is not monetizing sooner — Product Hunt reached cashflow break-even roughly 12 months after it first tried to make money, which came years too late. He also tried to expand horizontally into podcast, gaming, and book discovery, severely underestimating how hard it is to translate a community across categories.

Launches serve more purposes than customer acquisition: recruiting, fundraising momentum, partnership serendipity, SEO, and team morale (the overlooked one — a team member sharing the launch with their mom matters). The biggest mistake is writing like a PR person. Best tagline test: how do customers describe the product to friends? Gallery images that tell a sequential story outperform static screenshots.

Consumer is harder than B2B because monetization requires massive scale, you compete for attention against incumbents with built-in habits, and the user need is fuzzy. Hoover keeps a problem journal — recording annoyances, not solutions — to stay attentive to friction. Other signals: niche communities, behavioral shifts like distributed work, and technology unlocks that make something newly possible.

Momentum is reflexive: high momentum produces more high momentum. Delegation is the classic founder failure mode — Hoover was still personally editing the Product Hunt newsletter at 5 a.m. years in. Angel investing entry paths run from scout programs (easiest, least autonomy) to SPVs (deal-by-deal carry) to a full fund. For those without access, writing investment memos on companies you would have backed builds a track record before you have deals.

The rituals of great teams | Shishir Mehrotra of Coda, YouTube, Microsoft

Great teams share a small list of named, templated rituals every employee knows by their first Friday — that test, from investor Bing Gordon, underlies Shishir Mehrotra's book-in-progress on team rituals. As Dharmesh Shah frames it, culture is the product you build for employees; people describe culture by naming rituals, not values.

Coda's golden ritual is Dory/Pulse. Pulse forces everyone to write their opinion before seeing others', eliminating groupthink. Dory lets participants upvote questions before discussion so the most pressing thing gets addressed. Mehrotra later added a role column borrowed from Coinbase's RAPID — labeling each participant as approver, decider, or informed — preventing Pulse from drifting into consensus theater.

The eigenquestion concept emerged from a 2008 YouTube crisis. Instead of relitigating whether to link out to ABC.com for Modern Family, Mehrotra reframed an offsite around one prior question: will online video value consistency or comprehensiveness? The team chose consistency; every subsequent call — dropping external links, reclaiming the YouTube app from Apple — followed automatically. An eigenquestion is the one that, when answered, eliminates the most subsequent questions. Mehrotra teaches it through a teleportation prompt: given only two scientist answers, which two questions do you ask? Sharp candidates immediately surface the load-bearing variables.

For talent, Mehrotra uses PSHE (Problem, Solution, How, Execution): junior people execute handed playbooks; seniors identify problems no one assigned. Reference checks outrank interview signals — ask who on the past team identified the right problems, and the answer reveals whether the candidate set direction or merely ran efficient meetings.

On growth, Coda runs two loops: the Black Loop (share-create-share within teams; only makers pay, removing friction from sharing) and the Blue Loop (publish publicly to drive discovery). Nearly a third of new users enter through the Blue Loop, finding a useful doc before finding Coda.

Kickstarting and scaling a consumer business—Step 6: SCALE: Build your growth engine

Consumer startups have only three viable self-sustaining growth engines: virality, SEO, and paid acquisition. Sales is rarely economical at consumer scale. Most successful companies grow through just one — Tinder was nearly 100% word of mouth, Calm nearly 100% paid, Thumbtack nearly 100% SEO. Spreading effort across engines early is the central failure mode; the goal is to pick one and become world-class at it. Only after the primary engine plateaus does adding a second (usually paid) make sense.

Fit determines the choice more than preference. SEO suits products with user-generated public content (Yelp, Reddit) or proprietary data generating thousands of indexable pages (Grubhub menus, Thumbtack service listings). Thumbtack committed their entire 12-person team to SEO; results were negligible at six months, meaningful at eighteen, primary at thirty-six. Paid fits products where new users generate direct revenue with short payback. Virality fits products where sharing is baked into usage — Dropbox requires it, Pinduoduo structured group-buying so K-factor never fell below 1.

Marketplaces get two additional engines: direct supply sales (Grubhub reps walking into restaurants, OpenTable carrying hardware door-to-door, Etsy recruiting at craft fairs), and supply-driven demand, where the supply side markets the platform itself (DoorDash restaurants posting their own window stickers, Etsy sellers promoting their shops).

The Racecar framework names four supporting elements: the growth loop itself; turbo boosts — temporary accelerants like PR or influencer spikes (Cameo's Ronnie Radke produced 300% MoM growth); lubricants — conversion, activation, retention, and brand improvements that make every other part more efficient; and fuel — money for paid loops, content for SEO, users for viral ones.

All the ways to grow your product

The Racecar Growth Framework organizes growth into five components: kickstarts (seven early-user tactics), four self-sustaining engines (paid, SEO, virality, sales), lubricants (conversion, retention, activation, brand), turbo boosts (one-off spikes like PR or viral content), and two mid-stage accelerants — channel partnerships and geographic expansion. Channel partnerships (Google/Netscape, Netflix/Toshiba, Kayak/AOL) are high-risk/high-reward and typically contribute marginally. Geographic expansion is often the biggest single accelerant post-PMF — Instacart, Tinder, Facebook, and Rover all used it as the primary jump between kickstart and scaled engine, but it can only be done once.

The Racecar Growth Framework—expanded and illustrated

Startup growth has a natural sequence: use unscalable Kickstarts (cold outreach, campus visits, press) to reach the first 1,000 users, then let a self-sustaining Growth Engine take over—SEO, paid ads, sales, or virality, each recycling its output back into more growth. Lubricants (conversion, retention improvements) keep the engine efficient. Turbo Boosts (viral campaigns, influencer hits) spike growth at any stage. Once at scale, Mid-stage Accelerants like channel partnerships and geographic expansion reaccelerate. Layer a second engine before the first plateaus.

Category creation and brand building | Barbra Gago (Pando, Miro, Greenhouse, Culture Amp)

Category creation only makes sense when customers lack a shared word for what they need and no budget line exists. At Greenhouse, Gago tried renaming "ATS" to "recruiting optimization platform," got press coverage, but buyers kept calling it an ATS. They abandoned the push and invested in elevating the category — making applicant tracking genuinely strategic — which helped spark the modern people-ops movement.

Miro was the opposite. "Online whiteboard" was real but tiny; product, engineering, HR, and design teams each described the tool differently. "Visual collaboration" unified those use cases into one enterprise budget line. It emerged from customer listening, then required lobbying G2, Gartner, and Forrester to recognize it as distinct from diagramming. Competitor copycats confirmed rather than threatened it: a category only exists once multiple companies claim it.

The playbook: obsessive customer interviews to find convergent language; PR to test the framing; analyst relations to get listed on directory sites; content marketing to educate buyers there is something new worth budgeting for.

The Miro rebrand (RealtimeBoard to Miro, October–March): run it as a product sprint. Involve legal, sales, and product. Nod to the original — the yellow Post-it became the yellow M. The domain transfer cleared 48 hours before the South by Southwest launch with materials already printed.

Opinionated software enforces a best practice over flexibility. Greenhouse's structured recruiting — fixed pipeline stages, required interview kits — reduced bias by removing discretion. Gago's startup Pando applies the same logic to performance reviews: companies patch pay gaps without fixing the promotion process that generates them.

Brand is a system: visual identity, voice, photography-versus-illustration choices, and values shaping behavior at every touchpoint. At Miro, each letter embodied a value — agility mapped to a kinetic, wiggly shape — making the system extensible without retraining every new hire.

How Snyk built a product-led growth juggernaut | Ben Williams (VP of Product at Snyk)

Security tooling was historically run by security teams late in the dev cycle — slow feedback, developer frustration, poor adoption. Snyk's founders bet that developer-first security, built to keep developers in flow, could disrupt the model from below.

The first users came from a deliberately narrow focus: Node.js developers using open-source NPM packages, reached through the Velocity Conference in Amsterdam and community evangelism. Single persona, single ecosystem — narrow enough to ship fast, wide enough to matter (200,000 NPM packages, 2.5B monthly downloads, the average app 75% open-source).

The breakout acquisition loop was the GitHub fix-PR: connect a repo, Snyk scans and opens branded educational pull requests fixing vulnerabilities. Other developers in the repo see the PR, some sign up, connect their repos, loop repeats. Layered on: Snyk Advisor, programmatically-generated package health pages ranking on Google and funneling visitors toward sign-up; and free, ungated security education lessons.

Early self-serve monetization failed — individuals paid $100/month but enterprise purchases didn't happen. The fix was governance features (reporting, user management) that security teams require, plus language breadth beyond Node.js so a CISO's entire stack was covered. Sales hires followed those product changes.

The growth org is structured by outcome (acquisition, activation, monetization, platform), each team cross-functional with engineers, PM, designer, growth marketer, and decision scientist. Embedding growth marketers in product teams is uncommon but broadens idea generation and execution.

Activation is team-level: a team fixing a vulnerability within 30 days of creation, identified via ML as the strongest predictor of three-month retention — fixing, not logging in or scanning.

Growth strategy runs on a loop model — qualitative documentation of all loops plus quantitative data identifying the biggest constraint each quarter. The most important ceremony is the weekly impact-and-learnings review: PM-facilitated, focused on what experiments taught, no time on what shipped.

How Notion leveraged community to build a $10B business | Camille Ricketts (Notion, First Round Capital)

Community-led growth at Notion worked by achieving ubiquity — so many individuals talking about, teaching, and building around the product that enterprises stopped needing to vet it. Camille Ricketts, Notion's first marketing hire, frames the goal as *discovery* rather than awareness: the moment someone has heard enough to actively seek the product out. Her KPI was net new monthly visitors to Notion.com.

Three programs drove it. Ambassadors: starting in 2019 with 20 vocal Twitter users, inducting ~20/month to preserve intimacy — at 5,000 members communities go quiet because people feel they're addressing an auditorium. Influencers: sponsored creators drove measurable traffic; Lexi Barnhorn led this. Champions: a Slack for power users inside enterprise customers, supporting land-and-expand.

The ambassador flywheel became economic. By mid-2021 one creator earned $35,000 in four months from a single template; Notion supported creators with guides and peer networks. Some consultants now employ dozens of people — the Salesforce ecosystem model at prosumer scale. Notion Vietnam's Facebook group hit 250,000 members; the subreddit 210,000, both volunteer-run. Thirty in-person events happen monthly worldwide.

Notion's press break was a David Pierce Wall Street Journal piece calling it the one productivity app you'd ever need — visible in the traffic graphs. Product Hunt launches (2.0, Notion AI) each spiked acquisition.

On content, the frame is "content market fit": diagnose the audience's painkiller — not what they didn't know but what was causing anxiety or making them feel alone. First Round Review posts took ~16 hours each; quality and time correlated directly with performance. For founder social: wait until you have something genuinely useful to say rather than hitting a quota. Community is not right for every company — sales-led, long-contract products should invest elsewhere; pre-PMF enterprise teams benefit most from small customer advisory boards that evolve into a first evangelist layer.

Founder-led sales | Pete Kazanjy (Founding Sales, Atrium)

Founders who hand sales off too early lose the feedback loop that shapes both product and pitch — and almost always fire that first VP. Founder-led selling is the sequel to customer development: once the problem is validated, the founder must personally close the first 20–30 arm's-length deals before anything can be abstracted.

The signal to hire is a 15–25% win rate across 50–100 at-bats with genuine strangers. At that point the selling loop — first meeting → discovery → proposal → close — runs repeatably enough to live outside the founder's head. Until then no sales leader can help; the motion isn't documented anywhere.

When you do hire, skip the VP. Take two early-stage AEs who've sold to the same persona at a comparable startup (Greenhouse alumni for recruiting tools, Figma alumni for design tools). Supply documented materials first — discovery question bank, demo script, objection slides. Measure leading indicators from month one: first meetings booked, second-meeting conversion, progression to proposal. No second dates within a month is a signal; don't wait nine months for lagging revenue.

Two execution frameworks: sharpen ICP to three layers — company profile, end user, and budget holder (usually different people); treat the sales motion like software, updating talk tracks and slides after every call that stumped you. The underlying craft is rapid rapport, directed discovery, and staying silent after asking for money.

On PLG: even Atlassian had sellers — they just priced low enough to delay needing many. End users lack budget authority, so someone must eventually reach the purse strings. Every B2B company builds a sales org; the only question is when.

Two recommended books: Goldratt's *The Goal* for systems thinking applied to revenue pipelines, and Bill Walsh's *The Score Takes Care of Itself* for leading-indicator discipline over outcome fixation.

Virality is a myth (mostly)

Products don't grow by friends-telling-friends cascades — they grow through one-to-many broadcasts. A 2012 Yahoo study found 95% of Twitter content reaches people directly from its source or one degree out; nothing truly diffuses like a virus. Derek Thompson calls the real mechanism "broadcast diffusion": popularity is driven by the size of the largest broadcast, not the depth of a sharing chain.

Real products confirm this. Spotify broke out when Zuckerberg posted and Sean Parker seeded influencers. Twitter's first spike came from Om Malik's blog post. Instagram's launch was timed to simultaneous press coverage. Notion's turning point was a single Wall Street Journal story. Clubhouse's 1,500→10M run was sparked by Naval, Andreessen, then Elon Musk hosting sessions — each a one-off event, not a cascade. Growth stopped when broadcasts stopped.

Even where k-factor briefly exceeds 1.0, it reverts to linear quickly — a series of short S-curves, not sustained exponential growth. Keep optimizing word-of-mouth and referrals to amplify each broadcast, but treat PR, influencers, and press as non-optional: they are the ignition.

Five steps to starting your product-led growth motion, part 2

PLG is data-led growth: without instrumentation, free users generate no actionable signal. Three infrastructure layers are non-negotiable — product analytics (Amplitude, Mixpanel) plus a customer-360 database merging usage with CRM and firmographic data; an experimentation platform (buy Optimizely or Eppo before building — homegrown requires underestimated data-science resources); and lifecycle tools (Customer.io, Braze) that trigger messages on behavioral events, not lead-nurture cadences.

Team structure runs three phases: start with a small growth squad or a cross-functional tiger team for bigger moves like launching freemium (MongoDB's path). Once PLG proves fit, formalize a growth org. The head of growth typically reports to a CPO; at companies like GitHub and Cloudflare, a CRO reporting line reduces quota conflict and aligns everyone around self-service ARR.

Lessons from Airtable’s unconventional growth strategy | Zoelle Egner

Airtable's early growth was built on a counterintuitive bet: invest in customer success before sales, and treat word-of-mouth from individual champions as the primary distribution mechanism.

Zoelle Egner joined as employee 11 in 2015. Generic messaging ("build your own software") failed — people heard complexity, not capability. The answer was specificity. The team built a Slack integration surfacing new signups with company and title, then manually emailed promising candidates. The target was the "tinkerer" persona: someone who would build a content calendar, then walk to a colleague and build a UX research tracker, becoming an internal hero without a salesperson involved. Buyers and champions looked completely different. Once embedded in a team's workflow a base was nearly impossible to remove; retention was exceptional and internal virality — 10 users growing to 1,000 within a company — was the real engine.

Templates served as enablers, not acquisition tools. The gallery was never an SEO play; Airtable consciously chose not to invest there. Templates narrowed the surface area so champions could see their specific use case and gave them something to hand a colleague. CS conversations fed a conveyor belt: each customer build became a template and blog post, scaling insights into reusable content.

On brand: Airtable bought remnant billboard inventory in Manhattan's fashion and media districts for a few thousand dollars — not for leads, but to signal legitimacy to IT departments about to approve six-figure spend. Every small touchpoint accumulated trust.

Things to skip: conference sponsorships rarely pay off; Gartner category creation is a massive lift buyers don't reward. Elevating a profession works better — Gainsight built the customer success career identity, which people defend. PR helps with hiring and cold-outbound credibility, not acquisition. Egner's unofficial metric: how many users got promoted using Airtable.

Growth inflections

Growth inflections cluster into three types: product improvements, external events, and doubling down on the primary growth engine — with the last likely producing the most durable gains.

Product changes dominate the case studies. Figma's biggest inflection came from Team Libraries in 2017, making it the obvious winner over Sketch. Facebook re-accelerated twice: translated interfaces in 2008–09 broke the English-speaker ceiling; then a full mobile shift in 2010–11 pushed it past its own 700M forecast to 2B+. Netflix unlocked growth by combining "no due dates, no late fees" with subscription pricing — traffic exploded within days. Snap drove successive waves through ephemeral messaging (2011), Stories (2013), and face filters (2015). Lyft cut driver onboarding from 21 to 7 days with a custom applicant-tracking system, enabling 24 city launches in a single day.

External events need no product change. Tinder spiked when Olympic snowboarder Jamie Anderson mentioned it at Sochi; PR amplified the cycle. YouTube's first viral hits were an SNL clip and a Nike-uploaded Ronaldinho video. Discord tipped from one planted Reddit comment in a Final Fantasy subreddit. Slack inflected on media consensus — Stewart Butterfield on Forbes in August 2015.

Growth-engine bets: Airtable's templates seeded an SEO flywheel that peaked at a third of total traffic; Pinterest claims 100–1,000x organic growth; Dropbox's Samsung OEM partnership drove 100 million signups; PayPal's $20 referral bonus produced 7–10% daily growth.

Behind the scenes of Calendly’s rapid growth | Annie Pearl (CPO)

Calendly's viral engine works because 70% of signups come through recipients of scheduling links. The flywheel started accidentally: founder Tope built the first version with a Ukrainian contracting firm whose CS agents used the product to schedule calls with parents in K–12 education; those parents spread it to their schools. The product launched free partly because billing infrastructure wasn't ready, and superior UX plus zero price powered early growth.

When Pearl joined, Calendly had ~150 people, no sales team, and all ARR from PLG. By recording time: ~600 people, sales-led growth at 20% of ARR, outpacing PLG. The transition required matching hire profiles to the motion (inbound/PQL-driven, not outbound hunters) and to the actual buyer (department heads, not CIOs).

The most consequential internal change was clarifying who to build for. The product had served freelancers, educators, sales, recruiting, and customer success teams simultaneously — making prioritization nearly impossible. Pearl anchored on three ICPs: sales, recruiting, and customer success professionals in external-facing roles. That clarity propagated into OPA meetings (PMs debate opportunity hypotheses without leadership present), PRD templates, and roadmap allocation. A Venmo integration — high demand from solopreneurs, wrong ICP — is the concrete example of what gets cut.

Strategy follows the Playing to Win framework: winning aspiration, where you'll play, how you'll win. Calendly's horizon model ran 70/30 (horizons one/two) in year one, 50/50 in year two, 30/60/10 by year three as enterprise and departmental use cases matured.

Planning commits only to work within sight: discovery, solutioning, then engineering estimation with a real ship date — no dates promised until both problem and solution are known.

The team-business — multiple users scheduling collaboratively — is growing faster than the solo-user base, and Pearl reads that as the real next growth curve: departmental and multi-department deployments rather than individual subscribers.

Product-led marketing

PLG math forces CAC below $1 per unique visitor — at 6% visitor-to-signup and 5% signup-to-paid, a freemium product earns only $1–$2 per visitor in first-year revenue. Two channels dominate new-user acquisition for top PLG products (Airtable, Miro, Snyk, Zapier): organic search (40% of new users) and product virality (16%).

Organic search runs four plays: free sidecar products with a conversion nudge (Snyk's open-source vulnerability database); job-to-be-done templates with SEO-optimized pages (Airtable's 200+ templates); programmatic "how to connect X+Y" landing pages at scale (Zapier's 70,000+); and documentation hubs built around high-volume feature terms (Hotjar ranking #1 for "heatmaps").

Virality splits into external (sharing with outsiders, as in Loom) and internal (colleague spread). Tactics include designing social features into workflows (Calendly's group meetings), removing friction with free-tier bridges (Loom's Creator Lite), and building community around the product (Webflow's 75,000-member forum).

The north star metric is activated signups — users who reach the aha moment — not raw traffic or signups.

How to build a cult-like brand | Laura Modi (Bobbie)

Bobbie — the only female-founded organic infant formula company in the US — was built on the insight that 83% of parents use formula but nearly all feel ashamed of it. Laura Modi, formerly director of hospitality at Airbnb, launched Bobbie in 2021 to create a formula parents could feel proud of, one that matched modern nutritional science and looked good on a counter. The core positioning: "a better-for-you infant formula without the guilt."

The 2022 competitor recall emptied US shelves and Bobbie's subscriber count doubled in a week. But head of growth Shireen ran the inventory math and gave Modi six days before they'd fail existing customers. She shut the website for six months — no new signups, growth team renamed the "slowth team," actively emailing subscribers to cancel if they wanted. The decision made Bobbie the only formula brand that never ran out, cementing loyalty that no ad spend buys. A stranger stopped Modi at Davos in tears to describe what it had meant.

Brand architecture runs content → community → commerce. Bobbie launched Milk Drunk, a separate content platform, years before needing it — building SEO authority on functional searches like "how long does formula last?" until Milk Drunk appears between the CDC and WebMD on the first page. Paid marketing is treated as a drug, kept deliberately small. Modi attributes 60% of growth to product and brand alone; the remaining 40% is distribution and word-of-mouth.

She hires "optimistic doers" and deliberately picks non-specialists: Bobbie's head of marketing is a former Emmy-winning news anchor who didn't know what CAC meant. Naivety, in Modi's framing, is the operating condition for genuine innovation.

Momentum has to be manufactured. Set an arbitrary launch date and say so out loud.

Mastering paid growth | Jonathan Becker (Thrive Digital)

Performance marketing fails as a single-channel bet. Jonathan Becker runs Thrive Digital managing $500M/year in ad spend for clients including Uber, framing marketing as portfolio management: channel diversification reduces exposure to any one platform's algorithm shifts or CPC volatility.

Creative is now the primary lever because Google and Facebook have automated nearly everything else. Unpolished mobile-shot UGC beats brand-studio assets — an influencer on an iPhone outperforms a produced spot. One furniture client tripled ROAS by adding a dog to product photos. Testing discipline: run two near-identical creatives differing on one variable, normalize by click-through rate or impressions-to-conversion to correct for unequal impression distribution.

The most common B2B mistake is optimizing for cost-per-lead over lead quality. The fix: pipe CRM revenue data into a channel database, build a lead scoring model, and bid in real time on audiences likely to generate high revenue — even when that revenue arrives 6–12 months later.

Attribution has no single source of truth. Cookie-based last-click was the default; iOS 14.5's IDFA removal destabilized it overnight. Best practice layers cookie attribution, media mix modeling (tools like Recast), and customer surveys — treating it as an ongoing investigation, not a resolved number.

TikTok is where Facebook was six or seven years ago: cheap CPCs, weak attribution, heavy creative demand. It works for D2C brands with founder-led content or strong influencer networks; B2B is unproven but coming.

Technical grounding matters more than experience — Thrive recruits from physics, mathematics, and engineering. Agencies and in-house teams aren't substitutes; agencies need an internal POC to function.

AI has displaced trench work — bid management, manual analysis — shifting staff toward modeling and creative strategy. MidJourney and Dall-E produce mockups in 1% of prior time; ChatGPT drafts RFP responses at 80% quality, cutting a five-person week to an afternoon.

The power of strategic narrative | Andy Raskin

The most effective B2B pitches don't open with a problem and solution — they open with a shift in the world that makes the old game obsolete. Andy Raskin calls this a strategic narrative, and its structure comes from screenwriting: every story begins with a change that forces the protagonist to act.

The framework has five pieces. First, name the shift from old game to new game — Salesforce: "software is over, cloud is the new game"; Zuora: "transactions giving way to subscriptions"; Gong: "opinions giving way to reality." The naming must be blunt and compact even at the cost of precision. Second, name the stakes: winners are already playing the new game and clinging to the old risks extinction. Third, name the object of the new game — Zuora's was "turn every customer into a subscriber," a rallying cry that doubles as company mission. Fourth, name the obstacles: the specific challenges that make winning hard. These reframe what would be feature comparisons into emotionally resonant plot problems. Fifth, show how your product overcomes those obstacles.

The payoff is threefold: salespeople pitch a movement instead of features; the narrative filters the product roadmap (Gong declined feature requests that reintroduced opinions; 360Learning uses "does this help upskill from within?" as its bar); and it scales founders' judgment across go-to-market without them being in every room.

360Learning had been pitching "collaborative learning" — a label that invited constant competitor comparisons. Reframing around the shift from top-down training to upskill-from-within ended the "how are you different from X?" question entirely.

Category creation is adjacent but weaker: Gong's Amit Bendov says they could have called their category "strawberry intelligence" — the story was what moved buyers, not the name. CEO sponsorship is non-negotiable; product-led versions lack the air cover that makes marketing, sales, and roadmap cohere around a single story.

What working at Figma taught me about customer obsession

Figma's 150%+ net dollar retention, 90% gross margins, and $400M+ ARR trace back to one discipline: the product exists to serve customers, not the other way around. Sho Kuwamoto, VP of Product for seven years, frames this as a "service-oriented mindset" — think of Figma as a hotel where software is just the delivery mechanism for helping people do great design.

Three practices sustained this. First, customer support was everyone's job from beta onward; engineers spent up to 20% of their time in support, which paradoxically accelerated velocity because the team learned what mattered faster. Second, product decisions inverted jobs-to-be-done: instead of "what makes the product better," the question became "what serves the user's highest-level goal," cutting through competing metrics. Third, Figma launched with a narrow feature set — no components, styles, or multiplayer — because bad feel creates a credibility hole that takes years to escape. For high-frequency tools, feel drives adoption; Figma held a 60fps render bar and matched designer keyboard muscle memory from day one.

Leveraging growth advisors, hiring well, mastering SEO, and honing your craft | Luc Levesque (Shopify, Meta, TripAdvisor)

The right growth advisor can say one sentence that changes a company's trajectory — years of experiments compressed into seconds of communication. Luc Levesque (chief growth officer at Shopify, previously recruited by Zuckerberg to grow Messenger, Instagram, and WhatsApp) treats vetting an advisor more like a Series A bet than a consultant hire.

On advisors: wait for product-market fit first — accelerating a product users don't love burns first impressions at scale. Compensate in equity with a three-month cliff; no early impact, both parties walk. Vest front-loaded so the advisor is incentivized to transfer knowledge fast. To find one: start with VCs, use existing advisors to vet candidates (underused shortcut when you can't assess growth depth yourself), and look at who strong growth companies — Canva, Pinterest, Miro — hired.

On hiring: look for signs of excellence — repeated outstanding outcomes across any domain. A strong signal: a former boss who left a company and came back to poach someone, staking their own reputation on the pick. For critical hires, involve the whole executive team and the candidate's spouse; Zuckerberg did both over seven months to land Levesque.

On SEO: divide products into small-page sites (need a content strategy) and large user-generated or marketplace sites (huge optimization surface, faster lift). Entire travel companies traded on single keyword rankings. AI boxes atop Google results will gut informational query traffic; transactional queries are safer. Hire an internal SEO owner first, surround with advisors; agencies are a last resort. Impact timelines: three to twelve months for new content, potentially immediate for pages already near page one.

On craft: one hour of structured daily self-reflection — a red/yellow/green dashboard across work, parenting, and health, plus a running experiment log. The Zuckerberg lesson carried forward: optimize for impact, not activity.

Building high-performing teams | Melissa Tan (Webflow, Dropbox, Canva)

High-performing teams need four things: clear goals tied to measurable outcomes, a team-first culture over individual credit, an ownership mindset where people exhaust every option before feeling blocked, and deliberate fun. Results orientation and deep people investment aren't in tension — measuring success clearly and feeling genuinely cared about reinforce each other.

Dropbox illustrates both sides. Hiring first-principles thinkers — including people with no sales background to staff the sales floor — drove go-to-market innovation because those hires had no playbook to default to. What went wrong: delaying the B2B enterprise motion let Box fill that space, and growth sat on top of product rather than integrated from the start. The lesson: decide go-to-market structure and pricing early, before self-serve momentum makes correction painful.

On developing talent, give direct feedback within weeks — stated with intent ("I believe you can do this differently") and a concrete offer of support. Kim Scott's "care personally, challenge directly" from Radical Candor, workshopped at Dropbox before publication, captures this exactly. Developing internally de-risks scaling: institutional knowledge compounds in ways outside hires can't replicate.

When hiring PMs, the highest-signal step is a prep call before the final presentation: review the draft, give feedback, watch how they incorporate it. Candidates who ignore feedback rarely get an offer. Live screens include pulling up the actual product and asking what the candidate would test.

For growth teams, the most common pitfall is starting without a go-to-market blueprint: product-led vs. sales-led, value metric for pricing, how growth integrates with product. The "flying formation" — a DACI (Driver, Accountable, Contributor, Informed) plus explicit meeting cadences and metric ownership — keeps growth from being bolted on. The first growth hire should think like a portfolio manager: analytical, testing channels from scratch rather than a single-channel expert where expertise creates false precision.

Lessons from scaling Ramp | Sri Batchu (Ramp, Instacart, Opendoor)

Ramp reached $100M ARR in two years — fastest in FinTech history — with under 500 employees. Sri Batchu attributes early traction to "cap table as growth strategy": co-founders placed early-stage founders on the cap table who became immediate customers. Revenue now comes mostly from mid-market and enterprise; growth is still overwhelmingly net-new acquisition rather than expansion.

The channel mix is standard — outbound, paid, field, SEO, lifecycle. The edge is technology density inside each. A dedicated growth-engineering team embeds AI into the sales workflow (lead sourcing, message drafting, response prioritization) and owns pipeline quota rather than product metrics.

The north star is "dollars of SQL pipeline." Every sub-team's local metric maps through a translation layer to that number, refreshed every six months — a common currency that removes most prioritization debates and focuses judgment on genuinely marginal calls.

On experimentation: growth tests succeed about 30% of the time, so the culture demands conclusive failure, not just fast failure. B2B can't get large sample sizes quickly, so the fix is maximizing treatment effect — throw every relevant tactic at a hypothesis simultaneously. If the kitchen-sink version fails, retire the idea. This stops underpowered tests from being re-run every time new leadership arrives.

Payback period using contribution margin beats both CAC and LTV/CAC. CAC optimization attracts cheaper but less valuable customers; LTV is too assumption-laden for a young company.

B2B channel sequencing: founder-led sales → first hires → content/community/events → PR → paid/brand → SEO. SEO lands late because domain authority takes time without existing media presence.

For hiring, use SimilarWeb to identify which companies excel at the function you're filling, then source from those teams. On pay: widen bands — 10x operators should earn 10x — and talent density depends as much on managing out poor fits as on recruiting.

What 5 years at Reddit taught us about building for a highly opinionated user base

Passionate users behave like parents defending their children — the 2015 Reddit moderator strike that ousted the CEO in under a week shows what ignoring that reality costs.

Four tools emerged from five years at Reddit. First, the Trust Vault: a measurable reserve tracked via Edelman-style surveys ("How much do you trust our staff to do the right thing?"), once or twice a quarter. When moderator scores fell, the team shipped targeted wins (MEOWs) before touching anything controversial.

Second, disciplined listening. Loud feedback isn't representative. During Reddit's first desktop redesign in 13 years, the team replaced a sidebar based on vocal complaints, then faced a larger backlash demanding it back. Weight feedback by breadth (10%+ of users?) and influence (can this person shift others?). Explain why when deprioritizing a group.

Third, a private advisory council — ~80 rotating members on 12-month terms — consulted before public announcements. Public threads incentivize winning arguments; private calls produce honest feedback and turn council members into launch advocates.

Fourth, triage via a breadth-×-depth 2×2. Wide and deep: do it. Narrow and shallow: skip. For middle cases, explain why when saying no — asking for feedback then ignoring it burns more trust than never asking. When a vocal minority faces a change daily and the fix is cheap, throw a bone: classic and condensed density options during the redesign deflected most complaints.

During the five years this playbook ran, moderator protests were rare and Reddit crossed 100,000 active communities.

How today’s top consumer brands measure marketing’s impact

No single marketing measurement method can tell you whether a sale would have happened without the ad — the counterfactual is unobservable — so leading brands triangulate across three techniques. Research across 42 brands (McDonald's, Uber, Airbnb, Netflix, HelloFresh, DoorDash, and others) shows over 40% use at least two methods together and around 20% use all three.

Multi-touch attribution (MTA) tracks click paths via cookies and UTM codes. Google Analytics (on 60% of the top 100,000 websites) defaults to last-touch, which over-credits lower-funnel channels like search. DoorDash still uses last-touch despite its known bias; Uber built a custom system ("Euclid") to weight impression-level data across the full conversion journey. MTA is useful for daily campaign tweaks but cannot establish causality.

Marketing mix modeling (MMM) correlates aggregate spend with sales using statistics — developed in the 1960s, now resurging after iOS 14 cut MTA's tracking reach. Meta's open-source Robyn library democratized it; Resident (Nectar mattresses) used Robyn to achieve 20% revenue lift at the same blended CPA. HelloFresh uses Bayesian MMM specifically to quantify offline channels like TV. MMM struggles at campaign granularity and takes three to six months to build in-house.

Conversion lift studies (CLS) are the closest proxy for true incrementality: randomly withhold ads from a geographic region and measure the gap. Netflix uses city-level geo-testing for billboards. Uber measured a 6.57% statistically significant conversion drop when suspending one Google UAC campaign — the same methodology that exposed $100 million in fraudulent spend by proving those conversions would have happened regardless.

Triangulation in practice: use MTA for daily optimizations, MMM for cross-channel budget allocation, and CLS (once or twice a year) to calibrate the MMM. McDonald's ran a Meta geo-test to establish ground truth, then tuned its MMM until both methods agreed on Facebook's causal contribution.

What is good free-to-paid conversion

Freemium self-serve products (Canva, Typeform) convert at 3–5% (good) and 6–8% (great); add a sales-assist motion and thresholds rise to 5–7% and 10–15%. Free-trial products run higher — 8–12% good, 15–25% great — because they attract buyers close to purchase; 44% of free-trial companies have sales reach out to over half of sign-ups, double the freemium rate. Developer-focused products convert at roughly half the median. Four levers move conversion: (1) sales outreach triggered by product-qualification signals (multi-player use, usage growth, role, use case); (2) strong day-one onboarding — 40–60% of users never return after day one; (3) replacing sign-up volume with activated sign-ups as the marketing KPI; (4) the reverse trial — full premium access then a downgrade if unpaid — which converts at twice the classic freemium rate.

The ultimate guide to Martech | Austin Hay (Reforge, Ramp, Runway)

MarTech is systems architecture at the crossroads of engineering, product, and marketing — not tool procurement. The job is owning how data flows through first- and third-party platforms, designing that flow one to two years ahead, and controlling cost as the company scales. At roughly 100–150 employees the village approach breaks down: someone must own schema, contracts, permissioning, and liability. B2C aligns the function to growth; B2B/B2B2C anchors it to revenue operations, since Salesforce is the gravitational center.

The B2C stack split around 2020 when warehousing got cheap. CDP-centric architecture (Segment → integrations) stays valid for engineering-light teams. Engineering-heavy teams route data into Snowflake, model in dbt, then activate outward via reverse ETL tools like Hightouch or Census. Either path requires an explicit policy for when to pull from the ingestion layer versus the warehouse — skipping it creates duplicate flows and attribution chaos.

Mobile attribution broke with iOS 14, ending deterministic IDFA matching. The response is probabilistic modeling: build a model from the ~30% of users you can identify and extrapolate. For web, collect first-touch and last-touch UTMs plus ad-network click IDs on every user profile and page-view event from day one, in first-party cookies. Deferring makes multi-touch reconstruction impossible later. Most companies need better MTA before MMM.

"Build and buy" beats "build versus buy": acquire a third-party tool for 90%, build custom infrastructure on top. Vendors with large customers investing around their platform become more responsive on features and SDK changes.

Hiring signals: intellectual curiosity and engineering scrappiness — six months of self-teaching is sufficient. Best interview question: ask how the candidate prepared; systems thinkers answer unprompted. Red flag: proposing familiar tools before diagnosing the problem.

B2C stack: Amplitude (CDP + analytics), Customer.io → Braze, Snowflake, Hightouch, AppsFlyer. B2B: same core, Branch for web attribution, Salesforce as CRM.

How to hire your first growth team

Growth teams don't build new products — they distribute existing value. Acquisition is the right first focus 90% of the time: that's where the largest drop-offs occur and wins compound into retention and monetization downstream.

Timing is decisive. Pre-PMF, founders should only develop a growth model hypothesis — product-led, marketing-led, or sales-led — because that choice shapes the product's architecture from day one. During early traction, founder-led growth is the right vehicle. The first dedicated hire belongs only after PMF is validated and distribution methods confirmed, matched to whichever bottleneck is most painful: Acquisition PM, Activation PM, or Growth Marketer.

Don't hire a growth leader first. Even experienced leaders need 6–12 months to understand the local problem; under pressure they copy-paste from prior companies with poor results. Hire a "builder" — a generalist executor within one domain — first.

Two structural archetypes: centralized (Dropbox, GitLab, Snyk) optimizes for velocity but risks the "absolution effect" where the rest of the org stops feeling accountable for growth. Decentralized tribes (Amplitude, Pinterest, Airtable) spread ownership but lose velocity to competing departmental priorities. Both Amplitude and SurveyMonkey oscillated between models. When uncertain, start decentralized and switch when velocity becomes the bottleneck.

An inside look at Deel’s unprecedented growth | Meltem Kuran Berkowitz (Head of Growth)

Deel went from under $1M ARR in July 2020 to $295M by early 2023 — arguably the fastest SaaS growth ever — while staying EBITDA positive. Meltem Kuran, the second marketing hire, built growth almost entirely on cheap, non-paid channels that still account for roughly half of pipeline today.

The foundational tactic: set keyword alerts on Reddit, Quora, and closed Slack/Discord communities for terms like "hire someone in Germany," then send real people to answer substantively — never pitching, never automated. Gauge each community's ceiling first; a 1,000-person subreddit yields at most 100 customers. VC networks proved high-value partners because founders trust them for tool recommendations.

SEO runs on the same logic. The quality test: after reading, does the user return to Google? If not, the search is over — that's what search engines reward. Research starts with ~700 terms scored by buyer intent: green (convert-ready), yellow (50/50), red (research-only). Work greens first by volume. Clearscope checks readability (target: fourth-grade level). Eight people publish five new articles and five updates weekly — updates matter because compliance and payroll regulations change constantly.

Channel sequencing: website speed first, then content, then paid ads. B2B awareness campaigns are a trap early on — six to eight months of bottom-funnel work comes first.

Team structure separates functional experts (paid ads, content, brand) from regional managers, so specialists compound skills rather than being fragmented. First three growth hires: product marketing, copywriting, and a data analyst — the last unusually early, but correct. Growth KPIs are revenue, not leads; candidates who'll only commit to lead volume don't get hired.

Culture runs on "Deel speed" (urgency as a first principle), default optimism, and "little hands" — no one too senior for the nitty-gritty. Values were named about a year in, after they already existed in practice.

How to identify your ideal customer profile (ICP)

Picking the wrong audience is as fatal as picking the wrong problem — most B2B founders got their ICP wrong initially, and several (Front's Mathilde Collin, Census's Boris Jabes) cite it as their biggest early mistake.

The practical standard: land on at least three attributes, almost comically narrow. Gusto used six — five or fewer employees, in California, no benefits, salaried only, with deductions, paid eight days after payroll — then expanded one constraint at a time. Gong targeted US English-language software companies selling via video conferencing for $1k–$100k: roughly 5,000 companies globally. Snyk went after Node.js developers who were security-conscious. Linear filtered its waitlist to founder-led product companies using Google Auth and GitHub.

Discovery signals are consistent. Track who gets most excited — Canva spotted social media managers six months in. Track conversion rates by segment — Stytch saw 5–10% with growth-product personas versus 40–50% with technical buyers, then dropped the former. Watch for enthusiasm shifts — Census kept pitching business ops until data scientists responded with clearly more energy. Hex's Barry McCardel listened for "the nod": when he described pain as "screenshotting charts into Google Docs," the right prospects visibly recognized themselves.

Two counter-intuitive lessons: outbound signal is cleaner than investor-referred leads, which generate fake interest from founders who just wanted to make friends. And company size matters as much as role — Ramp found Series A/B companies wanted profitability and on-time books, while early-stage startups asked for perks.

ICP evolves. Segment moved from IC engineers to engineering directors to enterprise digital-transformation teams across three phases. Notion cycled through personas for years before landing on product/engineering/design/data as its company beachhead. If nothing is working, compress the time horizon: only pursue people who can buy this month.

Inside Figma's go-to-market: Building community, driving bottom-up growth, and crafting unforgettable launches | Claire Butler (first GTM hire)

Figma's growth came not from traditional marketing but from making individual designers love the product so deeply they became internal champions who spread it through their organizations — a bottom-up motion that ran without a sales team for the first three years.

Building credibility with a technically skeptical audience came first. Designers filter out marketing language instantly, so Figma's early blog posts were deep technical pieces — how Evan built a design tool on WebGL, posts on vector networks and grids citing Joseph Müller-Brockmann — content Butler could not have written herself, her quality bar. The first non-engineer hire was a designer advocate: a passionate user bridging community and product, gut-checking every marketing move and feeding feedback into engineering.

On Twitter, Dylan built a scraper mapping the design community's influence graph — clusters of iconographers, graphic designers, product managers — and used it to DM people for feedback, not to sell. Going where designers already were was the core distribution logic. Technical content seeded into Designer News and Hacker News gave designers reasons to return before multiplayer shipped a year after launch.

To spread adoption inside companies, Figma flipped its freemium constraint: unlimited collaborators on free (not unlimited files), removing the paywall from the viral step. Viewers are always free; only editors pay. Designer advocates joined sales calls — never quota-carrying — and deals closed far more often when one was present, a dynamic called the Tom Factor. Design systems, initially the biggest enterprise blocker, became the primary upgrade driver: deep content, a dedicated conference (Schema), and DesignSystems.com converted the obstacle into the reason to move from Pro to Org.

Champions don't stop mattering after the sale. Figma amplified them — conference speaking slots, social promotion — turning internal champions into thought leaders and closing the flywheel back to credibility.

How to win as a PM and grow at any company | Oji Udezue

The biggest predictor of PLG success is picking a problem sharp enough that customers become your marketing department. Oji Udezue (Typeform, Twitter, Calendly, Atlassian) organizes B2B opportunity around a workflow-frequency quadrant: breadth (how many departments use the workflow) against frequency (how often). High-frequency niche workflows — Jira, Salesforce, Calendly — compound fastest because mission-criticality drives daily engagement. Low-frequency plays like forms can win by carving out a high-frequency foothold inside a single department, then expanding. The zone of benefit frames the companion problem: people don't notice improvements below 3x. At that threshold customers switch; below it, inertia wins. Calendly beat earlier scheduling tools not through its viral sharing loop — that was a consequence — but because it compressed scheduling far enough that sharing became irresistible.

Virality is customer-augmented marketing: customers marketing for you. Slack had no referral program and couldn't connect users across floors of the same building, yet spread by lunch-table word of mouth. Synthetic virality (referral tags, sharing prompts) only amplifies a product that already earns organic word of mouth; on a mediocre product, it accelerates churn instead.

Customer discovery is three distinct practices: targeted discovery (conversations on a specific workflow), continuous conversations (PMs and designers auto-scheduled with customers weekly, no opt-in friction), and customer listening (processing standing signals — NPS verbatims, support tickets, G2 reviews, closed-won notes). The third is underbuilt industry-wide and requires no discovery conversations.

Frameworks are mental shortcuts — useful only if you understand the empirical relationship underneath them, so you can adapt by stage. What works at ten people is wrong at scale.

From Bridgewater: evaluate people on skills, attributes, and values. Google's skills-only interviews predicted performance at roughly 50/50. Forest time — one structured day per month off execution — corrects aim before a misaligned direction wastes the entire downstream build.

Brian Balfour: 10 lessons on career, growth, and life

Building a great product is necessary but not sufficient — distribution is the separator. Startups win, per a16z's Alex Rampell, by acquiring distribution before incumbents copy. That game has gotten harder: incumbents copy faster, organic channels (SEO, LinkedIn, social) are shrinking, and AI floods markets with near-identical competitors.

New distribution platforms are the relief valve, following four steps. Step Zero: consensus forms around a category with no clear winner. Step One: the leader identifies its moat and opens a third-party platform, trading distribution for use-cases and engagement. Step Two: the gold-rush. Step Three: closure — organic reach suppressed, top use-cases absorbed first-party, developers pushed toward paid. Facebook did this in five years; Google over decades; LinkedIn twice. Cycles are getting shorter.

Balfour's prediction: ChatGPT is the next major platform. The moat is context and memory — models are converging on quality, so whichever accumulates the most personal context wins. ChatGPT leads on MAU (roughly 10x Claude) and shows a retention smile curve, the historical signal of escape velocity. Apple is best-positioned architecturally but not executing; Google likely has inflated MAU from accidental clicks.

There is no opting out — if you don't integrate, competitors will and expectations shift. Late-stage companies can spread bets; startups must pick one and go all in. Criteria: prioritize retention depth over raw MAU, user monetizability, and the developer value exchange. Once in, plan the exit — own a critical workflow layer, accumulate specialized data, or build network effects the platform can't absorb.

On AI adoption: companies seeing results impose hard constraints — headcount capped at one-fifth of peer norms, no new hires until AI alternatives are disproven. Roughly 10% exit non-adapters and are furthest along. Adoption is a system constrained by its slowest link. Accelerating engineering without accelerating product and design just shifts the bottleneck.

Scaling your B2B growth engine

Top B2B startups averaged two years from founding to $1M ARR, or about 1.5 years after their first customer. Outliers like Ramp and Linear got there in months; Loom waited four years intentionally to achieve ubiquity before monetizing. The practical target: $1M ARR within 18 months of signing your first customer.

Organic inbound is the dominant growth channel across the 20+ companies studied. It splits into self-serve (users activate without sales contact: Loom, Figma, Linear, Gusto) and sales-assist (inbound leads handled by reps: Looker, Ramp, Vanta, Persona). Outbound is the primary channel for Gong and Zip. Content/SEO, paid ads, and partnerships supplement but rarely anchor. The practical rule: optimize whichever of the top three is already working, and accept that every B2B company eventually builds a sales team. Databricks tried a zero-touch PLG motion in 2015 and revenue flatlined within two quarters before they reversed course.

On pricing, the near-universal lesson from Amplitude, Front, Zip, Notion, and Gusto is: charge sooner and charge more than instincts say. Founders systematically underprice because building from scratch distorts perceived value. Notion charged early, became profitable at eight people, diluted under 3% across three fundraising rounds, and never spent the capital raised. Sprig's CEO watched his first sales hire quote seven-figure deals he would never have imagined. Segment doubled prices on an advisor's instruction and kept doubling. Early pricing should be simple — Stytch's mistake was not revisiting their model for 18 months, not that they started simple. Figma benchmarked against Sketch at $12/month and never changed it. Snyk rejected test-count and lines-of-code as billing metrics before landing on "contributing developers in the last 90 days." Coda's "maker billing" charges only creators, leaving editors and viewers free, specifically to keep the share dialogue free of dollar signs and protect the viral loop.

Introducing DRICE: a modern prioritization framework

RICE alone is insufficient: a two-stage process — fast T-shirt-size RICE to cut a preliminary shortlist, then a 30-minute "DRICE" (Detailed RICE) deep-dive on each finalist — doubles experiment impact rates. RICE scores Reach, Impact, Confidence, and Effort at S/M/L to produce a ranked shortlist of roughly 2x what you can build in a quarter. DRICE then converts those relative scores into a concrete hypothesis, a bottom-up dollar impact model, and a real engineering breakdown. At Dropbox, DRICE surfaced a Business Migration Tool that initial RICE had ranked low; it became the biggest activation win of the quarter. Teams using DRICE showed twice the impact rate of those using RICE alone. Secondary payoffs: dollar-denominated forecasts simplify resource conversations with executives ("X engineers will produce $Y next quarter"), and transparent scoring criteria motivate engineers to champion their own ideas competitively — undermining the HIPPO (highest-paid person's opinion) dynamic that otherwise dominates planning.

Lessons on building a viral consumer app: The story of Saturn

Saturn—a social calendar for high schoolers—reached #4 overall in the App Store by anchoring on single-player utility first. Co-founders Dylan Diamond and Max Baron, UPenn students at launch in 2019, grew to millions of users across nearly 20,000 schools with no school partnerships.

The core insight: pure-social apps decay nonlinearly. Lose a few creators and the product collapses for everyone remaining. Saturn's calendar works even solo, keeping D30 retention in the mid-30s—on par with Snap, TikTok, and Facebook.

Hyper-personalization drove early adoption. They white-labeled separate apps for the first 17 schools (iWeston, etc.), colorized to each school's branding. More than half of Weston High School joined in the first three hours. At school 18, they consolidated into one app while preserving the personalized feel.

The ambassador program—students supplying schedule data and seeding growth—was explicitly unscalable. They ran it through the first 1,000 schools anyway, then productized each function: a 2021 waitlist pulled 100,000 students from 10,000 schools in 90 days; crowdsourcing replaced manual calendar upkeep once the network hit sufficient density.

Founders' lived experience mattered: high school schedules are genuinely byzantine, and age proximity made ambassador trust-building possible. Competing apps sold to administrators; Saturn's student focus was the differentiator.

Long-term discipline meant skipping features that spike installs but harm retention, trusting that a new freshman class arrives every year to deepen the network.

The full-stack PM | Anuj Rathi (Swiggy, Jupiter Money, Flipkart)

Most new users are lazy, vain, and selfish — they resist habit change and demand immediate value. That frame, from Scott Belsky, drives Anuj Rathi's core PM philosophy: build for the reluctant non-user. At Swiggy, this meant threading one value proposition from brand exposure through marketing through onboarding. At Jupiter Money, the same principle governs cross-selling: earn the second product introduction by serving the first well.

Amazon's working backwards press release becomes a coordination tool. Rathi writes three divergent PR FAQs per initiative so leadership compares actual trade-offs rather than rubber-stamping one pitch. FAQs embed mandatory checklists by context: compliance sign-offs at a fintech; multi-sided impact at Swiggy, where a change to delivery-partner earnings ripples onto restaurant commissions and consumer fees. Always champion one option; have the alternatives documented.

The Four BBs framework structures resource allocation: Brilliant Basics (tech health), Bread and Butter (backlog improvements), Big Bets (cross-team initiatives), Breaking Bad (identity-level pivots — Swiggy becoming a "convenience company"). Which bucket gets focus points is a CEO-level call. Rathi proposes three allocation scenarios with explicit consequences before asking leadership to choose.

Marketplace dynamics break standard tools. OKRs fail in three-sided systems because every lever moves the other two simultaneously; A/B tests mislead because network effects couple both cohorts. Big-bet framing — pulling levers whole — works better. Stabilize all sides first, then declare a north-star: Swiggy's was the end consumer.

PM fitness requires three things: raw problem-solving ability (coachable via domain knowledge), drive (hardest to develop), and influence (non-negotiable — if improving at it doesn't excite you, the role will make you miserable). When teams fail, the root cause is almost always setup — bad org design or misaligned OKRs. Rathi estimates 70–80% of failures sit here — the org's architecture is already written into the product via Conway's Law.

Geoffrey Moore on finding your beachhead, crossing the chasm, and dominating a market

Pragmatists — the mainstream market — will not accept a visionary's endorsement. They buy from peers, and until peers exist the market stalls. Companies fall into the chasm not because the product fails but because they use the wrong playbook.

Four phases, four playbooks. Early market: visionary sponsors fund you outside normal budget; pitch the vision, land a marquee name. Bowling alley (crossing the chasm): shut the laptop and let the prospect talk — the sale is diagnosis, not demo; positioning is that you alone are committed to their exact problem. Tornado: budget appears market-wide simultaneously; land-grab share. Main Street: the product is commoditized and services carry the innovation.

Target-segment formula: big enough to matter, small enough to lead, fit with your strengths. Same geography × same industry × same profession × same use case. Win 30–40% of one segment and an ecosystem forms around you; spread across segments and none tips. Documentum started in pharma (500,000-page FDA filings, $1–2M per lost patent day) then expanded to petrochemicals, oil-and-gas leases, Wall Street — each hop sharing a use case or customer base.

Playbook mismatches invert results. Qualifying on budget kills early-market deals. Discounting signals low confidence in a risk-bearing decision; pragmatists want over-commitment to their problem, not a price cut. The compelling reason to buy is their pain, not your pitch. An enterprise salesperson defaults to horizontal coverage — the chasm needs a diagnostic sales engineer.

Product-led growth cannot cross the chasm; it works in the Main Street expand phase where risk is too low for one to form. PLG companies still need a consultative motion. The model is B2B-native; consumer markets follow different dynamics.

Entrepreneurs who build original software are a scarce resource — the goal isn't a billion dollars, it's enough to live well and then have impact.

The ultimate guide to willingness-to-pay

Pricing is the most under-leveraged growth lever — a McKinsey analysis found a 1% improvement increases profits by 11%, yet 50% of software companies surveyed have never run a pricing study. Behavioral scientist Kristen Berman breaks down four quantitative WTP methods.

Van Westendorp asks four open-ended threshold questions (too cheap / bargain / expensive / too expensive). Simple, but suffers hypothetical bias: in John List's donation study, pledges ran 2.5× higher than actual payments. Use only for familiar categories with bias-reducing framing.

Becker-DeGroot-Marschak (BDM) adds skin in the game: participants name their price knowing a random draw may trigger actual purchase. This mostly eliminates hypothetical bias — validated in a Ghana water-filter study — but the mechanism confuses many participants.

Multiple price list (Gabor-Granger) asks yes/no to a list of prices. A 2023 meta-analysis by Gao, Huang, and Jung found it underestimates WTP by priming attention to opportunity cost.

Discrete choice presents product bundles at varied feature-price combinations, repeated 5–7 times. Optum used it and killed a planned launch after finding WTP far below assumptions; Fivestars discovered a new pricing tier the same way.

Decision rule: familiar purchases → Van Westendorp or BDM; unfamiliar or high-ticket → discrete choice. Always include an incentive-compatible element or responses overstate willingness. For in-market tests, start high and adjust down: Apple did it with the iPhone; Twitter Blue dropped from $20 to $8. Framing shapes WTP as much as the price: "sauceable" pasta commands 80% more, and Miro charges $960/year for a whiteboard by calling it "the visual workspace for innovation."

How to successfully launch on Product Hunt (when it's right for your startup)

Most Product Hunt launches fail — Leo Bosuener of Social Growth Labs, who has put 60+ startups at #1, turns away 70% of applicants. The platform suits brand awareness, early-adopter feedback, and a high-authority SEO backlink, not sustained acquisition. Campaigns cost 50–120 hours of prep; paid ads deliver better ROI for first-time message testing. B2C tools see 500–1,500 signups; B2B products 50–300, and only with fully self-serve onboarding.

Seven execution levers: (1) Set conversion goals — signups and trial starts — not a ranking target. (2) Warm up your network 30 days out; supporters need seasoned PH accounts because same-day new-account votes get filtered. (3) Launch at 12:01 a.m. Pacific and stagger outreach in three waves (12:01 a.m., 7 a.m., 2 p.m. PT) — steady velocity wins; simultaneous voting from one room risks a shadow ban. (4) Choose launch day by goal: Mon/Fri for rank-to-traffic balance; Tue–Thu for maximum traffic; weekends for best #1 odds. (5) The algorithm weighs upvote velocity, upvoter karma, and geographic diversity — raw count doesn't decide rank. (6) Four assets matter: action-verb tagline, UI screenshots with embedded captions, a simple Loom walkthrough (3–5 min), and a maker's comment that opens with the pain point and closes with a trial CTA. (7) Substantive power-user reviews convert far better than upvote counts.

Five myths: famous hunters no longer trigger follower notifications; Product of Week rarely beats capitalizing on Day momentum; highest upvotes doesn't guarantee top rank; existing products can relaunch on major updates; direct links to your PH page are encouraged.

The ultimate guide to PR | Emilie Gerber (founder of Six Eastern)

Press earns its keep through second-order effects, not direct signups. For B2B, the real return is third-party validation — a logo in an outbound email, a link an AE attaches to a sales sequence. Perplexity is the B2C exception: zero paid marketing, pure press, immediate conversion spikes — the product takes one try to prove itself.

Publication fit matters more than prestige. TechCrunch is the default for funding rounds (nine stories a week out of ~50 actual) — pitch an exclusive, lead with amount, investors, one-line incumbent framing. Axios is deals-only. Business Insider takes pitch-deck exposés and requires revenue metrics. VentureBeat owns AI. Fast Company wants future-of-work angles and accepts op-eds. Forbes contributor network is lower-bar for early-stage companies. WSJ and NYT require on-record valuation and revenue — relationship targets until Series C+.

Cold outreach works as well as warm introductions. Three-sentence pitches outperform longer ones. Never lead with category creation — "X taking on Bill.com with Y approach" gives reporters instant framing. The RAMP bill-pay launch landed CNBC only by framing as a threat to Bill.com's market share. Contrarian angles travel: when Shopify's meeting-cancellation got universal praise, Clockwise placed an op-ed arguing blanket deletion doesn't fix meeting culture. When nine AI companies share a stance, the tenth gets the call.

Press releases are largely obsolete; a founder blog post does the same job more shareably. Percentage growth stats from near-zero are ignored; specific revenue figures with context land.

Podcasts and newsletters belong in every PR program. Pitch hosts on LinkedIn with "open to guest ideas?" before any formal pitch. Ask customers what they listen to. Morning Brew's Coworking series runs via Google Form at ~50% success.

Agency retainers start around $10K/month; one-off projects run $8–20K. Six weeks lead time is optimal; flex the date to get the right reporter.

The secret to Duolingo’s exponential growth

Duolingo grew daily active users from 5 million to nearly 30 million in five years by treating urgency as a compounding asset. Because ~90% of DAU growth comes from word of mouth, a 1% retention improvement on 100,000 DAUs adds 1,000 users in week one and 1,072 the next — the gain grows every week it's live. Delaying an Android port by 102 days (to polish a feature) cost roughly 153,000 average DAUs per day during the gap plus 3,000 missed word-of-mouth users daily. The operational rules: call experiments done as soon as data is sufficient, ramp rollout fast, port wins immediately, prioritize high-ROI work first because it compounds longest.

The strategic foundation is that Duolingo users consciously want to build a habit — unlike Instagram or games. That insight concentrated investment in notifications and the streak system. Notifications are treated as a golden goose whose channel must be protected: Groupon's trajectory (positive experiment after positive experiment → mass email rot) is the cautionary case. Duolingo tracks efficiency ratios (Streak Saver: 1 DAU per 3.6 sends; XP Happy Hour: 1 per 130, killed), monitors unsubscribe rates, gives Android per-channel controls, and cuts lapsed-user reminders at the "elbow" (seven days) rather than maximizing short-term returns.

For mechanics, copy first then innovate. The fourth leaderboard iteration — modeled on Gardenscapes and Golf Clash — lifted D1 retention 1%, D7 by 2%, D14 by 3%, and time-on-app 17%. Innovation is reserved for areas where Duolingo already leads or after a complete MVP cycle. Finally, prioritize by total affected users: a 0.1% improvement across all notification recipients outweighs a 30% gain on Korean Stories users by roughly 50x.

How to accelerate growth by focusing on the features you already have

Most product growth comes from optimizing existing features, not launching new ones. Ken Rudin — growth lead at Facebook, Google, Zynga, and ThoughtSpot — argues that when 80% of a feature works and 20% doesn't, many users get 0% of the value; fixing that 20% costs less and hits harder than a new build. At Google Search, only 15% of users knew about high-frequency query types (sports scores, weather). Raising awareness of what already existed nearly doubled engagement and generated millions in incremental ad revenue.

The ARIA framework structures this work. Analyze: run correlation analysis to find which features most predict retention or monetization, then measure adoption, step-completion, and success rates to identify high-impact features with low uptake. At Roblox, 20% of Avatar Shop visits ended with no avatar change — adding sort and search lifted success rate and improved new-user retention. Reduce: cut steps, prefill smart defaults, swap typing for tap-to-select, and give users editable templates rather than blank canvases. Introduce: surface features in context — when users need them, not in blanket onboarding — and explain the concrete benefit. Assist: fill empty states with guidance, offer templates that embed expertise, and write error messages that explain the fix.

ARIA is iterative — run all four steps repeatedly on each key feature.

Unorthodox frameworks for growing your product, career, and impact | Bangaly Kaba (YouTube, Instagram, Facebook, Instacart)

Career impact is environment times skills. Bangaly Kaba (Facebook growth PM, Head of Growth at Instagram, VP Product at Instacart, Director PM at YouTube) built this after feeling stuck at Facebook: score six environmental variables (manager, resources, scope, team, compensation, culture) on 0–2. Manager outweighs the rest — a good one can move nearly all the others. On skills, communication beats execution; poor executors who tell a great story keep rising. For mentors: describe a challenge to a contact, get introduced to someone who solved it; several on rotating Fridays beats relying on one.

The principle Kaba carries across companies is "understand work." The anti-pattern is identify-justify-execute: marshal data for an idea, build, ship to flat metrics. The fix is understand-identify-execute: reserve roadmap capacity to answer unknowns first. At YouTube he started at 60/40 understand-execute; teams reached 85/15 over time. At Instagram in 2016 the onboarding funnel had logging only at top and bottom — first quarter was instrumentation plus cheap parallel tests; win rate hit 60–70% across 12–20 experiments per team per quarter.

Instagram's growth ran on three compounding loops beyond virality: celebrity partnerships drove media coverage; a web launch (George Wang's idea) lifted growth 10% immediately and created SEO pages; news article embeds amplified both. The 2017 connections pivot was decisive — early Instagram recommended celebrities to all new users, who posted into silence and churned around month eight. Switching to friend-to-friend connections doubled retention over 18 months.

For culture change, Kaba uses the "managing complex change" matrix: vision, skills, incentives, resources, action plan — each gap produces a distinct failure mode; action plans are easiest first. His Instacart failure: arriving with a systems-and-process vision when the company needed tactical execution. Lesson: talk to people who have left — they give the unfiltered account you need to triangulate.

The surprising truth about what closes deals: Insights from 2.5m sales conversations | Matt Dixon (author of The Challenger Sale and The JOLT Effect)

Most B2B sales deals die not to competitors but to indecision — 40–60% of qualified pipeline closes as "no decision." The conventional response, dialing up FOMO (missed benefits, competitors pulling ahead, end-of-quarter discounts), backfires 87% of the time. The reason: buyers aren't afraid of missing out; they're afraid of messing up. This omission bias, grounded in Kahneman and Tversky's loss-aversion research, makes commission-of-error losses feel far worse than losses from inaction. Only 44% of no-decision losses are genuine status-quo preference; the other 56% are buyers who want to buy but can't — paralyzed by fear of personal blame.

Three fears drive this FOMU: configuration anxiety (wrong options, wrong contract terms?), post-signature surprise risk (a Gartner report or competitor move makes the decision look bad afterward — one startup lost a seven-figure deal two weeks after closing when a new Magic Quadrant ranked them poorly), and ROI execution risk (the buyer's name is on projected returns that may not land). Large incumbents amplify all three: more options, more public scrutiny, higher spend.

The JOLT method addresses this. J — Judge indecision: use "pings" — soft hypotheses about the specific anxiety you suspect — and listen for the echo. Buyers self-report as decisive; 87% of those in the dataset showed moderate-to-high indecision. O — Offer a recommendation: too many options cause paralysis; prescribing a specific path creates the "delegation effect," where the buyer shares blame for a bad outcome with the recommender. L — Limit exploration: volunteer your product's genuine weaknesses early and demonstrate personal expertise rather than hiding behind specialists — this shifts buyers from wanting to be experts to trusting you as theirs. T — Take risk off the table: reset ROI expectations downward ("build on 5X, not 10X"), present an implementation roadmap before signing, and frame professional services as an insurance policy.

The Challenger Sale addresses the earlier problem: getting buyers to want to move at all. Rather than asking what keeps them up at night, challengers teach buyers what should. Dentsply failed selling a cordless dental wand at 3× the price until it reframed its pitch around hygienist injury rates and workers' comp costs — problems only its product solved. The discipline: identify what only your company does, then build backward to what must be true for a buyer to pay a premium for it.

When and how to run a billboard campaign

Out-of-home advertising works for B2B SaaS primarily as a brand awareness lever, not a pipeline generator — and the ROI logic holds even when attribution is murky. Stytch's 2024 San Francisco campaign (80 bus ads, 130 transit shelter ads, 8 street billboards, 3 on Highway 101) produced a 25%+ uplift in branded search traffic, a 70%+ increase in branded search clicks globally, a sustained 10–15% lift after the campaign ended, and explicit billboard attribution from 10%+ of subsequent sales opportunities.

The decisive variable is frequency × resonance. A single placement can work if it goes viral — Vanta's SOC 2 billboard hit 1,000+ Reddit upvotes and landed their biggest enterprise customer; Bland's billboard demo video reached 25 million views by pairing a provocative "Still hiring humans?" message with a sandwiched product demo. Without virality, two-to-seven daily impressions across mixed formats (billboards for recall, transit/street furniture for frequency) is the reliable formula.

Execution specifics: plan six months out to secure good inventory; Highway 101 billboards run $20K–$80K/month; cap OOH at roughly 20% of marketing budget (Writer's rule); time around developer conferences; amplify every placement on social immediately; buy branded paid search to intercept people who search after seeing the ad.

Ethan Smith: The power of internal linking for SEO

Internal linking is the highest-ROI SEO quick win — over 90% of sites underuse it. Google discovers pages through links; poor link spread leaves pages unindexed. Fix: audit with Screaming Frog, identify high-traffic "crawl points," connect related content to them using 5–10 contextual links per page with descriptive anchor text. Editorial (human-written) content now outperforms programmatic SEO; use AI for topic research and structure, but not writing — factual errors harm rankings and users.

Summary: April Dunford on product positioning, segmentation, and optimizing your sales process

Positioning means being the best in the world at delivering specific value to a well-defined set of customers — and most "positioning problems" are actually internal misalignment. The five-step framework: identify competitive alternatives (including "no decision" — 40% of B2B losses), isolate unique attributes, translate to differentiated value, find customers who care most, then choose the market frame that makes value obvious. Firmographic segmentation precedes personas; in enterprise deals, the internal champion who builds consensus matters more than the final check-writer.

How to consistently go viral: Nikita Bier’s playbook for winning at consumer apps (co-founder of TBH, Gas, advisor, investor)

Viral consumer growth is reproducible as science, but lasting social networks are black swans — one per decade. Nikita Bier built tbh and Gas, two apps that hit number one in the US App Store with three-to-four-person teams and no paid acquisition, selling them to Facebook and Discord.

The core method is searching for latent demand — people contorting themselves through a broken proxy to get some value. TBH's signal was Sarahah, an Arabic-language app at number one in the US because Americans were using it to fish for compliments. Replacing open-ended anonymous messages (which produced bullying) with positive-only polls ("who has the best smile?") guaranteed the compliment and drove 60 messages per user on day one versus the typical three or four.

Testing discipline mattered more than the idea. Bier seeded each app into a single school — following students on Instagram, running targeted ads — not as a scalable strategy but to eliminate confounding variables and get clean signal on the core loop. Live chat ran 24/7, with a rep pasting insights into Slack.

Gas required rebuilding the entire growth flywheel: iOS regulatory changes killed server-side SMS invites, so each layer — will people send messages, spread within a school, hop schools, pay to reveal who polled them? — had to be solved fresh. The app reached $11 million in revenue and 10 million downloads on cloud credits with no investors.

Key findings: teen invite rates drop 20% per year of age from 13 to 18; adults stop adopting new apps at 22; naming and icon gender affect invite rates (renaming "Crush" with a pink icon to "Gas" with a black flame lifted male invitations). Time to value must be under three seconds. iOS 18's granular contact permissions will entrench incumbents by killing contact-graph growth for new apps.

Rethinking SEO in the age of AI | Eli Schwartz (SEO advisor, author)

AI Overviews are absorbing top-of-funnel search — the informational queries where someone asks "best beach vacation within two hours" or "top CRM tools" — and delivering a paragraph answer without requiring a click. That kills the long-form content game SEO used to reward. The real opportunity shifts to mid-funnel: queries where a user already has direction and is looking for a specific solution. Tinder's local pages ("online dating in Dubai") still work because they solve an active problem — loneliness in a specific city — that no summary can resolve.

The correct frame is to treat SEO as a product question, not a marketing task. A PM's job is to identify what a self-discovery user would actually search for, then build something that converts that user. Most B2B SaaS shouldn't do SEO at all: Mixpanel and Google Cloud are examples where no one Googles their way into a purchase — committees and demos close those deals. Zapier worked because users already searched "Gmail + Salesforce integration" and the page resolved that search with a usable outcome. Tinder worked because programmatic city pages connected to an immediate solvable need.

SEO is not free. A typical agency is $10K/month; add CMS, engineering, and design and you approach $1M/year. The right test is whether that beats the same budget in paid or trade-show spend for your conversion path.

Key myths: guest-post link-building on low-authority sites is useless — real links are a brand byproduct. Traffic alone is a worthless KPI for non-media companies. Keyword tools (Semrush, Ahrefs, Similarweb) can be off by 10x; use them for relative comparison only — a top-down TAM approach is more reliable for sizing opportunity.

Google holds 98% of mobile search. Default agreements with Apple and Chrome make it structurally insurmountable for ChatGPT or Perplexity regardless of quality.

Meta's Head of Product (and 29th employee) on working with Mark Zuckerberg, early growth tactics, why PMs are like conductors, and more | Naomi Gleit

Naomi Gleit, Facebook's 29th employee and Meta's longest-serving executive after Zuckerberg, built her career by showing up uninvited to the office repeatedly until a marketing role opened, then doing PM work voluntarily after hours until formalized — earning a standing ovation from engineering when she crossed over.

On Facebook's early growth team (with Alex Schultz and Javier Olavon), the core insight was that churn and resurrection lines dwarfed new-user acquisition, making retention the real lever. The seven-friends-in-ten-days activation metric mattered less for statistical precision than for giving everyone one shared goal. In January 2009 the team paused all roadmap work to instrument every registration step, finding 20% of users never confirmed their email; the fix was letting any notification click count as confirmation. The broader principle: apply product-and-engineering rigor to marketing-owned problems — including community-powered translation now covering 100-plus languages.

Gleit's "Naomi-isms" center on extreme clarity and canonical everything. Extreme clarity means shared understanding of facts, not agreement — achieved via numbered (never bulleted) lists, canonical nomenclature defining every term (she caught a team conflating "consistency" and "accuracy" for content reviewers, which are different things), and live visual editing during meetings so everyone sees the same agreed outcome. Canonical everything means one doc per project with named work streams, single-threaded owners, canonical meeting cadence, and canonical chat, linking outward to sub-docs.

On meetings: pre-read 24 hours ahead; decisions as three options plus a recommendation evaluated via red/yellow/green traffic-light table (rows = options, columns = criteria); notes reply-all within 24 hours after.

On Zuckerberg: a "learn-it-all, not know-it-all." His public persona is the private Mark she has always known. Small Group (his leadership team) is defined by disagreeable givers — long tenure, mission-driven, willing to push back — meeting in one open strategic session and one structured operational review per week.

Breaking the rules of growth: Why Shopify bans KPIs, optimizes for churn, prioritizes intuition, and builds toward a 100-year vision | Archie Abrams (VP Product, Head of Growth at Shopify)

Shopify deliberately optimizes for churn because its revenue model inverts normal SaaS logic. Most income comes from payment processing tied to merchant GMV, not subscriptions — so a handful of breakout merchants per cohort make the entire acquisition wave profitable. The growth team's north star is total cohort GMV over three-to-five years — power-law winners justify the losses.

That long horizon reshapes measurement. Every experiment runs a 5% global holdout; new-merchant cohorts stay tracked after the winner ships with auto-pings at 3, 6, 9, and 12 months. About 30–40% of experiments showing short-term lift produce no incremental GMV a year later. Payment-failure dunning campaigns are a named example: merchants who let payments lapse weren't committed, so long-term GMV was flat. Contrast: pre-filling the online store editor with default sections showed neutral activation lift at launch but meaningful GMV gains six months later — better stores converted visitors faster. Neutral-result experiments are shipped when intuition favors them.

The anti-pattern this targets: teams with conversion-rate KPIs quietly constrict the funnel step above them to inflate their metric. Shopify counters by tracking absolute counts reaching each stage, never rates. Monetary friction experiments — trial length, credits, pricing — unlock entrepreneurs who would quit before gaining traction.

Structurally, product splits into Core (100-year vision, no KPIs, taste-driven), Merchant Services (medium-term), and Growth (~600 people). Core requires a group-lead "okay-to" before anything ships — head-of-core Glen runs founder-mode discipline without being the founder. Growth can touch any surface but must meet Core's quality bar; the recurring tension is the no-wizard principle — Core insists simplified onboarding live inside the real product, not a separate carousel.

At Udemy, 90–99% discounts worked because buying a course satisfies an emotional job — feeling like you're making progress — independent of completion, driving high repurchase rates.

How GiveDirectly increased donations by over $3 million/year through experimentation

Removing friction and fixing defaults generated $3M in incremental annual donations for GiveDirectly after one year of checkout optimization. Adding PayPal, Apple Pay, and Google Pay was the biggest win at $1.3M/year (a 14% lift); ACH direct debit also cut fees by $50K/year. Rewriting email opt-in copy to be explicit about frequency doubled subscription rates from 25% to 55%, worth $400K/year. Setting monthly giving as the default — rather than using a modal that depressed one-time conversion — added $500K/year. A polished homepage with an embedded donation form lifted conversion from 1.98% to 2.67%, worth $700K/year.

The one failure: donor-recipient matching (named individual, quarterly field-collected updates) showed no conversion uplift (p = 0.69) and only a short-lived retention boost. Switching to village-level matching required no engineering and produced 58% higher email click-through rates than country-level messaging.

How a great founder becomes a great CEO | Jonathan Lowenhar (co-founder of Enjoy The Work)

Being a founder is a state of being — instinct, grit, courage. Being a CEO is a craft. Investors describe great founders with words like tenacity; great CEOs with skills — hiring, planning, financial management — and nobody teaches founders those skills on the job. "Founder mode" gave founders permission not to learn the craft; Enjoy The Work closes that gap.

Six failure archetypes recur: robot (bans emotion while demanding passion), pleaser (avoids hard calls), perfectionist (beautiful product delivered minutes before bankruptcy), angry CEO (talent exits when options appear), laissez-faire (great hires, no direction), and ready-fire-aim — the most common. Fix: plan backward from one north star — fundraise, exit, profitability, or wind-down — quantify what must be true, and shorten feedback loops.

On exits, the Magic Box Paradigm (from Ezra Roizen) replaces the banker-auction: never post a for-sale sign. Find a champion inside a likely acquirer who falls in love with the fantasy of combining your tech with their distribution. Three stages: learn the fantasy, prove it cheaply (champions want to say yes), quantify with a future-value model. Instagram sold for $1B at zero revenue on Facebook's fantasy of future ad expansion. With Corp Dev, move every negotiation async.

On hiring: define 12-month success first, find people who have done that thing, and prefer candidates pulled between roles by former fans. Distinguish architect (builds first playbook), optimizer (scales a team), and scaler (finds leverage) — each fails in the other role.

Go-to-market has four buckets: ICP with kill criteria, positioning against alternatives, demand-gen experiments ranked by impact-vs-effort, and a sales playbook. Start with a white-hot center; adjacencies come later.

When facing existential decisions, the quiet inner voice accessed through stillness is more reliable than the fear-driven lizard brain. Distinguishing intuition from reaction is the meta-skill no framework replaces.

Growth tactics from OpenAI and Stripe’s first marketer | Krithika Shankarraman

Copying what worked at Stripe or OpenAI is the wrong move — those outcomes were products of a specific competitive moment that no longer exists. Be diagnostic instead: figure out where your funnel leaks before choosing what marketing help you need. High close rates but low volume means fund demand gen. Lots of interest but weak conversion means a positioning problem — a product marketer beats a demand gen hire.

Krithika Shankarraman's four-step DATE framework codifies this: Diagnose the actual problem; Analyze competitors to find gaps rather than to copy them; Take a differentiated path; Experiment and discard what doesn't scale. At Retool, paid social generated vanity impressions but no pipeline, so she doubled down on customer storytelling — enterprises like Netflix using the product were proof no copycat could replicate. At Stripe, the threat to Connect was companies becoming payment facilitators themselves; the response was a reverse-RFP content piece explaining the painful DIY process and positioning Connect as the relief valve, winning SEO without claiming to be a PayFac.

At OpenAI, awareness was never the problem. The gap was use-case clarity: people knew the name but not what to do with it. Marketing's job became engineering "use-case epiphany" moments. When ChatGPT Enterprise launched its contact-sales form, lead volume 40X-ed overnight — she built the first lead-scoring model herself in Python.

Brand and velocity aren't in tension. Brand is every touchpoint — product, support, recruiting — and clarity about what you stand for makes organizations faster, not slower. Strong brands mean new launches inherit trust; weak ones start from zero.

The T-shaped CMO (deep in one pillar) is giving way to a comb-shaped one — analytical, creative, fluent across product marketing, demand gen, and brand, using AI to close gaps. Taste and genuine product conviction stay non-automatable.

Naming expert shares the process behind creating billion-dollar brand names like Azure, Vercel, Windsurf, Sonos, Blackberry, and Impossible Burger | David Placek (Lexicon Branding)

Brand naming is a discipline, not intuition — and companies almost never recognize the right name on sight. David Placek of Lexicon Branding (Pentium, Blackberry, Sonos, Azure, Vercel, Windsurf) argues that comfort is the enemy: unanimous team approval usually signals a name too safe to win. Andy Grove chose Pentium over "ProChip" because internal polarization told him there was energy in the word.

The process runs in three stages — identify, invent, implement. Identify starts from behavior, not positioning: how the company acts now, how it wants to act, how the marketplace should respond. Invent uses small teams of two, never large brainstorms. Three teams each get a different brief: one gets the real assignment, one thinks they're naming a competitor's product, one names something in a different category entirely. Off-brief teams produce most winning names because constraint removal frees range. Windsurf came from a team listing metaphors for flow and dynamics, not anyone trying to name an AI IDE.

Forty years of linguistic investment produced proprietary sound-symbolism data: V is the most alive letter in English (Corvette, Vercel); B signals reliability (Blackberry's double-B was deliberate); Z and S carry noise (Sonos, Azure). Compound names multiply associations — Windsurf, Powerbook, Blackberry each let 1+1 equal 3. Vercel layers "ver" (veritas, verde) with "cel" (accelerate) for processing fluency the brain leans into.

For founders without resources: draw a four-point diamond. Top — define winning. Right — what you have to win. Bottom — what you need. Left — what you have to say. Generate 1,000–1,500 names without evaluating, then speculate on what each could become. Test by presenting a name as a competitor's launch — if people say "they're not like the other guys," it's working. .com is an area code; get the right name first.

'Sell the alpha, not the feature': The enterprise sales playbook for $1M to $10M ARR | Jen Abel

Enterprise sales at the $1M–$10M ARR stage is a learning exercise before a revenue exercise. Jen Abel of Jellyfish argues founders hold three advantages no hired rep can match: they carry the vision, command attention as the org's highest-ranking contact, and catch "budding moments" — signals that reshape product-market fit — that a salesperson filters out.

The outreach formula: relevancy (why this person, this role, right now), a counterintuitive claim rather than "we're better than X," focus on the unsolved problem not the solution, and brevity that fits a phone screen. Abel's opener — "zero-to-one sales talent doesn't exist" — illustrates the hook. Cold-calling returns higher response rates than email.

On the first call, lead with vulnerability: you're early, still learning, want their read on the problem. Pretending completeness produces polite feedback; admitting otherwise produces honest feedback. The goal is a second call booked before the first ends — a prospect who won't give calendar time on the spot is signaling disinterest.

Roughly 40–50% of enterprise deals require a service contract before a technology contract — 90-day engagements educating buyers who lack any process to adopt the product. This gets the logo, shows intent, and positions the startup to shape buyer thinking before a competitor does.

Procurement is its own sub-sales cycle: know the signatory before the contract reaches them; fill out their paperwork yourself; simplify vendor categorization so legal doesn't default to high-risk; split the deal into service and technology contracts to incentivize IT due diligence. Cycles run six to twelve months, but a $60K initial deal can reach $280K within four months once inside.

The most common failure point is not closing — it is qualification. Every apparent bottom-of-funnel problem traces back to reaching the wrong people with the wrong message about the wrong problem.

Ecosystem is the next big growth channel

When AI commoditizes product speed and content output, the growth differentiator becomes ecosystem: reaching audiences through intermediaries who already own their trust. Every traditional B2B channel is degrading — LLMs cut search traffic, AI SDRs flood inboxes, endless new tools kill virality, events face fatigue.

The mechanism is a flywheel: partners create and distribute content, you amplify it, credibility compounds. Supabase grew from 1M to 4.5M developers in under a year via open-source community plus integration partners. Gamma attributes over half its growth to micro-influencers paid on virality bonuses. ElevenLabs hit $100M ARR with 50 employees by paying Voice Library creators over $1M, spurring viral "AI presidents" content.

Implementation requires senior buy-in — product must build integration features, not just marketing. Map partners beyond the obvious (Clio uses state bar associations). Filter by coverage (TAM reach) and composition (audience overlap). Start narrow: Arrows focused on HubSpot, grew 40×, then expanded. Deals must win for you, the partner, and the customer.

What world-class GTM looks like in 2026 | Jeanne DeWitt Grosser (Vercel, Stripe, Google)

When AI floods a market with near-identical products, the buying experience becomes the differentiator. Jeanne DeWitt Grosser, COO at Vercel and former Stripe sales chief, argues that 80% of enterprise buyers act to reduce risk rather than capture upside — frame the sale around what happens if they don't change.

The biggest structural shift she names is the GTM engineer. At Stripe in 2017, Grosser attempted "Project Rosland" — a company-universe database feeding customized outbound templates — but the AI didn't exist. Vercel rebuilt it in six weeks, one engineer at 25% time. The agent handles inbound as well as ten SDRs did; nine freed SDRs moved to outbound. Cost: ~$1,000/year versus $1M+ in salaries. The method: shadow the top performer, encode their workflow into an agent, keep a human to QA responses, and hold conversion flat before cutting headcount.

The same team built a "deal-bott" on Gong transcripts. Vercel's largest Q2 loss — logged as "lost on price" — actually failed because no economic buyer was reached. The bot now pushes risk signals into per-customer Slack channels; the team burns down coaching gaps weekly like engineering bugs.

On segmentation, size alone is insufficient. Stripe layered growth rate and business model onto the size axis (marketplace → Connect; B2B → billing). Vercel adds Google Crux traffic rank — OpenAI is mid-market by headcount but top-25 by traffic, triggering enterprise treatment — plus workload type, since e-commerce and crypto need different language.

GTM-as-product means every touchpoint should feel designed. At Stripe, the first post-qualification call was a whiteboarding session; prospects drew their payments architecture, left with an asset, and Stripe learned the deal.

PLG has a ceiling; no company reaches $100B on self-serve alone. Starting outbound too late is the most common mistake — 12–18 months to become a predictable engine.

Elena Verna 4.0

Lovable — a vibe-coding tool that hit $200M ARR in under a year with 100 employees — forces a rethinking of the growth playbook. Elena Verna, head of growth, estimates only 30–40% of what she learned at Miro, Dropbox, and Amplitude still applies; she now spends 95% innovating new growth loops and 5% on optimization, the reverse of every prior role.

What moves the needle: building in public (engineers announce their own ships, bypassing marketing), influencer marketing that outperforms paid social 10x, and giving product away aggressively — credits flow freely to hackathons and any user who wants to run a Lovable event. LLM costs get reclassified as marketing spend, not margin to defend. Retention is on par with B2B SaaS benchmarks; NDR is strong because builders keep buying credits. Activation gets almost no dedicated effort because the agent team already treats first-generation quality as the product.

The internal standard is "minimum lovable product," not minimum viable. Specs arrive paired with a Lovable prototype. All brand work lives inside product interactions; there is no brand marketing team.

On product-market fit: the cycle has collapsed from years to roughly three months. Each new foundation model raises the capability ceiling and consumer expectations shift just as fast — a permanent treadmill of scale, recapture fit, scale again. OpenAI losing 6+ percent share in a week during Gemini 3's launch illustrates the stakes. A parallel concern: women make up only about 20% of Lovable users despite the approachable branding, with adoption data showing a widening gender gap across vibe-coding. SheBuilds — women-only hackathons with free 48-hour access — is one response.

On hiring: Lovable screens for passion-as-hobby over credentials, high agency, and comfort with chaos. Failed startup founders and AI-native new graduates are the most valued profiles; paid work trials replace standard interviews.

A guide to advanced B2B positioning

Most weak B2B positioning traces back to one of four fixable breakdowns, diagnosed by April Dunford across 300 companies.

The first is cross-functional disagreement about who the real competition is. Marketing inflates competitors with high ad spend; product adds horizon threats never seen in real deals; sales rarely counts the status quo as a competitor. The fix: reframe as "If we didn't exist, what would a prospect do?" In B2B, roughly half of lost deals go to no-decision, and Matt Dixon's JOLT Effect research shows about half of those are buyer indecision — ignoring the status quo leaves the largest loss category unaddressed.

The second is product pessimism. PMs pulled into bad-fit deals develop a skewed view of what the product can do; clean wins that closed without PM help go unexamined. Teams maintain overlong competitor lists and dismiss wins as buyer error. Run positioning around strengths, include experienced AEs, and choose a moderator willing to challenge pessimistic claims.

The third is poorly defined differentiated value. Teams either stop too early (features, not outcomes) or abstract too far ("saves money," which any competitor claims). The right level stays specific: a database running queries 1,000x faster only resonated when reframed as customers answering support tickets 70% faster. Objection material — deployment ease, TCO, change management — belongs in the sales process, not the positioning statement.

The fourth is not knowing what entity is being positioned. Single-product companies have no real separation between company and product positioning. Multi-product companies must settle go-to-market strategy first, then choose a model: cascading umbrella (products sold separately), lead-product-plus-add-ons (Salesforce's CRM-first approach), platform or family bundle, or segment-specific positioning with possible separate legal entities.

Snapchat CEO: Why distribution has become the most important moat | Evan Spiegel

Distribution, not product-market fit, is the decisive differentiator in consumer technology — and most founders underinvest in it. Snapchat succeeded in 2011 partly because mobile and the App Store created a rare open moment when users eagerly downloaded new apps; that window has closed. TikTok bootstrapped distribution by spending billions to subsidize both viewers and creators simultaneously. Threads leveraged Meta's existing graph. Today, new entrants without an inherited distribution surface face near-impossible odds.

Software features are not a moat — Snap learned this 15 years ago when every innovation (Stories, face filters, lenses, swipe navigation) was cloned by larger platforms. The response: build ecosystems and platforms that cannot be simply copied. Snap's AR lens platform has millions of developer-created lenses; its creator relationships constitute a network that takes years to replicate. Hardware (Spectacles/Specs) extends this logic into a domain where vertical integration makes copying structurally expensive.

On product development: Snap runs a 9–12 person design team that operates flat, non-hierarchical, and at high velocity — hundreds of ideas reviewed weekly. The governing principle is "if you want a good idea, you need lots of ideas." Preciousness around any single concept is trained out from day one. Design functions as an intentional bottleneck on shipping, enforcing product coherence across an app used by nearly a billion people. Stories emerged not from a feature request — users asked for a "send all" button — but from listening to underlying frustrations: social pressure, permanent records, reverse-chronological feeds. The team built something those frustrations implied but no user had articulated.

The contrarian view Spiegel offers: humanity's pace of adoption will constrain AI deployment more than technical limits. Societal pushback on AI changes is being systematically underestimated by technology leaders who assume adoption follows capability automatically.

This Week #3: Finding product/market fit, interviewing for a Director of PM role, and structuring discussions with senior leaders

Early PMF signals are passion (customers visibly excited, asking when it ships), skin in the game (willing to pay now, even for consumer apps), and a clear "why" uncovered by repeated "why" probing. For Director-PM interviews, the five decisive dimensions are long-term strategic thinking, people leadership (per Andy Grove, your output is your team's output), stakeholder management, impact at scale, and communication. For senior-leadership check-ins, the McKinsey SCR (Situation-Complication-Resolution) framework structures any medium; choice of deck vs. doc follows audience norms and your own style.

How to increase virality

Only invest in virality if it's your primary growth engine — most products grow mainly through one of four engines (virality, performance marketing, content, or sales), and optimizing a secondary engine early is wasted effort. Virality fits when the product is better with friends (Slack, Snapchat), inherently shareable, or genuinely remarkable. K-factor (invites × conversion rate) has three levers: sharing rate, conversion, and engaged-user count. Six strategies raise sharing: nurture word-of-mouth through remarkable experiences; incentivize referrals (Dropbox storage, Pinduoduo discounts); make the product better with more users (Figma, WhatsApp); make it only functional with others (Calendly, Zoom); generate shareable content pointing back (TikTok watermarks, Spotify Wrapped); and increase experiential serendipity (Hotmail signatures). Growing engaged users — Facebook's People You May Know, TikTok Duets — adds more fuel to the cycle.

How to interview product managers

Hiring one strong PM requires talking to roughly 23 candidates — only 13% of inbound applicants clear the first screen, forcing heavy reliance on sourcing and referrals. The recommended funnel: recruiter screen (50% pass), hiring-manager screen (40–50% pass), a full day combining a project with three to six skill-based 1:1s, blind-vote panel, references, offer.

For IC PMs, the six skills to test are communication, collaboration, execution, strategy, impact, and product sense. Keep questions and scoring criteria consistent to reduce bias. Sample questions: "walk me through a project you're proud of that took 3–9 months" (execution); "what's the most impactful thing you shipped, and would it have been as impactful without you?" (impact).

The project is the highest-signal component — Merci Grace at Slack found project performance correlated with multi-year success. Two formats: take-home before the day (more thinking time, inequitable for candidates with family constraints) vs. 90-minute day-of (equalizes conditions). Real prompts include Chime's rogue-account identity scenario, Slack's new-team activation problem, and Airbnb's live stat that hosts ignore or reject 50% of booking requests. Evaluators look at how candidates decompose the problem; the discussion afterward is often more revealing than the deliverable.

To close, treat the offer like an investor pitch: Runway's Siqi Chen runs the same deck for candidates as for Series A investors.

The role of AI in product development | Ryan J. Salva (VP of Product at GitHub, Copilot)

GitHub Copilot's defining insight is that programming languages are constrained vocabularies — fewer words than English, cleaner grammar — making them well-suited for LLM training. The product originated when OpenAI hammered GitHub's servers cloning public repos to scrape training data; GitHub handed over the Arctic Code Vault dataset instead, and the collaboration produced CodeX, the model underlying Copilot.

The UX crystallized around keeping developers in flow. Early experiments tried side-panel suggestions, but switching panels broke concentration. The team landed on inline gray-italic multi-line autocomplete, built with VS Code adding the extensibility point. Latency experiments set the optimal delay at ~200 milliseconds. Copilot now writes roughly 40% of Python code and 25–40% across other languages.

Incubation happened inside GitHub Next, a ring-fenced team for horizon-two and horizon-three work. The handoff to an operational EPD squad was triggered by signal — mind-blown emoji tweets and Hacker News threads during the technical preview — not a calendar date. Researchers seeded the first EPD squad and returned to Next only after replacement seats were fully onboarded. The product team, not the R&D team, must own the roadmap; outsourcing innovation to a research group permanently is a failure mode.

Scaling surfaced two unexpected challenges. GPU supply constraints made specialized training chips genuinely scarce. Community trust required sustained dialogue: training on public code raised questions about model poisoning, security vulnerabilities, and code ownership. Content filtering started with crude blocklists, then graduated to Azure's Responsible AI sentiment models that can read context rather than match strings.

Salva's position: Copilot augments, never replaces — developers stay in the loop, and existing safeguards (static analysis, unit tests) stay in place. Portfolio allocation for larger teams: 5–10% on audacious moonshots, ~25–30% on operations, ~60% on incremental improvement of in-market products. At startups, those ratios collapse; you are the moonshot.

AI and product management | Marily Nika (Meta, Google)

Every PM will eventually be an AI PM — every product will need personalization, recommendation, and automation built in, which means partnering with research scientists as core cross-functional teammates rather than peripheral specialists.

The critical mindset shift is comfort with uncertainty. Traditional PM work runs on a build-launch cadence. AI product work runs on: form hypothesis → train → wait → discover the results weren't what you expected. The shiny-object trap is the opposite failure: building AI because it's exciting rather than because a real user pain point and usable dataset both exist. The AI PM's defining job is to identify the right problem, then find a research scientist to model it — not start with a model and hunt for a use case.

Practical rules: never train a custom model for an MVP — fake the AI in a Figma prototype to validate demand first. Don't build proprietary models unless differentiated data is your moat; if every company trains on the same licensed dataset, quality converges. Data requirements vary wildly: 15–20 labeled images can classify a dog breed; voice recognition and NLP need thousands of hours. Deciding when accuracy is good enough to ship — 70%? 80%? — is the PM's call.

For getting started now: ChatGPT is useful for mission statements, user personas, and segment ideation — it surfaces user types intuition wouldn't reach. Google's AutoML lets non-coders train image-classification models by uploading labeled photos; a wind-turbine company cut inspection from three weeks to hours by running drone photos through it.

AI PM career trajectories don't advance by launch count — that expectation needs renegotiating with hiring managers upfront since research cycles are long and many bets fail. Staying current means reading arXiv preprints, not just product blogs — that's where ChatGPT's techniques lived for years before productization.

I built a Lenny chatbot using GPT-3. Here’s how to build your own.

Dan Shipper (Every) built a working chatbot on Lenny's full newsletter archive in a few hours using GPT-3, and explains how to replicate it without a coding background.

GPT-3 is a probability engine that completes prompts; ChatGPT's API isn't publicly accessible, so custom bots use GPT-3 directly. Personality and conversation history are injected by prepending a persona description and the full transcript to each new prompt.

The central problem is hallucination plus a 2021 knowledge cutoff. GPT-3 confidently said Substack's first publisher was Andrew Sullivan (wrong — it was Bill Bishop) and that founders solving their own problems is "absolutely" best practice (wrong — fewer than a third of top consumer startups did this). The fix is context-stuffing: paste the relevant source text into the prompt like an open-book notecard. It corrected both errors immediately.

The constraint is the ~4,000-token prompt window. The scalable solution uses embeddings via the open-source GPT Index library (Jerry Liu): articles are chunked and vectorized (at $0.0004 per 1,000 tokens); at query time the closest chunks are retrieved and injected into the prompt. A few lines of code complete the pipeline.

Any evergreen content archive — newsletters, books, company wikis, personal notes — can be made queryable this way, eliminating manual search and grounding answers in source material.

Product lessons from Waymo | Shweta Shrivastava (Waymo, Amazon, Cisco)

Waymo's design challenge is not autonomy but credibility — building a system that feels human while being measurably safer. The car trains on human driving data with bad behavior filtered out, producing motion riders normalize within five minutes. Body language is engineered: inching into a merge, decelerating on San Francisco downhills even when speed limits allow otherwise. Driving is social, governed by city-specific unspoken norms, and the models absorb all of it. In-car monitors showing what the vehicle perceives, plus reachable rider support, close the trust loop.

The L4/L5 distinction shapes the product strategy. L4 — fully autonomous within geofences, no human at the wheel — is live commercially in Phoenix (no waitlist) and San Francisco; L5 (any road, no map) is not the goal. Snow is the open technical gap. KPIs run on two tracks: commercial metrics (trips per week, WAU, cost per ride) and driver-performance metrics anchored to collisions per 100,000 miles against a human baseline, plus "stops and strands" — how often the car halted unnecessarily or needed remote rescue.

Three PM lessons transfer beyond Waymo. Amazon's PR/FAQ practice — write the launch press release before building — forces genuine value-proposition thinking. Knowing what you are not building matters as much as knowing what you are; products that serve every request lose coherence. Large-company complacency is the default failure mode: Shrivastava pushed past it at Amazon by launching Honeycode, a no-code platform outside AWS's infrastructure core, before competitors forced the move.

The tell that you are not listening is the absence of conflict. The "rule of seven" makes this concrete: after seven emails without resolution, stop threading and call.

Promotion advice: make your ambitions known to your manager, then focus entirely on business impact.

Lessons on building product sense, navigating AI, optimizing the first mile, and making it through the messy middle | Scott Belsky (Adobe, Behance)

Product sense begins with empathy, not passion for solutions. Passion locks you 30 degrees off from what customers need; empathy delivers the actual answer. The method is sitting shoulder-to-shoulder with customers observing their whole day — not just product use — capturing the context that raw data misses.

The "first mile" — the first 30 seconds with any product — is where most users are lost. In that window people are lazy, vain, and selfish: they want instant success, zero learning curve, and to look good to peers. Most teams spend their final build stretch only then thinking about this entry point; inverting that ratio separates strong product teams. The problem never goes away because each new cohort is less forgiving — Photoshop dropping from hundreds of dollars to $10/month forced a complete onboarding rethink for users with creative aspirations but no skills.

Belsky's counterintuitive lesson from Behance's 2008 launch: they shipped groups, a tip exchange, snapshot sharing, and full portfolios simultaneously — most complex at launch. Killing the tip exchange spiked project publishing; killing groups raised it further. When a product does one thing, everyone does that one thing and the core metric runs at 10x. The rule: optimize for the problems you want to have — users asking for more features — not problems you avoid by building everything upfront. Always ask what you can remove when you add something new.

On consumer durability: products without structural mechanics underneath are just R&D for platforms that already own distribution. Uber unlocked genuine excess capacity; Pinterest built around collecting interests rather than social anxiety and drove so much referral traffic that publisher sites added Pin buttons themselves.

On AI: generative tools expand the surface area a designer or PM can explore. A Hollywood director uses ChatGPT to generate five scenarios he won't use — the rejects clarify what he actually wants. AI collapses organizational stack, letting a PM answer data questions without routing through an analyst. ChatGPT mines the center of existing knowledge; finding the edges that become centers remains a human job.

The quit-or-continue test for the messy middle: ask whether you have more or less conviction now than at the start. More conviction despite repeated failure — you're in the messy middle, stay. Less conviction — quit; the fuel isn't there. In a resource-constrained era, resourcefulness compounds like muscle; resources spent like carbs leave nothing behind.

Gustav Söderström

The internet has moved through three eras — curation, recommendation, generation — and each transition required rethinking the entire product experience. Gustav Söderström, Spotify's co-president and CPO/CTO, treats generative AI as a fundamentally new paradigm, not an upgrade to recommendation-era ML.

Spotify's first purely generative product is the AI DJ: a digitized persona that speaks personalized commentary and sequences music for users with zero intent. Radio historically owned this use case; the DJ replicates it at scale. Guiding principle: do as little as possible and get out of the way.

A core ML design rule: fault-tolerant UIs must match actual algorithm performance. If hit rate is one-in-five, show five items so at least one is relevant. Midjourney's early interface — four low-res variants fast, not one slow result — illustrates this exactly. Single-play-button designs only work when accuracy approaches perfection.

On org structure, Söderström traces Spotify's retreat from the squad model: seven-person autonomous cells at scale produce a hundred strategies in a hundred directions. The fix was moving autonomy to the VP level — enough independent thinkers to avoid CEO bottlenecking, but senior enough for pattern recognition. Planning follows a 10% rule (Shishir Mehrotra's): two weeks per six-month execution period.

The homepage redesign illustrates the hardest lesson: feature launches are voluntary; redesigns are not. Spotify's home feed was ~90% recall, ~10% discovery; the redesign flipped it, and users couldn't find playlists. To separate "right change, habits disrupted" from "wrong change," compare new-user cohorts against existing ones. The fix preserved discovery as opt-in, protecting what Spotify does better than competitors: tracking multiple in-progress sessions from a single screen.

Books most recommended: 7 Powers (Helmer), The Complete Investor (Munger), plus physics titles including Sean Carroll's Something Deeply Hidden and Paul Davies's The Demon in the Machine.

The 10 traits of great PMs, how AI will impact your product, and Slack’s product development process | Noah Weiss (Slack, Foursquare, Google)

Great PMs are facilitators, not mini-CEOs — the most dangerous framing in the field. Noah Weiss (CPO at Slack, formerly Foursquare and Google) argues the job is to facilitate decision-making pace and quality, amplify the team, and drive impact. Early-career PMs should master execution and data fluency; senior PMs need clear writing — the only scalable way to influence a large org.

Working with product-minded founders requires shared principles as a common quality language, plus a U-curve of involvement: heavy at kickoff to lock strategy and anti-goals, then at the end to taste-test in real code rather than mocks.

On AI, Weiss draws on 15 years from Google's Knowledge Graph to Slack. The recurring failure mode: UIs that appear supremely confident while hallucinating. The fix is virtuous feedback loops — training data as a byproduct of normal usage, the way Netflix ratings improved recommendations automatically.

Slack's methods for staying close to users: Complaint Storms (walk a competitor's onboarding ruthlessly; starting on someone else's product calibrates the team before turning the lens inward); live threads during usability sessions so engineers react in real time; quarterly Customer Love Sprints — two-week cycles shipping small, high-delight fixes from a burndown list.

When self-service growth plateaued in 2019, the team scrapped the roadmap and spent two quarters purely learning. Two levers emerged: a trial letting free users taste the paid product, and a north-star metric — "successful teams" (five or more people using Slack most of the work week) — which predicted upgrade likelihood at 4x. Product-market fit is a stack of S-curves; each new segment requires rebuilding from scratch.

Slack's early growth wasn't a PLG playbook — the term didn't exist. It was building something people put social capital on the line to share with coworkers; enterprise accounts grew organically before sales arrived.

Relentless curiosity, radical accountability, and HubSpot’s winning growth formula | Christopher Miller (VP of Product, Growth and AI)

HubSpot's shift to product-led growth came not from a top-down mandate but from a small team taking radical ownership of problems nobody else claimed. When Miller's team noticed the self-service pricing page was dormant, they simply asked if they could have it, redesigned it around discoverability, value-prop clarity, and friction removal, and produced a step-function change in the funnel. That ownership mentality — treating every unaddressed problem as the team's problem — expanded their remit over time because they kept delivering.

The core PLG philosophy: give value before you extract value. Free software is engineered so small customers get sustainable value, but hitting its limits makes a paid upgrade obvious. HubSpot's flywheel runs attract → engage → delight → advocacy → more top-of-funnel, driven by peer referrals that dominate SMB purchasing. PLG doesn't mean eliminating humans — it means using the product as the primary growth vehicle, with humans as a backstop for moments (security reviews, data migrations) where self-service genuinely fails the customer.

Common PLG mistakes: hiring a growth head with no resources; expecting near-term returns from what is fundamentally R&D; bad data hygiene creating analyst bottlenecks; and giving up because the data set is small — ten customer conversations often surface the "why" that event-tracking never will.

Channel diversification matters. HubSpot's original engine was content SEO; the shift added a free product on top. With generative AI threatening pure-SEO funnels, HubSpot is investing in microapps — single-use tools (Website Grader, brand kit generator, email signature generator) that deliver standalone value and pull users toward the full platform. Covid was an unexpected accelerant: goodwill pricing drove a surge in free-and-starter growth.

PM traits Miller selects for: relentless curiosity, resilience (70–80% of growth experiments fail; grasping for wins produces bets too small to matter), coachability, and ambivalence toward solution complexity.

The future of AI in software development | Inbal Shani (CPO of GitHub)

AI won't eliminate software engineers — it shifts what they spend time on. Inbal Shani, CPO of GitHub, argues Copilot is a copilot, not a pilot: developers spend less than 25% of their day writing code; the rest goes to meetings, code reviews, legacy archaeology, and coordination. Giving developers 30 minutes back a day frees capacity for creative thinking, not headcount cuts.

Junior developers are being reshaped most sharply. Without AI they spend early years learning syntax. With AI handling that, they can engage with systems architecture and product thinking from day one — capabilities today concentrated in seniors.

GitHub's numbers: 1.5 million developers across 37,000 organizations; 55% faster code writing; 85% report higher confidence in code quality; code reviews complete 15% faster; 88% of Accenture's AI-suggested code was retained. Time alone is a weak success metric — fast bad code is still bad. The better frame is "time to value": task assignment to realized business outcome. Security AI gets measured differently — by secrets prevented from leaking, not lines produced.

The most underhyped AI application in software is testing. As AI generates more code faster, the gap between volume and test coverage widens. AI-generated suites — unit, load, security, penetration — could close it, but the industry isn't talking about it.

The biggest adoption mistake is asking "what should we do with AI?" instead of "what problem are we solving?" GitHub designed Copilot the same way: zero friction, zero prompting, no new workflow needed.

Shani predicts a hybrid future rather than monolithic LLMs. Safety-critical domains like aerospace and automotive need specialized tuned models. General-purpose LLMs and niche models will coexist.

GitHub Next, the research team that originated Copilot, avoids two failure modes: becoming a paper-writing group disconnected from production, and being hijacked for short-horizon work. It stays on a three-to-five-year horizon with tight feedback loops into product and engineering.

How to be more innovative | Sam Schillace (Microsoft deputy CTO, creator of Google Docs)

Disruptive ideas always look dumb at first — the more disruptive, the dumber. The diagnostic signal is polarization: products that split people into intense fans and people who want them dead are genuinely disruptive. Mild indifference signals incremental work.

Schillace built Writely — which became Google Docs — by noticing that `contentEditable` plus JavaScript could power a word processor in a browser. His co-founders said the browser couldn't support it. He built it in days, discovered collaboration was possible only because locking hadn't been implemented, then spent months on the three-way merge problem. The order mattered: seeing the value first created motivation to push through the hard engineering. Seeing the difficulty first would have killed it.

The core heuristic is separating why-not questions from what-if questions. Why-nots (no offline, browser can't do this, people won't trust the cloud) become engineering problems once the what-if is compelling enough. The canonical why-not for Docs was airplanes. His answer: connectivity on planes will come. It did.

On users: people are lazy, and adoption follows thermodynamics — a little more convenience produces slow growth, a lot more produces explosive growth. Total friction must be far less than the resulting ease. Writely launched with no email required; after two minutes they asked for one, no password.

On AI: generative AI is making pixels free the way the internet made distribution free. The shift — "your product is a feature of AI, not vice versa" — arrives when models handle planning, multimodal output, and personalized state, collapsing software interaction to communicating intent. His multi-agent work at Microsoft found one result: giving agents a shared whiteboard as working memory makes them measurably smarter.

Career: do the thing you feel guilty getting paid for, and do it hard. Motto: virtue from error — make something from your mistakes.

Inside OpenAI | Logan Kilpatrick (head of developer relations)

OpenAI moves fast because of two hiring filters: high agency and high urgency — people who see a customer problem and start building before anyone assigns it. The Assistants API illustrates this: engineers heard developers wanting higher-level abstractions, self-organized, and shipped it without a top-down directive.

On product strategy: OpenAI expands general capabilities and will build general-purpose agents, but won't go vertical — no AI sales agent, no legal product. Founders competing in the general-purpose assistant lane should expect direct competition unless radically differentiated. Founders going deep in a vertical (Harvey for law is the named example) occupy space OpenAI has no interest in, and keep benefiting from improving base models without funding their own R&D.

GPTs (custom ChatGPT builder, launched late 2023) are the vehicle toward the agent future. Today they're custom-prompt containers: upload files, add instructions, optionally connect external APIs. Zapier's integration — routing all 5,000 Zapier connections through a GPT without code — is the biggest non-developer unlock. The longer-term shift is from synchronous chat to asynchronous tasks: "go do this, let me know when done." Narrow vertical GPTs also serve as the on-ramp for users who find a blank ChatGPT too open-ended.

Prompt engineering reduces to one rule: context is everything. The model has no background on the user; generic input produces generic output. Already live in DALL-E: automatic prompt expansion where the system infers what you meant and shows you the upgraded prompt to edit.

Research team size stays intentionally small: in a GPU-constrained environment, a researcher who doesn't multiply everyone's throughput is a net loss. Planning uses H1/Q1 goals without OKRs; revenue is a proxy for compute budget, not a goal. Third-generation embeddings are five times cheaper, with large non-English gains — roughly 62,000 pages per dollar.

You should be playing with GPTs at work

Custom GPTs — task-specific ChatGPT instances trained on internal docs — are already delivering measurable gains across non-engineering teams. Italy's ChatGPT ban cut coder productivity 50%; Duolingo saw 25% faster developer velocity via GitHub Copilot; PwC projects $15.7 trillion in AI-driven economic value by 2030.

Building one takes minutes: describe the task in plain English, upload PDFs or spreadsheets, share the URL. Zapier integration lets GPTs take actions, not just answer.

Practitioners use them to: rewrite UI copy against brand guidelines (Ramp), let PMs "interview" customer persona documents (Chime), index all past user research for searchable Q&A, auto-generate SMART goals from company values, grade search relevance cheaper than human labelers (Faire), produce multilingual e-commerce copy at 95% lower cost (Nordiska Galleriet), and synthesize sales transcripts into feature ideas. The pattern across all cases: upload proprietary context once, share a URL with the team.

How AI will impact product management

Conventional PM wisdom has it backwards: AI will hit hardest at the high-value strategic skills, not soft execution work. Strategy, vision, goal-setting, and discovery — historically the most prestigious PM outputs — are exactly what AI is built for: ingesting vast data and surfacing non-obvious conclusions. AlphaGo's move 37 and Jensen Huang's claim that natural language replaces programming point the same direction.

The robot-scale rankings bear this out. Strategy, vision, and goal-setting (4 robots) face sharp displacement; spec-writing and discovery score 3. Roadmapping, design feedback, and most shipping/alignment work land at 1–2 — heavy people-work that resists automation longest.

The upshot: PMs should double down on influence, empathy, communication, and product sense — the conductor role that marshals people and AI alike — while learning to work with AI as leverage on strategy and data rather than abandoning those skills entirely.

Why not asking for what you want is holding you back | Kenneth Berger (exec coach, first PM at Slack)

Most professionals either hope others will read their minds or issue orders — both produce stuckness, conflict, and a career out of integrity with what you want. The fix: articulate, ask intentionally, accept the response.

Step one is hardest. People-pleasers suppress wants entirely; high-control types state demands too unrealistic to say out loud. The diagnostic: "dream behind the complaint" — every complaint implies a world where it's resolved; start there and test whether that future is genuinely inspiring. If the dream is too embarrassing to own, it's the wrong dream.

Step two means breaking your rut. For people-pleasers: say it without data — relationships carry weight PMs undervalue relative to A/B tests. "I disagree and I know it's not my call, but I want you to know that" often moves outcomes. For control types: insist on a hell yes — "What would it take to get to hell yes?" extracts real commitment instead of promises that evaporate by deadline.

Step three is an emotional-regulation problem. Most responses are no. The traps: over-accepting (this no means never) or under-accepting (they're wrong so it doesn't count). A no is this person, this moment, this approach — data on what to try next.

Berger illustrates all three with his year at Slack — fired three times by Stewart Butterfield: first from overconfidence (never articulated what success looked like), second from fear (people-pleaser through a year of one-on-ones without naming what he wanted), third by blaming the CEO rather than hearing the nos. The year was torture because he was out of integrity with himself throughout.

Closing: most high achievers believe fear of inadequacy drives performance. Berger rejects this. Discipline gets you to the gym for a week; sustainable motivation requires a genuine vision. Fear signals a tiger. There is no tiger.

Counterintuitive advice for building AI products

Most early AI apps suffer a "tourist" problem — high initial traction, shockingly low retention — because demo value isn't user value. The core discipline is unchanged: find a high-pain workflow, iterate closely with customers, hold a high bar for UX. What shifts is the starting question: ask "what's technologically possible?" before assuming buildability. Proprietary licensed data and superior interfaces matter more than models as models commoditize. Segment users by AI attitude (embracers vs. skeptics) rather than functional needs, or you'll average them into tepid-tea products. Label features "AI-powered" — it improves both adoption and usage quality. Invisible wins beat big features: at Incident.io, 75% of incident summaries are now AI-generated by targeting the right high-frequency chore. Prompt engineering is underrated — teams repeatedly find they need better prompts, not new features. Time-in-app may fall as productivity improves. Pre-computed speed is decisive. Expect model capability jumps to lift your product automatically.

How close is AI to replacing product managers?

With expert prompting, AI beats or ties experienced human PMs on core product management tasks — and most voters prefer the AI output even when they correctly identify it as AI. That is the finding from a blind experiment by Lenny Rachitsky and prompt engineer Mike Taylor (co-author of O'Reilly's *Prompt Engineering for Generative AI*), testing GPT-4o against real human answers from Exponent's PM interview database.

Three tasks were tested. On product strategy (YouTube Music's one-year plan), AI won 55% of votes despite 77% recognizing it as AI; critics said it felt like a feature list rather than genuine strategy — the key remaining human edge. On KPI definition (DoorDash metrics), AI won decisively at 68–86% depending on ties; verbosity was the main tell. On ROI estimation (Meta Jobs feature launch), the human narrowly won at 58%, partly because it included numeric Fermi estimates — a gap chain-of-thought prompting largely closes.

The prompting method: find a real human example of the task, have the model extract structural instructions from it, then build a template with a role ("As a PM for a major tech company…"), a chain-of-thought "Thinking" section, and the example. Adding a single example measurably improved reliability; the strategy prompt also instructed the model to make minor grammatical errors and add obscure references, narrowing detectability. Prompting techniques account for 30–60% accuracy improvements on benchmarks — explaining why studies using basic prompts consistently understate AI capability.

The experiment points to a near-term split: structure, comprehensiveness, and metric frameworks are automatable at intern-to-mid level. Genuine strategic tradeoffs, tacit organizational context, and unexpected conceptual connections remain human territory. The next step maps AI capability against Lenny's full PM skill taxonomy — shape the product, ship the product, sync the people — to track what percentage of the role is currently automatable.

State of the product job market, part 2

The PM job market is in slow recovery: 5,752 open roles in late 2024, up from a 2023 trough of ~4,000 but below the 2022 peak of 10,000+. Senior and lead roles are growing share fastest; entry/mid still holds over a third. Remote PM postings fell from 35% to 22.5%. The Bay Area claims over 20% of PM roles, up 25% year-over-year; Bengaluru leads outside the US. SQL tops hard-skill mentions; LLMs are the fastest-rising. ML engineers (+79% YoY) and data engineers (+55%) are tech's hottest functions; diversity and scrum master roles have nearly vanished. AI companies hire almost exclusively engineers. Product owner titles absent at tech firms—common only at banks and telecoms.

Five proven prompt engineering techniques (and a few more-advanced tactics)

Generic AI output — emojis, corporate-speak, averaged mediocrity — comes from treating the model like a mind-reader. Specificity fixes it. Mike Taylor (O'Reilly prompt engineering book, 100,000-student Udemy course) distills eight techniques.

Role-playing: assign an expert identity to pull domain vocabulary and framing. Style unbundling: ask the model to list a person's style as bullet points, then feed those back as the spec — selective control rather than wholesale imitation. Emotion prompting: appending "this is very important for my career" measurably improves output quality per published research, though it can occasionally backfire. Few-shot learning: provide two or three format examples before the real request. Synthetic bootstrap: when real examples don't exist, generate ten diverse instances first, then use them as few-shot inputs.

Three advanced tactics for when single-prompt approaches fail: chain-of-thought ("think step-by-step") surfaces reasoning before you act on it; RAG pastes relevant documents into context so the model draws on current or proprietary data; LLM-as-judge has the model rate multiple generated versions against defined criteria — faster and cheaper than human review.

Shreyas Doshi Live

Product leaders stay chronically busy not because they lack productivity systems, but because they keep building the wrong things and accumulating feature debt — the real fix is a genuine product strategy, not a better to-do list.

Doshi frames a career retrospective around four questions he wished he'd asked sooner.

Why am I so busy? Efficiency tools hit a wall once scope outgrows any technique. At Stripe, a pre-aligned product strategy for Connect compressed a four-to-six-week planning cycle to three days. The subtler cause: PMs treat features as two-way doors when in practice they're one-way. Once a feature ships, misses adoption, and lands at a QBR, room dynamics force the team to extend it rather than kill it. Busywork accumulates as un-killed feature debt.

Do I actually have good taste? Doshi spent six years at Google internalizing execution over strategy, then watched Twitter circa 2014 struggle despite strong assets and diagnosed the problem as a strategy deficit. He rebuilt what taste means: not aesthetic pixel judgment, but the ability to evaluate ideas stripped of social proof. Three failure modes recur — excitement over catchy metaphors (two-way doors vs. "reversible decisions"), alliteration bias (fail fast vs. fail quickly), and deference to authority or complicated math. Real first-principles thinking means shedding those shortcuts.

Why does my job feel so frustrating? Product work operates at impact, execution, and optics levels. As scope grows, corporations push leaders toward optics regardless of where their energy sits. Doshi's fix was structural: capping team size at roughly 50 and staying at stages where impact work dominated, rather than climbing a ladder built around someone else's preferences.

Am I really listening? Doshi names this the hardest question, pointing to Rick Rubin and Drucker on what genuine listening — beyond active-listening technique — actually requires of a leader.

Product manager is an unfair role. So work unfairly.

PMs are split between maker and manager schedules, accountable without hiring authority, and absorbing organizational dysfunction — and the "great flattening" of tech (fewer ICs, larger surface, rising AI expectations) is worsening it. Fairness isn't coming; the answer is building asymmetric personal systems.

Seven tactics from IC PM Tal Raviv (Patreon, Riverside, Wix): (1) Execute meeting action items live while screensharing — draft the Slack summary, write the Jira ticket, schedule the next meeting before the call ends, so nothing enters a to-do list. (2) Replace half of scheduled meetings with sub-60-second Looms; pitch the time ROI in the message. (3) Stay out of Slack until mid-morning, section channels by urgency, hide read channels, and use reminders to auto-follow-up on ghosted threads; redirect DMs to public channels so threads resolve without you. (4) Build a self-reliant team by pushing decisions back to domain experts and publicly rewarding cross-functional initiative — "product isn't a role, it's a team." (5) Keep a lightweight Notion scrapbook of customer signals by swim lane so discovery never starts blank. (6) Dictate into focused GPT prompts via Whisper speech-to-text for PRDs and tickets — but don't let AI summarize raw customer data; staying in the weeds is where PM intuition forms. (7) Treat disconnection as work discipline: uninstall Slack mobile, display an OOO return date in your display name, and use physical transitions for genuine psychological detachment.

How to become a supermanager with AI

Managers who can explain "what good looks like" to a person can teach the same standard to a model — making their coaching available on demand rather than rationed by calendar. Hilary Gridley, director of PM at WHOOP, reports her team's performance, efficiency, and NPS all hit all-time highs after a year of this approach.

She built "The Executive Editor," a GPT trained on her most common writing feedback, which grades emails and decks on structure, clarity, and tone — PMs iterate through AI-driven rounds before Gridley sees the draft. A second GPT runs scenario-based logic drills during meeting dead time, building executive-influence skills without training sessions.

For meeting synthesis, PMs write their own three takeaways before seeing AI notes, keeping synthesis an active skill. To build a manager-clone GPT: gather before/after examples of work you've improved, have the model identify the patterns, then generate a prompt that replicates your edits — covering 0-to-80 so your time goes to 80-to-100.

A 30-day micro-assignment program builds iterative prompting as habit — adding context, pushing "make this 100x more specific," anticipating objections — replacing copy-paste fluency with genuine AI collaboration.

Behind the product: Replit | Amjad Masad (co-founder and CEO)

Replit's bet is not making engineers faster — it's making everyone a software builder. The real bottleneck was never talent or ideas but the fragmented tooling maze: runtimes, package managers, deployment. Replit collapses that into one browser-based environment so a non-technical founder can ship a working app in minutes.

The live demo is the evidence. A natural-language prompt produces a working Node.js/Postgres feature-tracker — voting, status columns, admin controls — in under ten minutes at about 15 cents of compute; a decent engineer would need several days. The AI runs as a second user in Replit's multiplayer editor, takes screenshots to verify rendering, writes git commit messages, and issues SQL queries directly.

The enabling layer is what Masad calls an AI-computer interface (ACI): text-optimized tool surfaces built for how LLMs process information, not human UIs. Primary coding model is Claude Sonnet, in a multi-agent setup with manager, editor, and critique models. Current limits: database migrations and large-scale architecture (sharding, queues), where the agent struggles on major iterations.

For product teams the shift is concrete: PMs prototype v1s and hand working code to engineers instead of mockups. SpotHero's head of marketing built a live competitive-pricing tracker this way; a public company tested a v1 with real users before it reached the engineering roadmap. A Figma-to-React extension converts design mocks to runnable code, eliminating the handoff gap.

Masad's "Amjad's Law": the ROI on learning basic coding doubles every six months as each AI capability jump multiplies what code literacy unlocks. The scarce skill shifts from writing code to generating ideas and debugging agents. For founders the key disposition is roadmap agility — Replit scrapped its roadmap the day Anthropic shipped computer use and pivoted immediately. The five-year image: a billion-dollar company run by one person, with AI handling development, support, and maintenance.

Behind the founder: Marc Benioff

Salesforce's durability at 25 years and $350 billion comes down to treating the company as permanently at the beginning. Benioff never looks at the stock price, calls himself a "startup CEO," and credits decades of Zen meditation for maintaining beginner's mind (Shoshin) — the expert's mind closes off possibilities; the beginner's holds infinite ones.

The Steve Jobs relationship shaped Salesforce's architecture directly. Around 2001, Jobs gave Benioff three directives: grow 10× in 24 months, land a landmark customer like Avon, and "build an application economy." Benioff interpreted that as a marketplace, bought appstore.com, and launched AppExchange in 2005. When the iPhone launched, he gifted Jobs both the domain and the App Store trademark — a gesture Jobs received with characteristic understatement: "this isn't going to be very big, but thank you."

On Agentforce: Salesforce is already resolving 83% of its own support inquiries via agents, with human escalation down 50%, and runs roughly 2 trillion AI transactions per week across Einstein and Agentforce combined. Benioff frames agents as persistent, memory-carrying digital labor — the Minority Report Gap store scene, not the Matrix. His four-layer stack: automate customer touchpoints → aggregate data → agentic layer → eventual robotic/drone layer.

The hardest moment in the company's history was the 2023 layoff of 10% of the workforce after pandemic-era overhiring. His lesson: "there is no up-and-to-the-right." Kaizen (continuous improvement) and appetite for the next disruption before it arrives is the only strategy that compounds across decades.

For Agentforce's launch, Benioff is running Matthew McConaughey and Woody Harrelson ads, aggressive anti-Copilot positioning, and hiring 1,000–2,000 new account executives simultaneously. The meta-principle — borrowed from how Chris Rock tests material in small clubs before a Netflix special — is to find the winning tactic through experimentation, then scale it into strategy.

Why great AI products are all about the data | Shaun Clowes (CPO Confluent, ex-Salesforce, Atlassian)

LLMs are only as good as the data fed to them — 90% of the work in building any serious AI product goes into getting the right data, timely and well-structured, into the model. Models are largely interchangeable; context is everything. Shaun Clowes, CPO at Confluent, uses this to explain why B2B SaaS incumbents are harder to displace than AI optimists assume: what makes Salesforce or Workday sticky isn't the UI or data model but years of accumulated business-rule configuration encoding how each company actually operates. Those rules are what any agent must reference — headless Salesforce still needs Salesforce. New entrants can win by embedding outcome data into workflows (Ashby does this in ATS), but incumbent lock-in compounds.

Most PMs spend the wrong 80% of their time on delivery management and internal politics instead of thinking from outside the building. The fix: write everything from the customer's and competitor's perspective and actively seek the counterfactual. LLMs accelerate this — paste customer interviews and ask where your strategy doesn't fit what they said; paste a competitor's public docs and ask what their strategy probably is. Confluent uses semantic clustering of inbound feature requests to surface trends across thousands of items.

Data works better as compass than GPS: it disproves hypotheses more reliably than it generates answers. Before trusting an exciting data point, check upstream, downstream, and one level above — a 2% cohort effect is noise; a retention win that collapses in week three is worthless; an ASP drop can negate a volume gain.

PLG in B2B is worth sustaining because without a dedicated team, end-user experience gets no structural advocate. The compounding happens when the two motions reinforce each other — self-serve signups surfacing to sales, pipeline leads flowing back into the product funnel when not ready to buy.

A guide to AI prototyping for product managers

Non-technical PMs can build working, shareable prototypes in under ten minutes — the main skill is knowing which tool to reach for.

Three categories exist. Chatbots (Claude, ChatGPT) handle single-page throwaway builds but can't host code or span multiple pages. Cloud development environments are the PM's real workhorse: v0 for polished default styling, Bolt for fast flexible layouts (caveat: it runs the server in-browser, so no logins or persistent storage without Supabase), Replit for internal tools and Python data apps, and Lovable when you want GitHub sync plus Supabase/Anthropic integrations. Local IDEs like Cursor are for people who already code.

The practical workflow: paste a Figma screenshot into Bolt with "Match it exactly," then layer on hyperspecific prompts — "add a price slider with a blue line and a black node." Converting a PRD or whiteboard sketch to a clickable prototype follows the same one-prompt-then-iterate pattern.

When the AI derails, four tactics recover it. Reflection: ask it to list requirements or diagnose errors before writing code. Batching: build the smallest functional version first, starting with the data model. Specificity: name the file, component, and exact behavior you want changed. Rollback: every cloud tool has checkpoints; use them when the AI overwrites too much instead of making targeted edits.

What’s in your stack: The state of tech tools in 2025

AI has displaced communication tools as the defining category in tech workers' stacks: 90% of 6,500 survey respondents use ChatGPT regularly — more than Gmail (76%) or Slack (72%). Claude is at 35%, Gemini at 24%, and over half combine assistants by use case (ChatGPT + Perplexity for research, ChatGPT + Claude for thinking).

Cursor, launched in 2023, is already used by 17% of all respondents and 21% of engineers, ranking above JetBrains and Xcode in adjusted value scores. v0 and Replit reach 10%, Bolt 5%.

The Jira paradox defines project management: Jira leads at 53% adoption while simultaneously topping the "please let us switch" list. Linear, founded 2019, now matches Asana (founded 2008) in share and leads the wish-list of alternatives. Notion compounds the disruption — second in PM, fourth in docs, third in CRM — its flexibility ("good for everything") the consistent differentiator.

Figma holds 97% penetration among designers and is gaining in presentations alongside Canva, which is displacing PowerPoint among founders and marketers. Slack is simultaneously the most and least valued communication tool — indispensable but described as a "noise generator." In customer support, Slack ties Zendesk at 29% because early-stage teams use shared channels rather than pay for dedicated platforms.

Three structural conclusions: bundling (Jira, Teams, Google Slides) locks users in but can't permanently hold off better-crafted alternatives; craft and UX now beat feature depth; and users mix tools within categories rather than picking one winner per job.

OpenAI researcher on why soft skills are the future of work | Karina Nguyen (Research at OpenAI, ex-Anthropic)

The skills that retain value as AI advances are relational, not technical: creative judgment, aesthetic taste, listening, prioritization, and people management. Karina Nguyen — who built Claude 3's post-training and evaluations at Anthropic, then led Canvas and Tasks at OpenAI — argues from direct experience: she pivoted from front-end engineering to research because Claude was becoming a competent front-end engineer.

Model training is more art than science. The central bottleneck is evaluation design: defining "correct" without degrading the model elsewhere. Training Claude 3 on self-knowledge (it has no body) made it confused about whether it could set an alarm — illustrating how data conflicts silently corrupt behavior. The supposed "data wall" is real only for pre-training on internet text; post-training via reinforcement learning on synthetic tasks faces no ceiling because learnable tasks are effectively infinite, and frontier benchmarks like GPQA (PhD-level Q&A) are already saturating.

Canvas and Tasks were built in months (four to five for Canvas, two for Tasks) via synthetic training. The workflow: define two or three core behaviors, generate examples via o1 producing and critiquing its own output, build deterministic evals (pass/fail on trigger, time extraction, comment placement), and iterate on win rates against prior model versions. Product managers keep spreadsheets of "current / ideal / why" — which o1 converts into training signal.

AI will handle strategy, scientific ideation, and routine knowledge work. What stays hard: creative writing voice, visual aesthetic discrimination, and reading human intent — the core difficulty for computer-use agents, which operate on pixels and cannot reliably judge when to ask a clarifying question versus proceed.

Anthropic emphasizes craft and tight prioritization; OpenAI is more bottoms-up with greater product freedom. AI research progress itself is bottlenecked by management: allocating compute to the right bets is the highest-leverage human role in the lab.

The creator of WordPress opens up about becoming an internet villain, why he’s taking a stand, and the future of open source | Matt Mullenweg (founder and CEO, Automattic)

WordPress powers 40% of all websites, 10x the market share of number-two Shopify, built under the GPL's four freedoms: use for any purpose, inspect the code, modify it, redistribute with those freedoms intact. That viral license is why Mullenweg calls Meta's Llama "fake open source": a clause requiring a commercial license above 750 million monthly active users means you eventually need permission, destroying the guarantee. AI models are largely trained on open source code, one of the few corpora whose license explicitly permits it, making open source foundational to AI's training data.

The conflict with WP Engine centers on three failures: trademark misuse that led 20-40% of surveyed users to believe WP Engine was officially affiliated with WordPress; stripping core features like post revisions to cut database costs; and bad-faith negotiations while secretly preparing a lawsuit through Quinn Emanuel. Silver Lake's 2019 acquisition preceded this. Mullenweg went public at WordCamp US in September; WP Engine responded days later with the lawsuit plus a PR campaign he compares to the Blake Lively dark-PR playbook. Sentiment analysis found 8% negative on LinkedIn/Facebook/Instagram versus 52% on Twitter, attributed to algorithmic amplification. A fork of Advanced Custom Fields into Secure Custom Fields was reversed by preliminary injunction; a separate project is now under active Automattic development.

The WordPress trademark sits in a tripartite structure: the Foundation holds it, WordPress.org is owned personally by Mullenweg, and Automattic holds the commercial license. Automattic ($500M ARR, 1,700 people in 90 countries) follows a Berkshire-like model, preferring healthy companies to accelerate over turnarounds. WooCommerce, acquired from a 35-person South Africa team, now generates the majority of revenue. Tumblr, bought for $3M from Verizon, came with 185 staff, FTC investigations, and App Store bans; half a billion Tumblr sites are being migrated to WordPress at the back end.

Building Lovable: $10M ARR in 60 days with 15 people | Anton Osika (co-founder and CEO)

Lovable hit $4M ARR in four weeks and $10M in two months with 15 people — fastest-growing startup in Europe — by making AI-generated web apps accessible to the 99% who don't write code. Type a description, get a deployable app in 30 seconds.

Anton Osika started with GPT Engineer, a 2023 open-source demo (50K+ GitHub stars) showing LLMs could generate multi-file apps from one sentence. Lovable repackaged that for non-technical founders: Supabase for backend, Cloudflare for hosting, one-click deploy. Mid-ramp, the team rewrote their entire codebase after the original scripting-language backend buckled under load.

The core technical edge is systematic "unsticking": identify where the AI reliably breaks down, tune the system quantitatively against those failure modes. They tackled the three worst stall points first — login, data persistence, Stripe payments. A live-edit layer lets users change text or colors on the rendered output without re-prompting the agent, something competitors didn't offer. GitHub sync lets non-technical founders stay in Lovable while engineers drop into Cursor on the same codebase.

Growth has been almost entirely organic: post demos and milestones publicly, let the product sell itself. The 18-person team (12 who write code) runs weekly planning on a FigJam board, picks the single biggest bottleneck, and ships. The roadmap stretches three months but changes monthly. They work in-office and eat lunch together as a deliberate coordination mechanism. Hiring prioritizes raw cognitive ability, obsession, and generalist mindset; every candidate does a paid work trial. The job posting is Shackleton-style, explicitly filtering out anyone seeking comfort.

On the future of product work: engineering skill remains valuable as constraint modeling, but taste and problem clarity become the scarcer inputs. To reach the top 1% of AI tool users: spend one full week taking a real problem end-to-end with AI, asking it when stuck.

Inside Bolt: From near-death to ~$40m ARR in 5 months—one of the fastest-growing products in history | Eric Simons (founder and CEO of StackBlitz)

StackBlitz went from near-bankruptcy to $40M ARR in five months after launching Bolt in October 2024 — zero to $20M in the first two months, with a 15-to-20-person team and 1M monthly active users.

The technical foundation is WebContainer, a full OS running inside the browser on the user's CPU, built over seven years. Competing products use cloud VMs that take minutes to boot; WebContainer boots in roughly 100 milliseconds, costs Bolt nothing per session, and enables a generous free tier without the abuse risks that force server-based products to restrict theirs. There are not enough VMs on the planet to serve a billion users, but there are a billion devices.

Claude Sonnet was the other unlock. The team tried building Bolt a year earlier and abandoned it — output was too unreliable. Sonnet crossed a critical threshold: software is deterministic, so Anthropic could generate near-unlimited training data through reinforcement learning across every imaginable app type. Bolt's growth started the week Sonnet went live.

Sixty-seven percent of Bolt users are not developers. One non-technical user built a CRM in three weeks for $300; an agency had quoted $30,000 and six months. Simons argues org charts will be rewritten: PMs will own UI-level execution, developers will shift to harder problems, and large triads will collapse into smaller pods. Bolt Builders — a certified human expert marketplace at roughly $50/hour — handles cases where AI gets stuck.

Operationally, the whole company joins a daily 8AM Zoom for at least an hour to preserve near-zero fidelity loss during hypergrowth. Keeping burn minimal for seven years — not scaling headcount in 2020–21 — is what let the company survive to launch Bolt.

Upcoming: a Figma-to-full-stack integration (prefix any Figma URL with bolt.new) and a Slack bot that acts as an on-call developer in threads.

How to win in the AI era: Ship a feature every week, embrace technical debt, ruthlessly cut scope, and create magic your competitors can't copy | Gaurav Misra (CEO and co-founder of Captions)

Captions (10M users, $100M raised) runs on one tempo: every engineer ships one marketable feature per week — something a user would download the app just for. When time pressure hits, cut scope not quality: strip to minimum useful core, ship it, let complaints drive the next release.

Technical debt is leverage. Each shortcut costs ~1% of daily capacity, so carry it until interest crowds out new work, then pay it down in a dedicated infrastructure quarter.

The roadmap splits in two. The public roadmap is everything users requested; every competitor has the same list. The secret roadmap holds features nobody asked for, sourced through company-wide quarterly brainstorming. First hit: Eye Contact (co-developed with Nvidia) — corrects gaze to face camera when a creator reads off-screen. The demo went viral in dozens of languages and was replicated by nearly every video app.

Snap's model informed much of this. Spiegel kept a 10–12 person design team (at 6,000 employees, no PMs for years) who owned both design and product. Opening to the camera rather than a feed was a moat Instagram couldn't copy without collapsing its own metrics. Misra built Snap's design-engineering function — hybrids who shipped live prototypes for small-market tests before full engineering investment.

On AI video: technology splits between neural rendering (per-person training, avatar-style) and large diffusion models (generalized, talking-head unsolved). Captions pursues the latter. Safety framework: documentation video (faking reality is harmful) vs. storytelling (ads, social, entertainment). Product design makes the first hard and the second easy.

Founding failure: Captions launched in two days, hit the App Store top, then Misra ignored it for eighteen months. He returned to find $500K in the account, 2,000 unanswered support tickets, and positive reviews. Refocusing produced an inflection so steep the original growth curve looks flat on the same chart.

Beyond vibe checks: A PM's complete guide to evals

Evals — not prompt engineering — are the deciding skill for AI product managers, because they let you measure exactly what a change does to each step in a system rather than eyeballing the final output. Aman Khan (Director of Product at Arize AI, co-creator of Andrew Ng's evals course) distinguishes three approaches: human feedback loops (sparse, expensive), code-based checks (fast but weak on open-ended tasks), and LLM-as-judge (scalable, written in natural language, probabilistic — needs ~90% agreement with human ground truth to trust). Every LLM eval has four parts: set the judge's role, supply the context variable, define what good and bad look like, and ground the terminology. The workflow runs in four phases — collect 10–100 labeled real-world examples, write and validate an initial eval, iterate on both the eval prompt and the agent prompt (swapping GPT-4o for Claude becomes testable), then run evals continuously in production as leading metrics. Common failures: over-engineering early (noisy signal), skipping few-shot examples in the eval prompt, and never comparing eval scores back to actual user outcomes.

OpenAI's CPO on how AI changes must-have skills, moats, coding, startup playbooks, more | Kevin Weil (CPO at OpenAI, ex-Instagram, Twitter)

Today's AI models are the worst you will ever use for the rest of your life — that framing shapes everything Kevin Weil, OpenAI's CPO, argues. Because the technology rewrites what computers can do every two months, the right posture is "model maximalism": if a product barely works at the edge of current capability, keep building — the models will catch up.

Evals — unit tests for model behavior — become the core PM skill. Models answer the same question with 60%, 95%, or 99.5% accuracy by task, and the product you build must reflect which bucket you're in. Deep Research was designed alongside its evals: hero use cases became test cases, fine-tuning climbed those benchmarks, and only when scores moved did the team believe they had a product. AI capability is capped by the quality of evals humans write.

OpenAI runs PM-light — roughly 25 PMs for hundreds of engineers — because too many PMs fill rooms with decks. The goal is high-agency PMs who make decisive calls when no one else will. Research, product, and engineering work as a single team from day one; handing off after research completes produces inferior output.

Durable startup moats will come from fine-tuned models trained on proprietary data OpenAI can't access. Weil expects ML engineers to become standard on every product team. Ensembles of specialized models outperform single general-purpose calls; OpenAI's customer support routes across o-series and 4o-mini by latency and cost.

Chat persists as the dominant interface because unstructured language is the lowest-common-denominator medium — any structured UI is a subset. Libra is Weil's biggest career disappointment: WhatsApp-native instant remittances killed by regulatory pressure and Facebook's reputation, not technical failure. The Move blockchain lives on in Aptos and Sui.

Prompting tip: include worked examples — poor man's fine-tuning that shifts output.

Everyone's an engineer now: Inside v0's mission to create a hundred million builders | Guillermo Rauch (founder and CEO of Vercel, creators of v0 and Next.js)

The bottleneck in building software has never been compute — it's the gap between people who have ideas and people who can ship them. v0 exists to close that gap. Guillermo Rauch built Vercel to trivialize deployment, Next.js to make React accessible, Socket.IO for real-time — v0 is the same move applied to the whole product-building stack. His addressable market is 100 million people — roughly Slack's MAU base — most of whom discuss products they wish existed but can't build.

v0 reached 1.3 million users; its community feature crossed 20,000 shared projects in under a month with thousands of forks. Rauch built a flight-radar app — Mapbox, canvas rendering for tens of thousands of aircraft, live aviation API — in two hours on bad in-flight wifi for $20/month. Equivalent engineer work would have taken weeks.

The argument about what AI does to programming is precise: translation tasks are going away. CSS specialists who converted Figma files into pixel-perfect layouts were doing translation, and the LLM now does it better — v0's output is more WCAG-accessible than Rauch's own code. What survives is knowing the vocabulary of systems — saying "turbulence" for an animation effect, "canvas rendering" when performance degrades — because that vocabulary steers the model. Foundational infrastructure engineers also survive; LLMs orchestrate compilers, they don't rewrite them.

Taste is learnable. The mechanism is "exposure hours" — a Vercel operating principle — deliberate time watching real users interact with your product. Rauch color-codes his calendar to protect customer meetings where he uses their product live.

At Vercel, PMs build animated interactive prototypes in v0 covering every product state, and the engineering response is "just ship it." Marketing and sales produce demo tools without filing tickets. v0 makes it safe to step outside your job description and make the thing.

Building a magical AI code editor used by over 1 million developers in four months: The untold story of Windsurf | Varun Mohan (co-founder and CEO)

Windsurf reached 1 million developers in four months not by building a better plugin but by forking VSCode entirely — the same autocomplete models tripled acceptance rates once it could render inline refactors natively rather than generating workaround images around the cursor.

Codeium started in 2019 on GPU virtualization. By mid-2022, with $2M ARR and 8 people, they saw everything converging on transformers, pivoted overnight to coding tools, and shed infrastructure. The second pivot — building their own IDE — came when VSCode's APIs capped AI's surface area.

Enterprise came early: Varun ran concurrent pilots with Dell and JPMorgan Chase before hiring a VP of Sales in late 2023. Requirements: codebase understanding at 100M+ lines, FedRAMP compliance, and hybrid deployment so indexes stay on the customer's infrastructure.

Three custom model layers: an in-house retrieval model that chunks and ranks code (sending 100M lines to Anthropic would exceed context limits by orders of magnitude), a fine-tuned edit model trained on incomplete-code preference data from tens of millions of daily interactions, and Claude Sonnet for planning. Windsurf tracks every user edit alongside every agent action, inferring intent from cursor movement and propagating changes automatically.

Hiring stays "dehydrated" — add someone only when genuinely underwater, because overstaffing causes people to manufacture work. Engineers serve as their own PMs; product strategy is three people for 160 staff.

On headcount: Amdahl's Law — if only 30 of 100 engineering time-units is writing code and AI compresses that to zero, total time falls 27%, not 100%. For high-ceiling companies the rational response is to invest more. Windsurf's own go-to-market staff built internal tools instead of buying SaaS, saving $500K.

The core bet: self-cannibalize every 6–12 months. Incremental requests are table stakes; what determines whether Windsurf wins is the internal roadmap nobody outside is asking for yet.

Make product management fun again with AI agents

AI automations — Zapier Agents, Lindy, Relay, Gumloop, Cassidy — are the most immediately practical agent category for PMs: pre-built integrations, trigger-based execution, natural-language prompting, no code. The core claim, from workshops with 5,000+ PMs, is that recurring tasks requiring judgment but not real expertise (customer-call prep, update summaries, NPS triage) are what these tools absorb, freeing time for customer immersion and strategic work.

The right task is what you'd give a smart junior intern: ongoing, one-at-a-time recurring work, not one-time batches (those go to Claude/Gemini via file upload or to Slack AI/Notion AI). Do the task manually once to write concrete instructions with examples.

Five design principles determine whether an agent works. Start smaller than feels right — one competitor site, not five. Cap the downside: DM you a draft rather than post publicly, append suggestions rather than edit. Give the right context (who's on the CS team, what your prioritization framework is) but skip the three-year vision. Stay close to raw signals — AI summarizing everything degrades PM customer intuition; demand exact quotes and links to original tickets. Treat misfires as prompt bugs: a PM's agent hallucinated sprint numbers because his examples said "Sprint 5" but the board didn't use that term.

For platform selection, use whatever the company already trusts, verify it supports the action (many read Zendesk tickets but cannot create them), and use an AI-generated comparison table to find the option with fewest moving parts.

The near-term ceiling is the feedback loop: agents can post to Slack but can't verify whether output looks right, because reading channel contents is a more sensitive permission than writing to it. The gap closes through orchestration — agents that screenshot their own output, as Replit's agent now does — not better models.

The rise of Cursor: The $300M ARR AI tool that engineers can't stop using | Michael Truell (co-founder and CEO)

Programming is heading toward a world "after code" — not chatbot-style prompting and not unchanged TypeScript and Rust, but a higher-level representation closer to pseudocode where engineers specify intent rather than syntax. Michael Truell, CEO of Anysphere, believes "taste" — knowing what to build and how it should work — will become the defining engineering skill as AI absorbs the translation layer between human intent and executable code.

Cursor reached $100M ARR in 20 months and $300M ARR in two years on a steady exponential with no single breakout. The company started late 2021 after two catalysts: the GitHub Copilot beta proving AI could be genuinely useful in coding, and OpenAI's scaling-law papers showing improvement would compound. They spent four months on mechanical engineering tooling before pivoting to software, concluding competitors weren't ambitious enough about where AI was headed.

Building a full IDE rather than a plugin rested on the conviction that the programming form factor would change so drastically that existing editors' extensibility would be insufficient. Truell rejects both pure-agentic and chatbot-only visions in favor of humans staying in the driver's seat with fast iteration loops. The most successful users chop tasks into small pieces rather than handing off large specs at once.

Custom models turned out to be central — against early assumptions. Cursor's autocomplete uses proprietary models for sub-300ms diff prediction; separate models handle codebase retrieval and fast application of edits sketched by frontier models. Every "magic moment" involves a custom model, built by fine-tuning open-source weights.

Hiring ran too slowly and over-indexed on credentialed-young profiles; later-career engineers often fit better. The current process centers on a two-day onsite work test. Truell expects software demand to expand dramatically as AI lowers build costs, keeping engineers in demand through a multi-decade transition more consequential than the internet.

Microsoft CPO: If you aren’t prototyping with AI, you’re doing it wrong | Aparna Chennapragada

Prototyping with AI is the new baseline — writing PRDs before building means moving too slowly. Microsoft CPO Aparna Chennapragada argues prompt sets are the new PRDs: the fastest path from idea to shared understanding is a working demo. Time to first demo has collapsed; the bar to ship has risen. The flood of AI-generated prototypes raises both floor and ceiling — breaking out requires taste and editorial judgment. PMs who add value through process won't last; those who make taste calls will.

NLX — natural language experience — is the new UX. Conversational interfaces have invisible grammars requiring deliberate choices. The prompt is a UI element; editable agent plans are a new construct. Verbosity of reasoning needs tuning: too terse loses trust, too verbose feels like watching a cron job. Follow-up suggestions need the same calibration. All active design decisions, not defaults.

Agents have three properties: autonomy (delegate goals, not steps), complexity (multi-step, not one-shot), and natural interaction. Practical example: ask a research agent to read every attendee's prior positions and synthesize an optimal persuasion pitch — new insight, not saved time.

Microsoft's "Frontier" program runs as a fake internal company staffed by early adopters on cutting-edge tools, institutionalizing "living one year in the future" without forcing the whole enterprise to change pace.

For zero-to-one bets, the two-of-three inflection test: step-function tech change, consumer behavior shift, business model innovation. Google Lens cleared the first two; Robinhood hit all three. Premature metric fixation is the companion failure — at a thousand users, CTR and retention are noise.

The hardest adoption problem is stale priors. Models that failed a year ago are unrecognizably better; that gap is an arbitrage. Google Now taught her being early is the same as being wrong — the interface was right, the intelligence wasn't. Today the problem inverts: exceptional intelligence, AOL dial-up chatbot interface.

Anthropic’s CPO on what comes next | Mike Krieger (co-founder of Instagram)

The biggest capability shift in Krieger's first year as CPO was Opus 4 delivering genuine novelty of thought. Running H2 product strategy through it returned angles he hadn't seen. SWE-Bench went from 50% to 72% in the window Dario predicted 90%; reading AI 2027 alongside his product strategy produced a "wait, am I the character?" moment.

Anthropic is patient zero for AI-written code: over 70% of pull requests are Claude Code-generated, Claude Code itself probably 95%+ self-written. The merge queue had to be re-architected from volume; PR review moved to Claude-reviewed-by-Claude with human acceptance tests. Real constraints are now alignment and coherent releases. What stays hard is structuring the question — decomposing backend vs. frontend — not writing code.

The clearest product leverage: embedding PMs in model post-training, not only UX. Artifacts with Claude 4 proves it — pairing the fine-tuning team with product designers produced something qualitatively different from off-the-shelf prompting.

On ChatGPT's consumer dominance: Krieger accepts it and leans into Anthropic's identity as the tool for builders and tinkerers. He spent year one with financial services API customers, not chasing consumer hits. For AI founders, durable spaces are deep domain knowledge with vertical GTM (legal, healthcare, biotech) or novel form factors incumbents can't pivot to.

MCP exists because every Anthropic integration was being rebuilt from scratch; two engineers proposed a shared protocol so the ecosystem benefits. The goal: expose every Claude primitive as MCP so Claude can write back to them. His wife asked Claude to add output to project knowledge; it couldn't.

On metrics: optimizing for session length or likability is a trap. The north star is whether Claude freed meaningful hours — a prototype done in 25 minutes instead of six — and whether it shows up in quiet moments that never register in thumbs-up data.

How to get your entire team prototyping with AI

The barrier to team-wide AI prototyping is not individual skill — it is the absence of shared component libraries and agreed handoff points across the development lifecycle. Without these, AI prototyping stays siloed, producing inconsistent visuals that fail in front of stakeholders.

Component libraries solve the consistency problem via three methods at increasing effort: screenshots fed to v0 or Bolt with a structured React/Tailwind prompt (no coding required); Magic Patterns' Chrome extension, which extracts live CSS from any webpage into reusable components; and code-based libraries using real front-end components mocked off the back end, optionally powered by the Figma MCP server (Get Code, Get Variable Definitions, Get Image) so Cursor can pull design tokens directly from Figma Dev Mode.

Once a library exists, the baseline-and-fork workflow prevents rebuilding from scratch: create a high-quality reproduction of the current product as a frozen baseline, then duplicate it freely to explore divergent ideas in parallel.

For lifecycle integration, fidelity must match context. Discovery warrants mid-fi, produced in ~20 minutes from a PRD prompt. Stakeholder alignment and user interviews need high-fi, taking 20–60 minutes of refinement. Engineering scoping uses the prototype for interaction documentation only — the generated code is not reusable. For accurate logos, paste the SVG source or a Brandfetch embed link directly into the prompt rather than letting the tool hallucinate a near-miss version.

AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff (Learn Prompting, HackAPrompt)

Prompt engineering is still alive: the gap between a bad prompt and a good one can mean 0% versus 90% accuracy on the same task. Sander Schulhoff created the first prompt engineering guide two months before ChatGPT launched, ran HackAPrompt — the first AI red-teaming competition, collecting 600,000 adversarial prompts now benchmarked by every major lab — and co-authored The Prompt Report, a 76-page meta-analysis of 1,500+ papers with OpenAI, Microsoft, Google, Princeton, and Stanford.

Techniques that move the needle: few-shot prompting (pasting examples of the target output — in one medical coding project this alone produced a 70% accuracy jump); decomposition (ask "what sub-problems need solving first?" before the main task); self-criticism (have the model critique then rewrite its output — diminishing returns past three cycles); and additional context at the top of the prompt, where it gets cached cheaply on repeated API calls. Ensembling — running the same question through multiple differently-prompted instances and taking the plurality answer — is effective for high-stakes product prompts.

What does not work: role prompting for accuracy tasks. Studies were re-analyzed and found statistically insignificant. Threat or reward language shows no reliable effect on modern models. Chain-of-thought phrasing is still worth adding on non-reasoning models: even GPT-4o skips its reasoning on roughly 1 in 100 calls at scale.

On security: prompt injection is not solvable — Sam Altman has cited a ceiling of 95–99% mitigation, not elimination. Prompt-based defenses and guardrails fail against attackers exploiting the intelligence gap between the guardrail and the main model. Fine-tuning to a narrow task is the most practical partial defense. The deeper concern is agentic deployment: a coding agent browsing the web could hit a page with injected instructions to write malware into its codebase. As agents gain real-world consequences, this unsolved problem becomes critical.

From ChatGPT to Instagram to Uber: The quiet architect behind the world’s most popular products | Peter Deng

At Uber, Peter Deng learned that sometimes the product doesn't matter: price and ETA were the product, not the pixels. The most valuable tech companies rarely started with a technological breakthrough; they applied existing tech to fundamental human needs and polished relentlessly. Instagram's Krieger and Systrom built on this with craft and conviction, not invention.

For companies building on AI, two durable moats: a proprietary data flywheel (Windsurf's accept/reject training signal is his example) and deep workflow fit. Product craft can override distribution advantages — Cursor and Granola broke through against Microsoft Copilot and Google Meet because the experience was compelling enough that users told friends.

On scaling from one to a hundred: plan chess moves in advance, build systems that let you go sustainably faster, and hire a growth team early — not primarily to drive growth but because growth PMs force instrumentation rigor that an analytics team alone never achieves. Instagram didn't know its user count when Deng arrived as first Head of Product.

His PM archetype framework, developed at Uber, sorts practitioners into five types: consumer PM (taste-first), growth PM (skeptical, data-first), business/GM PM (margin-first), platform PM (tools-for-others), and research/AI PM (model-depth). Everyone has a primary and secondary; the healthiest teams spike in different directions and argue productively.

Two hiring rules: if Deng is still telling someone what to do six months in, he hired wrong — the frame pressures both parties toward autonomy. His final interview across OpenAI, Uber, and Airtable covered only growth mindset, tested by asking candidates to describe a genuinely painful mistake and how it permanently changed how they work.

Career principle: optimize for learning over prestige — moving between Facebook, Instagram, Uber, Airtable, and OpenAI every two-to-three years once a role stopped teaching him new things.

An AI glossary

The transformer architecture (Google, 2017) enabled modern LLMs: its "attention" mechanism processes all words simultaneously rather than sequentially, capturing context at scale. Training works by next-token prediction — the model sees billions of sequences with the final token hidden, adjusting internal weights to learn grammar, facts, and reasoning.

Post-training shapes behavior: fine-tuning adds domain-specific training (medical literature, customer transcripts) to specialize outputs; RLHF builds a reward model from human preference comparisons, then uses reinforcement learning to align responses with human intent. RAG retrieves documents at run-time and injects them into the prompt, reducing hallucinations. Evals are unit tests for AI — predefined inputs checked against expected outputs, identified as the most underrated skill in AI product work.

Agents are a spectrum: more agentic systems plan independently, take real-world actions (commit code, update CRMs), draw on live data, and self-correct. MCP standardizes agent-to-tool connections. Synthetic data — LLM-generated text, diffusion-model images, synthesized audio — addresses training-data exhaustion as real internet content is fully consumed.

I’ve run 75+ businesses. Here’s why you’re probably chasing the wrong idea. | Andrew Wilkinson (co‑founder of Tiny)

Most entrepreneurs chase the wrong idea: crowded, glamorous markets where everyone wakes up wanting to compete. Charlie Munger's "fish where the fish are" points the other way — the archetype is a $30M business helping disabled people fill out government-assistance forms. Nobody dreams of building that, which is why margins survive.

Difficulty should match experience. Going after an AI company as a first venture is deadlifting 300 pounds on day one. Wilkinson's web design agency produced an immediate win; subsequent losses — pizzeria, cat furniture, DJ school — came from chasing harder models.

Any business that can't scale past the work itself is a job. "Lazy leadership" — delegate everything you hate as fast as possible — converts one into the other.

For idea generation, follow passion into an industry then find the profitable adjacent niche. He loved film, studied movie funding (terrible returns), and bought Letterboxd — a network-effect moat. Passion earns the knowledge; the niche is where the money is.

On people: "there are no problems, only people problems." If you've wondered even once whether to fire someone, do it. A single doubt is a leading indicator.

On AI: Lindy agents triage email by urgency, auto-archive low-value threads, and draft replies from multiple-choice prompts — replacing a full-time assistant. A separate agent pulls Perplexity research on each contact before meetings. Palm Treo analogy: we're in the preview phase; the iPhone moment is within five years, after which knowledge-work restructures.

On happiness: accumulating past $1B left the anxiety loop intact. The decisive intervention: an SSRI started at a near-homeopathic dose — within days, "someone turned down the volume on the nasty voice." ADHD diagnosis followed: 30% of entrepreneurs have it versus 5% of the general population, and treatment transformed focus and relationships — both outrank every financial outcome.

What people are vibe coding (and actually using)

Non-technical people are shipping vibe-coded tools they use daily, not prototypes they abandon. From 1,000+ replies to a single X/LinkedIn post, the dominant pattern is hyper-personal software that solves a specific friction no commercial app addresses: a carb scanner built to manage a child's diabetes (CarbScan.ai), an apartment buzzer that auto-answers deliveries based on custom rules (Buzzerbee), a Gmail add-on that holds email and releases it at scheduled times, a lash-style photo tracker, a one-screen tip-first bill splitter that beats Splitwise for pre-dinner math.

Cursor, Claude Code, Replit, and Lovable dominated as build tools. Many creators had zero coding background; Bolt and v0 appear frequently for the earliest scaffold. Chrome extensions are especially common since users spend most time in the browser.

Several hyper-personal apps attracted unexpected audiences: a clothing-layer weather app (85K users in nine months), a pickleball analytics tracker with U.S.-wide uptake. The gender mix in responses was notably balanced. The secondary finding: these aren't one-person experiments — friends, families, and strangers adopt them. The era of n-of-1 personalized software is real, and the gap between "I've always wanted this" and "it's live" has collapsed to days.

The AI-native startup: 5 products, 7-figure revenue, 100% AI-written code | Dan Shipper (co-founder/CEO of Every)

Every — 15 people, five products, seven-figure revenue, zero hand-written code — is Dan Shipper's proof-of-concept that AI lets a small generalist team operate at a scale that was structurally impossible three years ago. The Cora team (an AI email chief-of-staff, 2,500 beta users) is two engineers plus fifteen simultaneous Claude Code instances; they review code but write none. Shipper is explicit that this still requires knowing how to code — English atop scripting languages follows the same decades-long transition as C→Python — but no one writes by hand.

Three practices account for most of the leverage. First, a dedicated head of AI operations (Katie Parrott, background in content marketing and process design) meets with Shipper weekly, catalogs every repetitive task, and turns those into prompts and automations — separating workflow-building from the people doing the work. Second, "compounding engineering": every unit of work should make the next unit cheaper. Engineers build PRD-generation prompts, copy-editing Claude Code slash commands (one command runs style-guide edits across the entire codebase and opens a GitHub PR for the editor to approve), and a library of agent personas — Friday, Charlie, Claude — used in parallel because different models have different "taste." Third, internal pain points are the first version of products: Monologue (voice-to-text), Spiral (content automation with self-judging agentic drafts), and Cora all started as internal tools.

Product selection filter: services that used to require expensive professionals — chief-of-staff, ghostwriter, lawyer — democratized by cheap intelligence. The consulting arm (~$1M last year, growing) trains large companies on the same approach; the single best predictor of adoption success is whether the CEO uses ChatGPT daily.

His definition of AGI: when it becomes economically profitable to leave an agent running indefinitely — the "teenager" phase, always doing something useful without waiting for the next instruction.

Build your personal AI copilot

LLMs feel generic and disappointing for strategic work because they lack context, not intelligence. Feed them the background any colleague would need and they become a genuine thinking partner for long-horizon work.

The setup uses the "Projects" feature in ChatGPT, Claude, or Gemini (paid plans). Three steps map onto new-hire onboarding. First, write instructions — a persistent system prompt defining the AI's role, values, and coaching behaviors. Second, upload project knowledge: strategy decks, customer research, competitive landscape, org charts, past retrospectives, performance reviews. If documentation is thin, prompt the AI to interview you and generate a summary document to re-upload. Third, open a per-initiative chat thread, dictate everything you know in stream-of-consciousness speech-to-text, then ask "What is the single most important thing I should do next?" — and get responses that name your actual stakeholders because the context is there.

With context loaded, useful prompts become conversational one-liners. The copilot can challenge roadmap assumptions, draft user stories from verbal rambling, role-play stakeholder conversations, generate prototypes via Claude Artifacts or tools like Lovable, and brainstorm AI automations. The automation prompt enforces a hard constraint: event-driven triggers only ("when a new ticket arrives"), not batch schedules — reliability degrades on batch tasks.

Keeping the copilot current requires a "gossiping" habit: whenever something shifts — a stakeholder reversal, new data, a hallway conversation — dictate it into the thread by voice, as you'd vent to a desk neighbor. At initiative close, upload retrospective notes and personal reflections to project knowledge, compounding value over time.

The current ceiling is that you are the human API — manually exporting PDFs, copying text, bridging tools. The next generation would pull directly from project management and messaging systems and push proactively, like a chief of staff who notices your calendar and opens a role-play before you ask.

Pricing your AI product: Lessons from 400+ companies and 50 unicorns | Madhavan Ramanujam

Price is a measure of value, not a dollar figure — and 72% of product innovations fail commercially because companies treat it as an afterthought. Madhavan Ramanujam of Simon-Kucher argues for "product market pricing fit": have willingness-to-pay conversations before building, not after launch.

Core method: after pitching the value proposition, ask "what's an acceptable price, an expensive price, a prohibitively expensive price?" — never "what should I charge?" Rahul Vohra used this to land Superhuman's $30 price point. Porsche battle-tested every Cayenne feature against willingness to pay before a blueprint was drawn, yielding more than half of its profit. A two-sided marketplace discovered its beloved "Highlight Connections from Facebook" feature had zero paying segment — saving a wasted build.

Segmentation must reflect differing needs and willingness to pay, not demographics. King Charles and Ozzy Osbourne share the same demographic profile. Apple's iPhone ($299–$1,499) productizes to segments; Uber's pool/X/Comfort/Black maps to moment-level switching; Eventbrite's three tiers serve distinct event-operator needs.

How you charge matters more than how much. Michelin shifted truck tires from per-tire to per-mile — the cost became a variable pass-through on trucker invoices. Segment moved from API-count to monthly tracked users because marketing buyers couldn't evaluate APIs. Subscription fits ongoing value (LifeLock, Spotify); usage fits intermittent, episodic value; hybrid covers mixed cases.

Behavioral pricing exploits the irrational brain. Adding a $299 decoy alongside a $99 middle tier shifted mix upward for a 30%+ ARPU increase with no product changes. The Panini effect — displaying products as an incomplete puzzle — lifts multi-product attach rates from ~20% to 40–50% by triggering the compulsion to complete sets.

In downturns: de-feature a cheaper alternative to retain churning customers rather than discounting (discounts become the new permanent price). Try non-pricing moves first — more product, longer contract terms, or extended payment windows.

He saved OpenAI, invented the “Like” button, and built Google Maps: Bret Taylor on the future of careers, coding, agents, and more

Agents accomplish jobs rather than assist people, and that changes how software gets built, sold, and priced. Bret Taylor (co-creator of Google Maps, inventor of the Like button, CTO of Meta, co-CEO of Salesforce, OpenAI board chairman, CEO of Sierra) argues that autonomous agents with measurable output make outcomes-based pricing viable — Sierra charges per resolved interaction, not per token or seat. Customers see 50–90% automation with CSAT scores of 4.6–4.7.

On market structure: frontier models consolidate to hyperscalers (too CapEx-heavy for startups); tooling companies risk displacement by big-lab developer days; the real opportunity is vertical agent companies where moats shift from model orchestration to domain depth. Many AI founders should lean harder on direct sales — the buyer and user are often different people, which breaks product-led growth.

On coding: computer science stays valuable for systems thinking. The open question is what programming system — not Python, more likely something with Rust-style compile-time verification — best supports human oversight of AI-generated code.

Google Maps grew from a failed V1 (Google Local, a digital Yellow Pages a homepage link couldn't save) to 10 million users on launch day, 90 million when satellite imagery shipped. Build native to the medium, not a digitization of the prior thing. FriendFeed lost to Twitter because Biz Stone put Oprah and Ashton Kutcher on Twitter while FriendFeed polished features.

The Like button started as a "one-click comment" to clear single-word acknowledgments. A heart was rejected as wrong for tragedy posts; "like" won as the most neutral positive signal.

Sheryl Sandberg caught Taylor editing a subordinate's deck himself and reframed his approach: ask not what you enjoy but what is most impactful today. Founders default to their strengths as the answer to every problem — answering that question honestly is harder than it sounds.

Inside ChatGPT: The fastest-growing product in history | Nick Turley (Head of ChatGPT at OpenAI)

ChatGPT was not designed to be a product. It was a 10-day productization of a hackathon demo that OpenAI shipped before the holidays expecting to collect data and wind it down. Sam Altman tweeted it, it went viral, and the team spent weeks insisting usage would drop before accepting they had something. The original name was "Chat with GPT-3.5" because nobody believed in it enough to brand it.

Three early decisions compounded into advantages: no waitlist, so users could watch each other in real time and TikTok threads became a free discovery engine; free access, even though GPT-3.5 had been in the API for six months and anyone could have built something comparable; and a $20 price set via a Van Westendorp survey posted as a Google Form on Discord. The $200 Pro tier came later purely as a vehicle to ship capabilities — o3 Pro, GPT-5 Pro — that couldn't yet scale to all users.

Turley splits retention gains roughly into thirds: systematic improvement on actual use cases (writing, coding, advice); high-leverage capability unlocks, notably search, which eliminated the knowledge-cutoff problem and created outlinks that now send meaningful traffic back to publishers; and classical product work like removing mandatory login. The "model is the product" principle means model iteration and product iteration have collapsed into the same loop — ChatGPT broke the old pattern of waiting a year for a big-batch release.

The sycophancy incident — an update that made ChatGPT endorse whatever the user said — forced sycophancy-tracking evals into every subsequent release; GPT-5 improves on that metric.

At 700 million weekly active users ChatGPT still feels to Turley like MS-DOS: the chat interface was the simplest thing to ship, not the destination. Natural language input is permanent; turn-by-turn chat is not. The product has no Windows yet.

Why your AI product needs a different development lifecycle

AI products break traditional software assumptions in two ways: their behavior is non-deterministic on both ends (users express intent through open-ended prompts, models generate probabilistic outputs), and they introduce an agency-control tradeoff that conventional tools never had. The more autonomously a system acts—booking flights, resolving support tickets, opening pull requests—the less visibility and override capacity you retain. Shipping a fully autonomous system before that tradeoff is understood is how demos become production fires.

The CC/CD framework (Continuous Calibration/Continuous Development, named as a deliberate contrast to CI/CD) structures development around earning agency incrementally. The Continuous Development side has three steps: scope capability by version using the agency-control axis rather than feature sets (v1 = high control/low agency; v3 = full autonomy), set up the simplest measurable application with control handoffs baked in from day one, and design application-specific evals against a reference dataset of 20–100 curated examples. The customer support illustration runs from ticket routing (v1) to human-reviewed draft responses (v2) to autonomous resolution with fallback (v3)—the same ladder GitHub Copilot and Cursor climbed.

Post-deployment, the Continuous Calibration side runs evals on live interaction logs, manually reviews the lowest-scoring clusters to surface repeating error patterns (documented in a structured table), then applies targeted fixes—prompt changes, retrieval improvements, or architectural additions—before re-evaluating. Evals themselves often need revision, because real user behavior diverges from the reference dataset. The loop repeats until the system earns enough trust to move to the next agency level.

How 80,000 companies build with AI: products as organisms, the death of org charts, and why agents will outnumber employees by 2026 | Asha Sharma (CVP of AI Platform at Microsoft)

Software products are becoming living organisms rather than static artifacts — tuned continuously through interaction data, fine-tuning loops, and rewards design. As Microsoft's CVP of AI Platform overseeing 80,000+ companies building on Azure, Asha Sharma argues that proprietary post-training data is the new IP: once a model exceeds 30 billion parameters (per Nathan Lambert's leaderboard study), pre-training CapEx loses economic sense — you get more leverage by fine-tuning an existing model on your own use-case data. Microsoft's Dragon physician AI saw character acceptance rates jump from 30–60% to 83% after annotating 600,000 patient-physician interactions and feeding them into a continuous optimization loop.

The GUI era is giving way to code-native, composable interfaces because text streams interact better with LLMs — the same pattern as databases moving to SQL and cloud infra moving to Terraform, just faster. Builders who think in composability rather than canvas will have structural advantages.

At the organizational level, the org chart is becoming a work chart. With 15,000 Azure customers already running agents (millions of agent instances), agents route tasks, execute autonomously, and self-correct via evals and observability loops — with fewer management layers needed. The full-stack builder emerges as the dominant archetype: small cross-functional squads that can metabolize the entire product loop (rewards design, UI, evals, fine-tuning) without waiting on hand-offs between roles.

Planning adapts by thinking in seasons — defined by secular shifts (prototyping era → reasoning models → agents) — rather than fixed 6-month roadmaps. Within a season, teams run loose quarterly OKRs and 4–6 week squad goals, with deliberate slack built in for slope.

The biggest lesson from WhatsApp, Instacart, and now Azure: invisible infrastructure (reliability, data residency, privacy, latency) beats features every time. Satya Nadella's core leadership move — generating renewable optimism — is what sustains execution through constant disruption.

How Devin replaces your junior engineers with infinite AI interns that never sleep | Scott Wu (Cognition CEO)

Autonomous coding agents represent a qualitatively different paradigm from AI code completion — they work asynchronously, integrate with Slack, GitHub, and Linear, and accumulate institutional knowledge of a codebase over time. Cognition's Devin launched in March 2024 as what Wu calls a high-school CS student; by late 2025 it operates as a junior engineer. Cognition's 15-person engineering team runs up to five Devin sessions per engineer simultaneously, Devin merges several hundred PRs per month into their own codebase, roughly a quarter of all PRs are currently Devin-authored, and Wu expects that to exceed 50% by year-end.

The core workflow principle: give Devin tasks, not problems. Well-defined, verifiable work — a specific front-end change, a bug fix, adding tests — reaches near-autonomous completion. Larger scope requires steering at planning and review. Wu frames the shift as engineers moving from bricklayer to architect: defining architecture and trade-offs (roughly 10% of engineering time) becomes the job; implementation, debugging, and log-trawling (the other 90%) get delegated. He invokes Jevons Paradox — cheaper execution historically expands total demand — to argue AI will increase engineering headcount, not reduce it.

On defensibility, Wu prefers stickiness over moats: Devin's persistent codebase representation compounds over time, and the multiplayer dimension — colleagues annotating each other's sessions in Slack, reviewing PRs, onboarding new hires through Devin's wiki — creates compounding value generic tools cannot match.

High-compute reinforcement learning, not imitation learning, is what made coding agents viable; code's automated feedback loops are what RL training requires. The long-term product vision ends at engineers directing changes through their product UI, never viewing raw code.

On hiring, Wu flew to a junior MIT candidate's parents in North Carolina to negotiate a schedule letting him work nearly full-time while still graduating — the literal enactment of "hire the best people."

Building eval systems that improve your AI product

Generic AI quality metrics — hallucination scores, toxicity rates — produce dashboards teams ignore because those scores don't correlate with actual user failures. Evaluations must be grounded in error analysis before measuring begins.

Phase one: a single principal domain expert (a psychologist for a mental health bot, a lawyer for legal analysis) reviews roughly 100 sampled interactions via open coding — free-form critique plus binary pass/fail per trace, detailed enough to serve as few-shot examples for an LLM judge. Critiques are then grouped via axial coding into a prioritized taxonomy of under ten failure modes. Counting category frequencies shows where to invest: for one apartment-leasing assistant, conversation flow, handoff-to-human, and rescheduling failures dominated.

Phase two: objective failures get code-based evaluators (fast, deterministic, cheap); subjective failures get an LLM-as-judge built by splitting ground-truth labels into train/dev/test sets and validated with TPR and TNR — not accuracy, because a judge that always predicts "pass" can hit 99% accuracy on a high-success product while catching zero real failures. Known judge error rates can statistically correct reported scores.

Phase three: code-based evals run in CI against a golden dataset of regression cases, blocking regressions. LLM judges run asynchronously on production samples; critical paths get synchronous guardrails with near-zero false-positive rates. When production monitoring flags a new failure mode, error analysis restarts and the new case enters the golden dataset — a flywheel where every discovered failure permanently hardens the system.

The ultimate guide to AEO: How to get ChatGPT to recommend your product | Ethan Smith (Graphite)

LLM answer engines rank products by citation frequency, not position. ChatGPT, Perplexity, and Gemini run a web search (RAG), collect citations, then summarize them; the product mentioned across the most sources wins. This inverts Google's playbook.

Early-stage companies benefit most. Google SEO requires years of domain authority; AEO does not — a company cited in a Reddit thread or YouTube video can appear in ChatGPT answers immediately. Leads from LLM traffic convert at 6x the rate of Google Search (Webflow's figure) because users arrive already narrowed by multi-turn conversation; LLM now accounts for 8% of Webflow signups.

The playbook splits into onsite and offsite. Onsite: build landing pages targeting question clusters and cover the long tail — ChatGPT queries average 25 words versus 6 for Google, so entire question categories have never been answered. Help centers are underused: move them from subdomains to subdirectories, cross-link aggressively, and add pages for obscure integration questions. Offsite: citations come from YouTube/Vimeo (especially open for unglamorous B2B niches), Reddit/Quora (transparent participation — name yourself and employer, give a useful answer; fake-account networks get banned), and tier-one affiliate publishers like Dotdash Meredith (Investopedia, Good Housekeeping dominate LLM citations for consumer categories).

Track "share of voice" — percent of runs where you appear, across question variants and platforms — using any of ~60 commodity answer-tracking tools. Run controlled experiments: intervene on one question set, leave another untouched, measure after a few weeks, reproduce before committing. Most published AEO best practices are unverified assertions repeated until conventional.

Fully automated AI content does not rank. Graphite's study found only 10–12% AI-generated content across thousands of Google and ChatGPT citations. The structural risk is model collapse: if AI content dominated citations, derivatives of derivatives would loop in, collapsing wisdom-of-the-crowd diversity into one opinion on everything. AI-assisted editing is correct.

Why AI evals are the hottest new skill for product builders | Hamel Husain & Shreya Shankar (creators of the #1 eval course)

Building AI products without reading your own data is the core failure mode; "evals" is just data analytics applied to LLM applications, and the most neglected first step is manually reading individual traces.

The process Hamel Husain and Shreya Shankar teach starts with error analysis: sample roughly 100 conversation traces, write a plain-language note on the first obvious failure in each one (open coding), then move on. This surfaces failure types you could never predict — a real estate leasing agent hallucinating a virtual-tour offering, abrupt call-transfers that never confirm with the user, garbled text-message threads where the AI loses context. LLMs cannot do this step for you: they lack product context and will call the virtual-tour hallucination correct.

Once you have enough notes, use an LLM to cluster them into actionable categories (axial codes) — "human handoff failures," "conversational flow issues," "formatting errors" — then count occurrences with a pivot table. That counting is the entire analytical reveal: you now know what to fix first.

For failures that can't be caught with simple code checks, build an LLM-as-judge evaluator: one narrow, binary (pass/fail) prompt per failure mode. Binary scoring is non-negotiable — a 1–5 scale produces uninterpretable averages. Before shipping, validate it against your own labels using a confusion matrix; raw agreement percentage misleads when errors are rare. Run judges on sampled live traces continuously. Shreya Shankar's research ("Who Validates the Validators?") shows that rubrics always evolve as you see real output, so the process is inherently iterative. Four to seven judges is a typical end state; many failures get fixed by a prompt edit and need no judge at all.

Upfront cost is three to four days; maintenance is roughly 30 minutes a week. A-B tests and dogfooding are subsets of this discipline, not alternatives to it.

First interview with Scale AI’s CEO: $14B Meta deal, what’s working in enterprise AI, and what frontier labs are building next | Jason Droege

Meta's $14B investment bought 49% non-voting stock in Scale AI; Scale remains independent, Alex Wang moved to Meta's superintelligence team, Droege became CEO. Scale has ~1,100 employees, two lines each at hundreds of millions in revenue, and has grown every month since the deal.

What Scale does has changed as models advanced. Eighteen months ago, tasks were preference rankings between short stories. Today a single task is a world-class developer building an annotated website, or a PhD explaining a nuanced cancer topic. 80% of the expert network holds a bachelor's degree or higher; 15% hold PhDs. The shift from generalist labor to deep specialists is Scale's own trajectory accelerating.

The next frontier is RL environments: sandboxes where agents learn to navigate real software — Salesforce, healthcare platforms, calendar apps — reliably enough for production. Scale has built these for over a year. Core challenge: the permutation space across enterprise configurations is vast, so data value scales with how broadly it transfers.

Enterprise deployment takes 6–12 months before an important process is automatable. "Easy to learn, hard to master" captures the current gap. Most pilots fail from insufficient investment, not product limits.

On business building: Droege filters ideas through gross margin as a proxy for differentiation, plus network-effects and lock-in. At Uber Eats, the team reverse-engineered restaurant economics by weighing actual ingredients — finding ~70–80% incremental gross margin on incremental orders, justifying the commission model. Droege initially rebuffed McDonald's; the delay produced a stronger exclusive deal and hockey-stick growth. Uber Eats went from zero to $20B in 4.5 years; COVID took it to $50B.

On hiring: three criteria — curious problem-solving, cross-functional humility, leadership potential. He kept the Uber Eats team from zero to $20B, valuing complementary strengths over credentials. Motto: "The end is never the end" — survival precedes thriving.

Inside Google's AI turnaround: The rise of AI Mode, strategy behind AI Overviews, and their vision for AI-powered search | Robby Stein (VP of Product, Google Search)

Google search traffic is growing, not shrinking, despite ChatGPT and Perplexity. Robby Stein, VP of Product for Google Search, argues AI is expansionary: it creates new question-asking behavior on top of the immovable base of queries (phone numbers, directions, prices) that competitors haven't eroded. Google Lens visual search grew 70% year-over-year and already handles billions of searches.

Google's AI product stack has three layers. AI Overviews delivers a fast summary atop standard results. Multimodal search (Lens, camera) handles visual queries. AI Mode — built by a five-to-ten-person team roughly a year before launch — is the full-conversation frontier experience. It uses query fan-out: the model issues dozens of background Google searches to fact-check its response, then surfaces links to authoritative sources. AI Mode is not a general chatbot; it targets information and research, not creativity or productivity tasks.

The differentiator is integration with Google's proprietary data: 50 billion Shopping Graph products updated 2 billion times per hour, 250 million Maps places, Finance data, and the full web index filtered by spam signals. The on-ramp is natural — people already bring hard questions to Google; now the AI can answer them.

On AEO: query fan-out means searches still happen, so Google's helpful-content signals (user intent, originality, authoritative sourcing) remain valid. The new opportunity is content for the complex advisory questions AI now handles.

Stein's three product-building principles: understand why people hire your product (Clayton Christensen's jobs-to-be-done); use data to reach true root causes rather than surface symptoms (Instagram's Close Friends failed for two years because "favorites" was mistranslated, leading to tiny lists that generated no reciprocal DMs and killed the emotional connection the feature depended on); and design for clarity over cleverness (Don Norman's doors principle, the name "AI Mode"). Humility runs through all three as a fourth thread.

Everyone should be using Claude Code more

Claude Code's name is its worst marketing: non-technical people who treat it as "Claude Local" — an AI agent with direct access to local files, browser, and system — unlock capabilities cloud chatbots can't touch. It handles large file sets, runs longer tasks, and takes real actions. Two terminal commands install it.

Dan Shipper downloads meeting recordings and asks it to flag every moment he avoided conflict. Justin Dielmann runs it from his home directory as a file organizer — finding duplicates, cleaning downloads, auditing structure. A Clay PM fed it internal hiring docs plus competitor JDs and got a full job description, interview plan, and rubric. Derek DeHart connects it via MCP to Fireflies, Linear, and Notion as a product-research hub, validating assumptions against customer call transcripts. Gang Rui's slash command cross-references journal entries against Git commits to surface gaps between intentions and execution.

Other uses: auto-renaming invoice PDFs by vendor/date for taxes; generating changelogs from commit history in 10–15 minutes; scraping competitor ad libraries; building a daily "Claude CEO" briefing from Gmail, Brex, Mercury, and Linear.

How to measure AI developer productivity in 2025 | Nicole Forsgren

Speed and stability in software delivery move together — shipping smaller changes more frequently produces fewer bugs, a smaller blast radius, and faster recovery. This inverts the ITIL wisdom that two-week change-approval windows ensure stability. Elite DORA performers deploy on demand, achieve lead times under one day, restore in under an hour, and keep change fail rates at 0–15%. No significant difference exists between small and large companies; retail outperforms all sectors, likely because weaker performers were eliminated.

DORA's four metrics — lead time, deployment frequency, mean time to restore, change fail rate — measure outcomes. Improving them requires fixing underlying capabilities: automated testing, CI/CD, trunk-based development, loose coupling, and culture. The dora.dev quick check identifies which constraints are most likely given your performance tier and industry.

SPACE complements DORA for picking balanced metrics around any complex creative work. Its five dimensions: Satisfaction and wellbeing, Performance, Activity, Communication and collaboration, Efficiency and flow. Use at least three at once — activity metrics alone (lines of code, PR counts) are the perennial failure mode. DORA is an implementation of SPACE applied to the delivery outer loop.

On AI: developers using tools like GitHub Copilot now spend roughly 50% of their time reviewing generated code rather than writing from scratch. The gain isn't "same task in half the time" — that framing produces headcount arguments. The actual shift is cognitive reallocation: routine work is offloaded, freeing capacity for harder problems. Open questions include overreliance, novice-versus-expert differentials, and whether SPACE needs a trust dimension.

The four-box framework operationalizes measurement: write a word-level hypothesis (top two boxes), then map each concept to data proxies (bottom two). Disagreements surface at the proxy level, keeping bad metrics from corrupting good hypotheses. Google triangulates instrumentation against developer surveys; when they conflict, the surveys are almost always right.

Al Engineering 101 with Chip Huyen (Nvidia, Stanford, Netflix)

What actually improves AI applications is not tracking the latest models or agonizing over vector databases — it's talking to users, preparing better data, writing better prompts, and optimizing the workflow. Chip Huyen, author of *AI Engineering* (O'Reilly's most-read book of 2025) and former NVIDIA NeMo developer, built this into a viral table because teams recognize themselves in the wrong column.

Pre-training encodes statistical patterns from internet-scale text (traceable to Claude Shannon's 1951 work); post-training is where labs now compete hardest because pre-training data is near-exhausted. Post-training combines supervised fine-tuning on expert demonstrations with reinforcement learning: RLHF trains a reward model on human preference comparisons; verifiable rewards let math and code be checked automatically. Data-labeling companies face structural pressure: a tiny buyer pool with unlimited switching leverage.

RAG retrieves relevant context before answering. Performance gains come almost entirely from data preparation — chunk sizing, metadata, hypothetical-question augmentation, rewriting docs in Q&A format — not vector database choice. Evals are essential at scale or where failures are catastrophic; for low-stakes features, ROI often favors shipping something new instead.

AI engineers build products on top of existing models rather than training them. A randomized Cursor trial across a 30–40 person team found the highest performers got the biggest lift; elsewhere, senior engineers resist because they find AI output below their standard. Several companies are restructuring: senior engineers set guidelines and review code while juniors and AI produce it — leaving open how the next generation will develop systems-level judgment.

Test-time compute — generating multiple candidates and voting, or extending reasoning chains — improves perceived performance without changing the base model. Voice interfaces remain hard engineering problems: multi-hop latency and interruption detection are classical challenges, not foundation-model ones. A "billion-year" nihilism — nothing persists, so try things — is Huyen's working life motto.

How Block is becoming the most AI-native enterprise in the world | Dhanji R. Prasanna

Block's most consequential productivity gain came not from AI but from replacing a portfolio structure (Square, Cash App, Afterpay each running its own engineering) with a single functional org where all engineers report to one head. That structural change enabled coherent technical strategy and company-wide AI deployment. Dhanji Prasanna wrote the internal "AI manifesto" as a part-time consultant; Dorsey hired him as CTO after a two-day walk around Sydney.

The main AI output is Goose: an open-source, MCP-based desktop agent that connects to enterprise systems — Snowflake, Tableau, GitHub, calendar, AppleScript — through lightweight extensions writable in a few lines. AI-forward engineering teams report 8–10 hours saved per week; across all staff (legal, risk, support), Block estimates roughly 20–25% of manual hours saved. The biggest surprise: the highest gains appear in non-technical employees who use Goose to build internal tools without waiting months for an engineering team. The enterprise risk team built a full self-service risk system in hours. One engineer runs a Goose variant that watches his screen and Slack, opens PRs autonomously for features mentioned in conversation, and reschedules calendar conflicts without input.

The next Goose version targets hour-scale autonomous sessions rather than the current five-to-seven-minute median; overnight idle LLM capacity is treated as waste. Block is experimenting with parallel overnight builds — describing multiple approaches before sleep and discarding most in the morning — as a replacement for the classic "pick best path before committing resources" discipline.

On hiring: learning mindset matters more than current AI fluency. On code quality: YouTube stored videos as MySQL blobs on a pure Python stack and became Google's most valuable product. Code quality and product success have no meaningful relationship. Both Goose and Cash App started as single-engineer hack-week experiments; Block's first public Bitcoin product was a three-person hackathon with Dorsey.

The Godmother of AI on jobs, robots & why world models are next | Dr. Fei-Fei Li

Modern AI traces to a 2006 data-deficit diagnosis: neural networks were failing not because the math was wrong but because training data was nothing like the volume humans accumulate through experience. The fix was ImageNet — 15 million internet images across 22,000 object categories, open-sourced with an annual competition. In 2012, Geoff Hinton's Toronto team combined ImageNet with two consumer NVIDIA GPUs and made the first serious progress on object recognition. That trio — big data, neural networks, GPUs — still underlies ChatGPT; only the scale changed. As recently as 2016, tech companies avoided "AI" as a brand liability; by 2017 every company claimed it.

AGI is a marketing term, not a scientific one. Current models cannot count chairs in a video walkthrough of two rooms and couldn't derive Newtonian mechanics from modern celestial data — both trivial for a child. Scaling alone won't close those gaps; fundamental innovations remain ahead.

The missing piece is spatial intelligence: the capacity to create, reason about, and navigate genuinely 3D environments. Language models have a clean training-to-output alignment (tokens in, tokens out); robotics doesn't — robot actions require 3D-world data that passive web video can't supply, so the "bitter lesson" (more data always wins) is harder to apply. The self-driving comparison is sobering: 20 years from Stanford's DARPA desert win to commercial Waymo, and wheeled cars are simpler robots whose goal is to avoid touching anything.

World Labs (co-founded with Justin Johnson, Christoph Lassner, and Ben Mildenhall, ~30 people) launched Marble: a prompt-to-3D-world product where a sentence or image generates a navigable, exportable environment. A Sony virtual-production collaboration cut scene-creation time 40x. Other early use cases: robotic simulation training data, game development, and controlled immersive environments for psychiatric exposure research.

Stanford's Human-Centered AI Institute, co-founded in 2018, now spans all eight Stanford schools and has shaped federal and state AI policy through congressional bootcamps, the AI Index report, and a national AI research cloud bill.

Every person — musician, nurse, farmer — has a role in AI. Human dignity must sit at the center of how the technology is developed and governed, not just its productivity gains.

AI tools are overdelivering: results from our large-scale AI productivity survey

A 1,750-person survey of tech workers finds AI unambiguously overdelivering: 55% say it exceeded expectations, more than half save at least half a day per week, and only 17.7% report disappointment.

Founders benefit most — 49% save 6+ hours weekly, 78% positive ROI — because they use AI for strategic thinking (decision support, ideation, vision) rather than production tasks. Designers benefit least: 45% positive ROI, 31% below expectations (triple the founder rate). Engineers show the most mixed quality results: 51% say work improved but 21% say it worsened, the highest "worse" of any role.

Usage diverges sharply. PMs focus on PRDs and prototypes; designers on research synthesis and copy; engineers almost exclusively on code. The biggest unmet demand is upstream: user research (+27pp gap for PMs), documentation and code review (+24–26pp for engineers), product ideation (+29pp for founders). AI is delivering on production; the next wave is helping people decide what to build.

ChatGPT leads for PMs (58%) and founders (72%). Engineers split three ways — Cursor (33%), ChatGPT (31%), Claude Code (29%) — with low switching costs and no consolidation. PMF testing via the Sean Ellis question surfaces category winners: Cursor and Claude Code for coding, Lovable for prototyping, Granola for meeting notes (~2.5x loyalty-to-usage ratio). GitHub Copilot shows weak retention despite early-mover status.

Downsides are near-universal: 92.4% cite at least one complaint, led by generic outputs (56%) and hallucinations (52%), creating a review burden that partially offsets time savings. On agents: only 14% are active users; even among them, workflows remain 75%+ human-directed; blockers are organizational, not technical. n8n leads with 219 platform mentions versus Zapier's 85.

We replaced our sales team with 20 AI agents—here’s what happened | Jason Lemkin (SaaStr)

B2B sales is problem-solving: help a qualified buyer understand why your product fits their situation. Jason Lemkin (SaaStr founder, EchoSign CEO at $100M ARR) cites a portfolio company that fell from $5M to under $1M ARR after founders fired the sales team because they disliked sales.

Hire your first two salespeople — always two, for A/B testing — after closing 10 customers yourself and once sales exceeds 20% of your time. The criterion: would you buy your own product from this person? The right hire watched your explainer video, can articulate the customer's problem, and wants to sell. Have candidates demo your product live; 98% won't prepare — that's the filter. Both need 2+ years of B2B experience from a harder-to-sell product; someone from an easier sell melts under complexity.

Don't hire a VP of Sales until two reps hit quota. Hiring earlier means asking one person to find PMF, be rep one and rep two, and scale simultaneously — near-certain failure. The VP must still want to sell; ask "what do you do your first 14 days?" Process-first answers disqualify.

Sales orgs scale on rules of eight: eight AEs per manager, eight SDRs per manager. Comp is 50/50 base/variable; ramp new reps at 100% commission for one quarter, then target a 4-5x close-to-take-home ratio.

Product leadership belongs in major deals — a VP of Product can commit roadmap on the spot; a VP of Sales cannot. Give sales a fixed quarterly budget (~10% of engineering capacity) for feature requests — forces prioritization over reacting to the loudest deal.

The 14-day trial standard came from Salesforce reps wanting same-month closes, not customer evidence; Slack and Canva built to billions on indefinite free tiers. Forced annual SMB contracts carry the same flaw. Low churn compounds over decades; short-term monetization destroys it.

Aishwarya Naresh Reganti + Kiriti Badam

AI products differ from traditional software in two fundamental ways: non-determinism (inputs and LLM outputs are both unpredictable) and the agency-control trade-off (every step toward autonomy is a step away from human oversight). These invalidate most intuitions built on classical software.

The practical consequence: build autonomy in deliberate stages. For a customer-support agent, V1 classifies and routes tickets with humans in control; V2 drafts suggested replies that agents edit, logging those edits as implicit training data; V3 resolves tickets end-to-end. Shipping V3 first leads to endless hot-fix spirals — Air Canada's agent hallucinated a refund policy and the company was legally bound by it. A UC Berkeley / Databricks study found 74% of enterprises named reliability as their top blocker.

This progression underlies their Continuous Calibration, Continuous Development (CCCD) framework. The development loop: curate a baseline dataset, build the application, set evaluation metrics. The calibration loop: after deployment, spot error patterns you didn't anticipate, fix them, design new metrics for emerging failures. Pre-deployment evals catch only known failure modes; production signals — implicit (regeneration, drop-off) and explicit (thumbs-up/down) — surface the unknown ones. Neither alone is sufficient.

Three team-level success factors: leaders who rebuild intuition hands-on (one CEO blocked 4–6 AM daily for AI catch-up, coming back with questions for expert round-tables); a culture of augmentation over replacement-fear so subject-matter experts engage; and workflow obsession — knowing where deterministic code, ML models, and LLMs each belong instead of defaulting to agents for everything.

On architecture: supervisor-plus-subagent patterns work; flat peer-to-peer multi-agent protocols don't, because guardrails must be replicated everywhere. Coding agents outside the Bay Area are underrated — adoption is low despite demonstrated value. The broader moat argument: companies winning today accumulated painful iteration knowledge that can't be purchased or replicated quickly.

The power user’s guide to Codex: parallelizing workflows, planning techniques, advanced context engineering tips, automating code reviews, and more | Alexander Embiricos

Codex is not a coding autocomplete but a sandboxed shell agent designed to receive hard, real bugs — not toy tasks. The growth unlock from the original cloud-only product to a local VS Code extension and terminal tool came from recognizing that developers need interactive, paired feedback before trusting an async agent they can't see.

The architecture is three-layer: model + API + harness, all co-tuned. Codex uses only the shell (no semantic-search or bespoke tool calls), enabling faster specialization. Long-running sessions — overnight or 24-hour — rely on "compaction," where all three layers coordinate to summarize context before the window overflows and continue in a fresh one.

The Sora Android app is the clearest benchmark: two or three engineers built it in 18 days to internal launch, went public 10 days later, became the top app in the app store. Codex read the iOS app, produced a work plan, and implemented the Android port while examining both codebases simultaneously — porting across a known API surface is a particular strength. The Atlas browser team saw comparable gains: what previously took two to three engineers two to three weeks now takes one engineer one week.

The bottleneck the team is addressing is code review, not code generation. Writing code is the enjoyable part of engineering; reviewing AI-generated code at volume is the friction. Codex already runs automated review on OpenAI's training infrastructure, catching configuration mistakes. The next step is Codex babysitting training runs — watching loss curves, detecting anomalies, acting without waiting for a human to check dashboards.

On AGI timelines: the binding constraint is human review speed, not model capability. Unlocking productivity hockey sticks requires agents that validate their own output. Greenfield stacks could approach meaningful autonomy within a year; legacy systems will take years of incremental integration.

The non-technical PM’s guide to building with Cursor | Zevi Arnovitz (Meta)

Non-technical PMs can ship real software today using a structured Cursor workflow. Zevi Arnovitz, a PM at Meta with zero technical background, built StudyMate (a live, paying quiz app) solo on weekends and localized it from Hebrew to English in two days.

The workflow runs through six custom `/commands` saved in the codebase. `/create-issue` captures ideas mid-session into Linear via MCP without breaking flow. `/exploration-phase` has Claude read the codebase and map affected files before any code is touched — the step that separates structured building from pure vibe coding. `/create-plan` outputs a markdown file with status-tracked tasks; the format also lets work split across models (Cursor Composer for speed, Gemini for UI). `/execute-plan` runs the build. The most distinctive stage is three-model peer review: `/review` has Claude audit its own output, then Codex and Composer independently review the same diff. Their findings go back to Claude via `/peer-review`, framed as competing "dev leads" — Claude must either defend choices as by-design or fix them. `/update-docs` then patches markdown context files so future agents make fewer mistakes in the same area.

Each model has a distinct character: Claude is communicative and opinionated (the CTO); Codex is uncommunicative but solves the hardest bugs; Gemini produces strong UI despite a terrifying internal process.

When Claude makes a recurring error, the fix is asking what in its system prompt caused it, then updating that prompt — permanently raising the floor instead of repeatedly patching the same bug. The `/learning-opportunity` command applies the 80/20 rule to explain unfamiliar concepts in context, turning each session into deliberate skill-building.

At larger companies, the prerequisite is making the codebase AI-native through rich markdown context files. PMs can then handle contained UI changes, open a PR, and hand off for final review. Database migrations stay with engineers.

ChatGPT apps are about to be the next big distribution channel: Here’s how to build one

ChatGPT is becoming a transactional layer, not just a referral step — users can book flights, browse hotels, and order food through interactive widgets rendered inside the chat. The distribution opportunity is comparable to the 2008 App Store or early SEO: Adobe, DoorDash, Canva, Expedia, and Zillow are already live. Early movers compound their advantage because users build intent-to-app habits ("apartment search" → Zillow) that are hard to displace.

The architecture has three parts: the ChatGPT conversation, an MCP server exposing named tool functions (like `search_restaurants`, with plain-English descriptions ChatGPT uses to decide when to invoke them), and a React widget in a sandboxed iframe. Tool descriptions act as SEO metadata — precise wording determines when ChatGPT surfaces your app. Widgets can call tools directly via `window.openai.callTool()` without re-involving the model. One widget renders per turn; flows are sequential. Three display modes exist: inline cards, fullscreen, and picture-in-picture.

To build: an MCP server (Node.js or Python SDK), a widget, hosting on Vercel or Railway, and OpenAI app directory registration. Replit or the purpose-built Chippy tool are fast starting points. Simple apps take one to two days; complex multi-tool apps take a few weeks.

Marc Andreessen: The real AI boom hasn’t even started yet

The real AI boom hasn't started because AI is entering an economy that barely changed in 50 years. U.S. productivity growth ran at half the post-war rate and a third of the 1870–1940 rate — the physical world is frozen by regulation, cartels, and politics. Even tripling productivity growth only returns to the 1900 pace, which that era experienced as opportunity, not displacement.

Demography reinforces this: birth rates are below replacement in the U.S., Europe, and China; immigration is tightening. Without AI, depopulation means contraction. Mass unemployment would require 10–50% annual productivity acceleration — historically unprecedented. The realistic path is price deflation in affected sectors, equivalent to a purchasing-power raise for everyone.

Jobs outlast tasks. Product managers, engineers, and designers are in a Mexican standoff — each believes AI lets them absorb the other two roles, and they're largely right. Scott Adams's insight applies: competence at two or three domains is worth more than double because the combination is rare. The prescription is an E-shaped skill set: use AI as teacher — ask it to train and quiz you, not just produce output.

The most AI-native founders are questioning whether company structure needs to persist. Wave one: AI redefines the product. Wave two: 100 coders become 10 doing 10x output. Wave three, nascent: the founder orchestrates a nearly all-AI operation — Bitcoin approximated a one-person trillion-dollar outcome.

On moats, Andreessen is openly uncertain. Within 18 months of ChatGPT, Google, Anthropic, Meta, xAI, and Chinese labs all had competitive models. Claude Code built Cowork in a week and a half — impressive and frightening for defensibility. Industry structure in five years is unknowable.

Human IQ caps around 160; frontier models already test at 130–140. Human equivalence is a footnote — 200+ IQ models are coming, and more Einsteins is good news.

How to build AI product sense

Using Cursor for non-technical daily work — strategy, prioritization, writing — teaches AI product sense faster than years of ChatGPT, because coding agents make the machinery visible. Consumer UIs hide how AI works; coding agents expose it: you see the reasoning, the tool calls, the context window filling, and hit the same walls engineers hit in production.

Cursor is three familiar tools fused: a chat interface, a text editor, and a file explorer. Two differences from ChatGPT projects matter: you drag specific files into each chat, and the AI edits files directly — value lives in documents, not conversation history. LLMs can only produce text; Cursor plays "handyman," executing tool calls (read_file, search_replace) and returning results. MCP is a standardized connector so SaaS tools like Figma or Linear build one integration instead of one per model.

Model comparison builds real opinions. Claude refused Disney song edits on copyright grounds; OpenAI stumbled on the apply_patch tool call. Tool-calling skill is independent of reasoning quality — a model can be smart but clumsy with tools.

Building a personal productivity OS (Knowledge/, Tasks/, GOALS.md, AGENTS.md) makes three concepts tangible. RAG is "look things up before answering" — the agent searches files and prepends relevant content. Memory (AGENTS.md) is a text file prepended to every conversation, consuming tokens permanently; what belongs there versus on-demand retrieval is a design decision. Context engineering is the deliberate choice of what fills that window — memory, RAG, tool definitions, and conversation history all compete for the same space.

Context rot is the practical ceiling: performance degrades well before the window is full, hitting hardest in precision tasks. Watching Cursor's token counter on real work builds intuition benchmarks cannot. Once you can decompose any AI product into text, tools, and results flowing back into more text, the FOMO dissolves.

The rise of the professional vibe coder (a new AI-era job) | Lazar Jovanovic (Professional Vibe Coder)

Non-technical builders have a structural edge: not knowing what's "supposed" to be impossible means they just try it. Lazar Jovanovic, Lovable's first official vibe coding engineer, ships internal tools and public products using AI with no coding background. Lovable's community manager generated video in the tool before that was a feature; Jovanovic shipped Chrome extensions engineers said the platform couldn't build.

Once building is solved by AI, clarity is the only bottleneck. He spends 80% of time planning, 20% executing. His parallel-prototype method runs four simultaneous builds with escalating specificity — voice brain-dump, typed prompt, Mobbin/Dribbble screenshots, then code snippets from 21st.dev (AI reads code more precisely than English). The winner is obvious quickly; more tokens upfront saves hundreds downstream.

Context-window management follows. Vague prompts waste tokens on search and apology-writing instead of fixing. His PRD stack — masterplan.md (intent and feeling), implementation plan (build order), design-guidelines.md (with CSS to prevent AI over-creativity), user journey map, tasks.md — gives the agent everything before it acts. A rules file tells it to read all docs first and report back. Prompts shrink to "proceed with next task."

When things break: let the tool self-correct; add console logs to surface what it can't see; import the repo via Repomix into Codex or Claude for diagnosis; revert and re-prompt. After each fix, ask the agent how to prompt better and encode that into rules.md.

Career path: build in public on LinkedIn, give away every workflow, apply by shipping a Lovable app instead of a resume. Do the job before anyone hires you for it.

AI amplifies existing judgment — it cannot manufacture it. Deterministic roles (translation, data work) face high replacement. Design is his predicted next winner: it is emotional and models are not. Manual coding becomes calligraphy — rare, admired, no longer load-bearing.

Building AI product sense, part 2

Meta's new "Product Sense with AI" interview judges candidates on handling model uncertainty — noticing guesses, asking right follow-ups, deciding under imperfect information — not prompting skill.

AI features break not in demos but when real users bring messy inputs. Confronted with chaos, models invent structure: fed a chaotic Slack thread, a model hallucinates a roadmap, assigns wrong owners, and turns offhand comments into commitments.

Marily Nika's weekly ritual (under 15 minutes) runs three probes: ask the model to do something obviously wrong (hallucination patterns); ask something ambiguous (semantic fragility — understanding words but missing intent); ask something deceptively complex like "group 40 bugs into themes" (where reasoning first breaks). Each probe becomes a design requirement.

Before shipping, define minimum viable quality (MVQ): acceptable bar, delight bar (8-9 of 10 attempts without retry), and do-not-ship bar, calibrated to error cost and phase. Estimate the cost envelope: at $0.02/call, meeting summaries cost $0.30/user/month — viable; $5/user/month is a business problem.

Where behavior breaks, add system-prompt guardrails: ask instead of guess, offer choices when context is long, flag invented structure, impose light format to cut variance. One rule — "only assign an owner if someone explicitly confirms" — eliminated the biggest trust failure in a Slack-summarization product.

Sherwin Wu V2

At OpenAI, 95% of engineers use Codex daily, 100% of PRs are reviewed by Codex, and engineers using it heavily open 70% more PRs than those who don't. The engineering job has become management: engineers run 10–20 parallel Codex threads, steering agents rather than writing code. One internal team operates a 100%-Codex codebase with no manual fallback; agent failures are almost always context failures — the fix is encoding tribal knowledge into comments, .md files, and AGENTS.md.

AI amplifies top performers most, so managers should spend over half their time on the top 10% of engineers — anticipating blockers before they land.

The "one person billion dollar startup" undersells second-order effects. If one person can run a billion-dollar company, a hundred more can each run $10M–$100M companies, producing a proliferation of vertical B2B SaaS — bespoke software for narrow use cases. VC returns may compress, but this is excellent for high-agency individuals.

Customer feedback in AI can trap builders in a local maximum. Scaffolding customers want improved today — vector stores, agent frameworks — gets absorbed by the next model. Build for where models are going. A product at 80% of needed capability may click into something excellent when the next generation ships.

Most enterprise AI deployments are top-down mandates that fail. What works: an internal tiger team (often technical-adjacent non-engineers) that builds real workflows and creates bottom-up excitement; without this, employees don't know what to do even when AI is on their performance reviews.

Business process automation — repeatable, procedure-driven work in support, utilities, and operations — is larger and more neglected than Silicon Valley discussion reflects. OpenAI's platform stance (every model in the API, no competitor blocking) traces to the mission clause about spreading AGI benefits to all of humanity, since reaching everyone through one product is impossible.

How to do AI analysis you can actually trust

AI-generated customer analysis always looks confident — clean themes, punchy quotes, tidy summaries — even when the underlying evidence is fabricated. Caitlin Sullivan, a user-research consultant with 2,000+ hours testing AI workflows, identifies four failure modes and a prompt fix for each.

Invented evidence. LLMs generate quotes statistically, not by retrieval. ChatGPT stitches quotes from multiple sources, especially when asked for brevity ("≤12 words"). The fix has two parts: a QUOTE SELECTION RULES block (start where the thought begins, include hedges and qualifiers, cite participant ID and timestamp, never combine statements) followed by a QUOTE VERIFICATION prompt classifying each quote as VERIFIED / PARAPHRASE / NOT FOUND. In practice, most of ChatGPT's initial quotes fail that check.

False or generic insights. LLMs default to consensus — themes like "price is a factor" that apply to any product. The fix is structured context loading with four components: project context (the specific decision), business goal, product context (domain knowledge the model lacks), and participant overview (who is speaking). With that framing, the model moves from generic theme lists to per-participant verdicts and actionable tallies — in the worked example, 30% of churned users would be retained by a screen.

Signal that doesn't guide decisions. "22 respondents mentioned a screen" is not actionable. The fix is few-shot calibration: provide labeled examples at each level of a decision-specific scale with explicit rationale, so the model distinguishes "screen directly solves this" from "app UX fix is cheaper" from "screen is irrelevant to their actual complaint."

Contradictory insights. AI never flags its own uncertainty. A final VERIFICATION PASS — checking quotes, surfacing within-participant contradictions, flagging thin-evidence findings — catches errors the first pass presented with equal confidence. The trade-off: fifteen minutes of verification now versus six months building the wrong thing.

Boris Cherny

100% of Boris Cherny's code has been written by Claude Code since November. As head of Claude Code at Anthropic, he ships 10–30 pull requests daily with five agents running in parallel. The curve went from 20% in February to 30% by May to full automation by fall.

Claude Code started as a solo terminal hack that got two internal likes and was not an immediate external hit. Growth came through latent demand: users were forcing a coding tool to grow tomato plants, analyze genomes, recover wedding photos from corrupted drives, and run SQL queries. Unintended use at scale is the strongest signal to build a dedicated version. Cowork came from six months of observing that pattern.

The model now proposes what to build, reading bug reports and telemetry to surface fixes. Anthropic 4x'd engineering headcount since launch while achieving 200% productivity gains per engineer — compared to a few percentage points Cherny saw from years of code-quality work at Meta.

Core product advice: don't constrain the model with rigid orchestration. Give it tools and a goal; let it decide the path. Scaffolding might lift performance 10–20% but those gains vanish with the next model. Build for the model six months out — Claude Code felt underwhelming early because the model wasn't ready; when Opus 4 landed, growth went exponential.

Safety has three layers: mechanistic interpretability (monitoring neuron activation including deception-related patterns), controlled evals, and real-world deployment. Claude Code ran internally four months before release to study that third layer.

His analog: the printing press. Sub-1% of Europe were scribes before Gutenberg; in 50 years, printed volume exceeded the prior thousand years; literacy reached 70% globally over two centuries. His prediction: "software engineer" fades by end of 2025, replaced by "builder" — everyone codes, everyone manages product.

How to use AI for your next job interview

Interview prep fails not because candidates don't work hard enough, but because there's no feedback loop — companies don't explain rejections, so you can't tell if the problem is your resume, your stories, or bad luck. Top performers in a study of 30+ tech professionals fixed this by building AI workflows: Greg fed transcripts to Claude for line-by-line scoring on answers he thought went well; Ella had ChatGPT surface the exact resume-to-JD gaps a hiring manager would flag before any human saw it; Sean ran pre-interview simulations to test which stories landed. Logan used these methods to land a senior architect role at Anthropic after eight years without interviewing — in two weeks.

That research was encoded into a free Claude Code coach (GitHub: noamseg/interview-coach-skill), persisting state across sessions. Key commands: *prep [company]* generates a one-page brief with 7–10 predicted questions, story mapping, and hiring-manager concerns; *analyze* scores a pasted transcript on substance, structure, relevance, credibility, and differentiation (1–5), diagnoses root-cause patterns like "narrative hoarding," and produces a side-by-side rewrite of your weakest answer; *mock* runs a full 4–6 question interview without mid-session feedback; *negotiate* delivers exact salary scripts with fallback language.

The coach tracks the gap between self-ratings and actual scores over time — the calibration that surfaces blind spots. Default AI prep produces polished-but-forgettable answers; the counter is adversarial prompting ("find reasons to reject me") and leading with counterintuitive, experience-earned insights. Covert recording tools like Granola raise consent issues in two-party-consent states; stories polished beyond real experience will surface under follow-up.

Jenny Wen

The traditional design process — discover, diverge, converge, produce beautiful mocks — is dead, not by designer choice but because engineering velocity forced the change. When engineers spin up working features in hours via Claude Code or v0, designers can no longer lead with prototypes. Jenny Wen, who led design for Claude at Anthropic and directed design at Figma (FigJam, Slides), describes the role splitting into two kinds of work: supporting execution (reviewing, guiding, polishing what engineers ship) and setting near-term vision — now a 3–6 month prototype, not a 2-year deck.

Her own time has shifted from 60–70% mocking to 30–40%, with a new slice implementing polish directly in code. Figma still earns its place for exploring 8–10 directions simultaneously — a branching mode linear coding tools don't support once you're committed to one path.

On AI replacing design judgment: Claude isn't hireable yet — good at first passes, not at producing anything special. Taste will improve, but humans remain accountable for what to build, not just execution.

Three archetypes Wen hires for: block-shaped strong generalists (80th-percentile across multiple skills), deep T-shaped specialists (top 10% in one domain), and craft new grads — fast learners unburdened by entrenched process. Advice to young designers: build real things, don't wait for permission.

On management: pure people-management is fading; managers must now provide direction and stay technically current. Wen stepped back into IC work at Anthropic to retain empathy for how much the role has changed. Psychological safety and high standards aren't in tension — safety makes high standards easier to apply.

The legibility framework (from VC Evan Tana) guides her internal scouting: illegible ideas with energy are worth diving into. Claude Cowork's skills framework originated this way — a dense internal prototype nobody could fully articulate, but with clear gravitational pull.

From skeptic to true believer: How OpenClaw changed my life | Claire Vo

Running nine specialized agents on OpenClaw is not AI psychosis — it is the same logic as nine Slack channels: each agent gets bounded context so it doesn't drown in irrelevant noise. Claire Vo, CPO-turned-founder, started as a vocal skeptic (first install deleted the family calendar) and became a convert after recognizing product-market-fit: users complain it's broken, not that it's useless.

Her key unlock was the employee-onboarding mental model. Each agent gets its own Gmail address and calendar delegation — never the password to your primary account. Sam replaces a human contractor who spent ten hours a week on CRM work: nightly PLG scan, Exa enrichment, soft outreach emails, large-enterprise prospects held for founder review. Howie preps podcast briefings. Finn handles the family calendar — three kids, two schools, three basketball leagues — pinging the group chat each afternoon to sort pickup duties. Sage project-manages a Maven course launch, nudging Claire to post on LinkedIn and filing research into the course repo.

Practical setup: clean, separate machine; enable macOS screen sharing to avoid a dedicated monitor; give the agent its own local admin account; use Telegram via BotFather as the control channel; use flagship models (Opus, Sonnet 4.6) for security hardening against prompt injection. Install Claude Code on the same machine as a surgeon — it reads config files and fixes them in plain English.

Browser use is the sharpest edge: the open web is hardened against bots, so fall back to web-search APIs (Brave ships by default) when native control fails. For memory issues, prompt the agent to write action items before context compresses. The soul file and tools.md are the two files worth hand-editing.

The deeper point: management experience transfers directly. Scoping roles, building trust progressively, not micromanaging implementation — all predict agent success better than technical skill.

An AI state of the union: We’ve passed the inflection point, dark factories are coming, and automation timelines | Simon Willison

November 2025 was the inflection point: GPT 5.1 and Claude Opus 4.5 crossed a threshold where coding agents stopped producing buggy approximations and started reliably delivering what you asked for — spin up four in parallel, step away, return to working software. Willison (co-creator of Django, coiner of "prompt injection") calls professional use "agentic engineering" — distinct from vibe coding, which he limits to personal tools where bugs harm only yourself.

The logical endpoint is the "dark factory." StrongDM already runs one: nobody writes code, nobody reads code. QA is a swarm of simulated employees in a hand-built Slack/Jira/Okta simulacrum at $10,000/day in tokens. The same agents now do credible penetration testing — Anthropic discovered 100 Firefox vulnerabilities and responsibly disclosed them.

Experienced engineers benefit most because agents amplify existing judgment. Mid-career engineers are most exposed: not senior enough to steer agents well, already past beginner onboarding benefits. Start every project with a thin skeleton file (one passing test, correct indentation) rather than verbose instruction docs — agents pattern-match from code faster than prose. Use red/green TDD; agents tolerate the discipline better than humans, and the test suite prevents regressions as the codebase grows.

The structural security risk is the "lethal trifecta": private data + malicious-instruction exposure + exfiltration route. An email agent that reads your inbox and can reply satisfies all three. "Prompt injection" has no reliable fix — unlike SQL injection, you cannot enumerate every attack surface across every human language. Detection at 97% is a failing grade. The best partial mitigation is Google DeepMind's CaMeL: a privileged agent plus a quarantined one that tracks taint and escalates only genuinely risky actions to a human. His prediction: the normalization of deviance that preceded the Challenger O-ring failure has been accumulating in AI — a headline breach is coming.

Anthropic’s $1B to $19B growth run: how Claude became the fastest-growing AI product in history | Amol Avasare

Anthropic grew from $1B to $19B ARR in 14 months — 10x year-on-year since 2023 — and the growth team takes only partial credit. Research, Claude Code, and go-to-market contributed more. Amol Avasare (head of growth, hired via cold email to CPO Mike Krieger) leads a ~40-person org that spends 70% of its time on "success disasters": scaling failures caused by growth coming too fast.

Unlike conventional growth orgs that skew 60-70% toward small experiments, Anthropic flips the ratio toward large bets. The reasoning: if product value will be 100-1000x higher in two years from model improvement, micro-optimizations capture a shrinking share of the opportunity. The Chrome extension underpinning Cowork and Claude Code was a growth-team bet when no one else was doing it.

Activation is the hardest AI growth problem. "Capability overhang" — models advancing faster than users update their mental models — leaves people asking a near-AGI for the weather. The consistent fix across Mercury, MasterClass, and Anthropic: add intentional friction. Ask users who they are, route them to the right feature. Mercury spent a full quarter on onboarding quality rather than conversion metrics; it was Avasare's highest-impact quarter anywhere.

Anthropic's internal automation project, CASH (Claude Accelerates Sustainable Hypergrowth), uses Claude to identify experiments, build them, test, and analyze results. It performs at junior-PM level on copy and minor UI changes — only viable since Opus 4.5, improving with 4.6. Separately, Avasare runs scheduled Cowork agents that scan 20-25 charts each morning, detect Slack misalignment across teams, and file his expenses.

The coding focus traces to a 2021 internal memo by co-founder Ben Mann: better coding models accelerate research, producing better models — a recursive loop. On team structure, projects under two engineering weeks are owned by product-minded engineers; PM accountability kicks in above that threshold.

How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)

Anthropic ships features in days not because of model superiority but because of deliberate process. Cat Wu, Head of Product for Claude Code, credits three mechanisms: specific user-and-goal definitions so engineers decide without PM sign-off; labeling nearly everything "research preview" to drop commitment overhead; and a standing launch room where engineering, docs, and PMM publish a feature announcement the next day.

PRDs exist only for ambiguous or infrastructure-heavy work. The default is weekly metrics readouts plus a principles doc enabling autonomous decisions across five PM clusters: research, developer platform, Claude Code/Cowork, enterprise, and growth.

The most important emerging PM skill is calibrating to current model capability, not the AGI-perfect future. Wu's practices: ask the model to introspect on its failures; find five engineers who can articulate why a harness works; build targeted evals of ten good examples. Code review illustrates this—tested on earlier models, found too inaccurate, and only shipped as a PR gate when Opus 4.5/4.6 and Sonnet 4.6 were reliable.

Model upgrades regularly obsolete product scaffolding. The Claude Code to-do list forced earlier models to track call sites during refactors; Opus 4 made it unnecessary. Each model launch includes a system-prompt audit to remove interventions the new model no longer needs, and may unlock features that weren't accurate enough before.

Cowork's practical value is connecting Slack, Gmail, Calendar, and GDrive into a briefing and deck-generation engine. Wu generated a 20-page conference deck overnight—reviewing the proposed outline once before Cowork built the full output against Anthropic's design template.

The long-horizon roadmap extrapolates from one task to hundreds of remote parallel tasks, with interfaces shifting to help humans spot-check work. Product consistency is the explicit trade-off for velocity. The durable skill is product taste: knowing which of thousands of feature requests is worth building, and what the right UX is.

How to build a company that withstands any era | Eric Ries, Lean Startup author

Successful companies are destroyed not by competition but by their own success — the more valuable the business, the greater the temptation to extract rather than create value. Eric Ries calls this "financial gravity." *Incorruptible* is a structural diagnosis: the standard Delaware charter obligates boards to accept the highest acquisition bid regardless of buyer intent. Vectura, a UK inhaler-therapeutics company, was acquired by Philip Morris for £1.1 billion despite public outcry; Philip Morris took a $900 million write-down and dissolved it within three years.

Only 20% of founders remain CEO three years after IPO under standard governance. Every advisor says it's too early at each stage until it's suddenly too late.

The first defense is "harder is easier": commit to a principle before you need it, and trust compounds. Cloudflare gave away SSL encryption — their top revenue driver — top-of-funnel grew an order of magnitude, and Cloudflare became a $70 billion company. Groupon tested two daily emails against one, confirmed the uplift, ended up sending eight, and destroyed the company.

The concrete minimum: incorporate as a Public Benefit Corporation — a two-page Delaware filing replacing "any lawful act or activity" with a stated mission. Anthropic added a Long-Term Benefit Trust at Series C: outside AI safety trustees who appoint board directors and hold no equity. That is why Anthropic could decline a $200 million Pentagon contract and why refusing an unsafe model release is institutionally enforced, not merely cultural.

Ries's umbrella term for these structures — Perpetual Purpose Trusts, employee ownership trusts, nonprofit foundations — is "spiritual holding company." Such companies are six times more likely to reach year 50. The parallel to AI alignment is deliberate: corporations are the oldest form of emergent intelligence, and builder values propagate into structure exactly as architectural decisions flow through software.

Why we’re at the beginning of the AI hardware boom | Caitlin Kalinowski (ex–OpenAI, Meta, Apple)

AI acceleration in software is approaching saturation — the next frontier is the physical world. Caitlin Kalinowski (hardware lead at Apple, Meta/Oculus, OpenAI) argues a hardware boom in robotics, manufacturing, and physical AI is now beginning.

Hardware's irreducible difficulty: engineers get four or five total "compiles" across a product's life. Each iteration takes months; mass-production is the final compile. This demands front-loaded testing and working the riskiest parts first. KPIs — cost, weight, pixels-per-degree — must be locked early; shifting a price target from $300 to $150 mid-program wastes most prior work.

Supply chain is the deepest constraint on scaling robotics. The chain runs from rare-earth magnets to actuators (motors) to subassemblies, with each layer outsourced over 25 years to China, Japan, and Korea. Actuators are the chokepoint; the US has almost no domestic actuator industry. DRAM prices spike as cost-insensitive AI data centers compete with consumer hardware for constrained supply. Pre-buying is the only near-term hedge.

Humanoid robots remain advanced prototypes. Scaling requires solving supply chain, reliability, and domestic factory capacity. Most manufacturing is better served by dedicated task robots, which already displace humans on PCB and mechanical assembly lines. Robots strong enough to do meaningful work carry "no human within three feet" warnings.

VR was a platform investment: SLAM, depth sensing, and space-perception research now feed robotics and autonomous vehicles. AI helps with PCB routing and analysis, but real CAD requires physical reasoning — friction, contact, weight — that LLMs cannot do. World models may be the necessary base.

On hiring: mix experienced generalists with AI-native engineers in their 20s who build everything AI-first — they teach senior engineers how to think. Mission alignment unifies cross-disciplinary teams better than credential match.

Kalinowski left OpenAI after the DoD partnership announcement, citing governance speed and lack of defined guardrails.

The AI paradox: More automation, more humans, more work | Dan Shipper

More automation produces more work for humans, not less — the core paradox Dan Shipper (CEO of Every) argues from running a company that doubled headcount last year because of AI adoption.

Work will bifurcate into two surfaces. First: companies will converge on a single "super agent" in Slack rather than per-person agents. Personal agents fail because they need a dedicated caretaker — once maintenance stops, quality degrades. Shopify and Ramp have moved to the one-company-one-agent model kept running by a forward-deployed engineer. Second: the primary work surface shifts to an AI harness (Codex, Claude Cowork) with an embedded browser — email, documents, and analytics run inside the agent. Shipper reached inbox zero for 10 days by having Codex aggregate emails, render a summary page, and execute dictated instructions per thread.

The SaaS apocalypse is wrong — agents increase SaaS usage and request volume. Agents use the user's tokens rather than the vendor's — restoring margins. Every's SaaS spend is up year-over-year despite heavy automation.

On benchmarks: Shipper built a "senior engineer benchmark" from his own vibe-coded app that crashed at launch. Human engineers score high 80s/100; most models score ~30; GPT 5.5 with Opus 4.7 planning reached 62. The gap is that models fix what you tell them to fix — a human senior engineer looks at a codebase and decides unprompted to rewrite it. That higher-frame judgment is what rising benchmarks don't capture.

Models make yesterday's human competence cheap and commoditized. Humans stay ahead by using frozen competence to make something novel. Who thrives: PMs with product sense who go all-in on AI coding (Shipper cites a non-technical PM who ships faster than most engineers); full-stack designers who can now build what they design without handoff. Each model drop creates frontier room for people who keep riding forward.

Max Schoenig

Agency — not skill — is what separates people as AI removes skill barriers. Max Schoening, head of product at Notion, argues that before AI you could always blame a capability gap. That excuse is gone. What matters is whether someone treats the world as malleable and acts.

Designers and PMs should code to think in the material, not to ship features. Notion built an LLM-friendly playground to get non-engineers past terminal fear. The goal is understanding agent loops — a PM who can't reason about agent loops is less useful than one who designs them deeply without writing a line.

Malleable software — software working for users rather than its maker — is finally becoming real. Current apps are furniture you can't rearrange; the fix is communal, adaptable tools. Notion's AI agent works because context roams freely through a connected workspace instead of hitting walls. Static Figma prototypes fail for AI products because you must feel the interaction — Bret Victor's Stop Drawing Dead Fish.

Great products have one tiny exceptional core. Pull requests made GitHub; git push to Heroku master made Heroku; the Dropbox icon synced so reliably it doubled as a connectivity indicator. Adding features to compensate for a weak core never works. Schoening illustrates this with a 2014 competitor where he polished markdown folding for years while Notion launched with a broken editor — and won because the block model was right.

On AI's trajectory: models improve fastest at coding. Software eating the world simply accelerates — HR automates workflows, every domain becomes a coding problem. The SaaS apocalypse is exaggerated; "as a service" still means maintenance and expertise no one wants to replicate. Taste is learnable: the ability to predict whether an in-group will like something, built through reps with feedback, like training a model.

What Seven Years at Airbnb Taught Me About Building a Business

Culture, goal-setting, and focus compound into durable competitive advantage — Airbnb's growth to $30B proved this. Brian Chesky routinely doubled proposed goals and pushed for 10x; uncomfortable targets worked because ownership was always individual (a number without a name doesn't happen). Problem-framing mattered equally: vague statements sent simple projects in circles for months, while crisp problem statements let complex ones sail. The "Snow White" storyboarding process — envision the ideal user journey, identify emotional gaps, work backward — drove instant booking from 5% to a majority of reservations by reframing host reluctance into separable "can" and "want" problems. Org design is a product: autonomous cross-functional teams with one or two focused metrics outperform matrix structures. High-bar discipline extended to emails, docs, and hiring ("hell yes or no"). Narrowing team scope, then doubling down on the highest-opportunity quarter, consistently unlocked both impact and morale.

The Secret to a Great Planning Process — Lessons from Airbnb and Eventbrite

Bad planning fails not from unclear strategy but from undefined roles — who contributes, when, and who decides. Rachitsky (Airbnb) and Gilbreth (Eventbrite, 10+ years each) propose the W Framework: Leadership shares high-level Context; Teams respond with Plans; Leadership Integrates into one unified plan; Teams confirm Buy-in. The four-step sequence forces explicit handoffs, preventing the misalignment and abandoned plans that plague most annual cycles.

This Week #6: Cultivating good relationships with distributed co-workers, building trust to accelerate growth, and leveling up as a PM when you have extra time

Remote miscommunication forces engineers and designers to redo finished work. Hiten Shah's fix: write down every decision mid-call, screenshare directly in the ticketing or spec tool so feedback lands in real time, and postmortem every ship to iterate the process itself. For fintech, trust unlocks growth — surface social proof, certifications, and guarantees early; lean on referrals as the strongest top-of-funnel lever; and email users Wealthfront-style "here's what we earned you" reminders to arrest churn. A PM on sabbatical should pursue the skill long postponed, read to named gaps, and write to consolidate what they've learned.

Setting goals

Goals slot between strategy and roadmap — they confirm whether strategy is executing. At Airbnb, the north star was nights booked; when supply became the binding constraint, the metric became "new listings with at least one booking." Pick the metric by mapping levers in the growth model, then set threshold tops-down (what the business needs) and bottoms-up (what the team can deliver) — Brian Chesky's bar: "uncomfortable, but not impossible." Skip this and the silver-burrito trap hits: everyone imagines a different goal inside the same project.

Communicating bad news - Issue 26

When a project fails, the PM's job is to be the solution, not just the messenger. Tell stakeholders before they hear it secondhand; briefly recap why leadership originally approved the work, then share results; highlight any genuine learning. Most critically, arrive with a clear recommendation — iterate, kill, or ship with a claw-back plan. Keep identity separate from outcomes: projects fail, PMs don't have to.

Autonomy vs. direction - Issue 35

New managers expect more power to mean less persuasion — it means more, now including your own reports. Five moves resolve the tension: align on facts first, since most disagreements are information gaps (an ROI chart resolved an Airbnb referral dispute); distinguish consulting and collaboration from consensus, which Apple, Tesla, and Amazon avoided on their best products; overrule when you feel strongly — Amazon's "disagree and commit" shows why forcing full buy-in slows decisions; apply the 90/10 rule, letting the team run a doubt-worthy experiment as a morale valve; and always hold a point-of-view, because without one you can't frame the right questions.

Getting better at product strategy

Product strategy has three failure points — determining it, articulating it, and acting on it — and you can get two right while the third kills you. At Airbnb, the Instant Book team spent months tuning host reminders before reframing: if you can see a listing, you should be able to book it instantly. That vision anchored a three-track strategy (host tools, host incentives, guest nudges) that took bookings from 5% instant to 80% in two and a half years. Good strategy is problem-oriented, insight-driven, and focused on a small number of high-leverage bets (Rumelt). Articulation demands the rule of three and a memorable framework — the CAN/WANT split (features vs. incentives) is one example. Execution means repeating the strategy at every touchpoint, running all prioritization through it, and treating it as a living hypothesis that evolves with evidence.

Fostering a culture of experimentation

Airbnb's design reputation obscures its real edge: its tenth employee was a data scientist (Riley Newman), whose founder relationships made data a strategic asset from the start. Adopting "nights booked" as the north-star forced leaders to run experiments rather than ship tasks — accountability demanded knowing why things worked. Once the DS team hit ~25 people, it decentralized into product teams, shaping roadmaps rather than doing post-hoc analysis. Data was framed as "the voice of customers at scale," converting skeptics. Scrappy experimentation tooling followed, built when manual analysis grew too slow. The payoff: every attempt to add "sort by price" collapsed bookings, because cheap listings have worse hosts — intuition failed repeatedly; data caught it.

Three myths about sabbatical programs

Well-designed sabbatical programs retain employees rather than losing them: 80% return more loyal and creative. Uber's failed policy was a symptom of Kalanick-era dysfunction, not a cause. Retention design splits into two cycles — shorter sabbaticals around year four to catch high performers eyeing the exit (McKinsey externships, Endeavor country rotations), and long-cycle programs like Brighton Jones's fully-paid six-month sabbatical at year ten for employees you want for decades.

The "company falls apart" fear misreads the diagnostic value: when African Leadership Academy's founder Chris Bradford took leave before its biggest fundraising push, the resulting financial crisis exposed a dangerous single-point-of-failure dependency that would have surfaced catastrophically later anyway.

The subtlest myth is "I'm fine." Functional workaholism — maintaining productivity while depleting reserves — is invisible until six to eight weeks into real disconnection. Leaders who skip sabbaticals suppress uptake through the "work devotion norm"; studies show employees only take leave when managers do. Minimum viable policy: paid, three months, with 90-day off-ramp and reintegration.

Healing your co-founder relationship

Co-founder tension almost always lives in interpersonal soft skills, not strategy. Stanford GSB's Carole Robin prescribes a four-step sequence: first tune inward to identify your actual feelings (disappointed, resentful, frustrated); then distinguish organizational disagreements—direction, exit timing—from behavioral ones, which are harder and more personal; next look for shared responsibility rather than blaming "the other," anchoring on a superordinate goal like "what's best for the company"; finally "stick to your reality"—describe behavior and its impact on you without crossing the net to attribute motive or intent. Resolve with agreements treated as trial periods, with explicit plans for handling regressions.

Getting buy-in

Buy-in fails for three reasons: assuming your conclusion is obvious, feeling time-crunched, and unconsciously avoiding dissent. Shivani Berry (CEO, Arise Leadership) proposes three fixes. Co-creation beats persuasion: walk stakeholders through the data early, run a brainstorm on a straw-man draft, and let the group land on the recommendation — Berry's team reached the same search-feature conclusion she had, but everyone owned it. Presence means eliminating apologetic filler: Salesforce engineer Natalie dropped "sorry" for one week and colleagues immediately said her ideas landed stronger. Storytelling means cutting backstory to only what a stakeholder needs to decide, then leading with their stake in the outcome.

Five habits of highly annoying product managers

The five habits that most reliably erode team trust are: treating frameworks as gospel rather than context-dependent tools; prioritizing without explaining the why, so engineers question whether their work matters; calling everything "quick," which burns goodwill when estimates slip; chasing consensus instead of making a call at 70% information (Bezos's threshold); and over-speccing features, crowding out designer and engineer judgment. Fix: buy-in over perfect process, written rationale in PRDs, and conversation over documentation.

Product management career ladders

Two-thirds of companies run separate IC and manager tracks for PMs, with the fork at Senior PM (L6). The standard IC sequence: APM → PM → Sr. PM → Principal PM; manager track: Group PM → Director → Sr. Director → VP. Top evaluated competencies, in order: leadership, impact, scope, execution, communication. Over 80% have a Senior PM level; two-thirds have APM and Principal PM roles. Stripe, Airbnb, Lyft, and Facebook resist title inflation — Stripe keeps "Product Manager" at every IC level.

How to create a winning product strategy | Melissa Perri

The product owner role was never meant to be product management. Software developers invented it — no PMs attended the 2001 Agile Manifesto meeting — to answer one narrow question: what to build next? The Scrum Guide gave them a "product owner" to manage the backlog and write user stories. Not customer discovery. Not strategy.

Large non-tech companies — banks, insurance, pharma — adopted Scrum en masse in the 2010s. They took business analysts and project managers, put them through two-day certification classes, and declared transformation complete. SAFe (Scaled Agile Framework) amplified this with a large map prescribing how to coordinate hundreds of teams; executives loved it because it looked like a plan. What it produced: product owners spending 40 hours a week writing user stories for a login API that already worked, developers idling until backlogs were filled, quarterly big-room planning locked to roadmaps with no customer discovery. A Dutch water company went bankrupt because teams were so consumed in SAFe process work they couldn't ship the invoicing system needed to collect payments.

The distinction that matters: product owners are trained on process (stand-ups, sprint cadences, backlog grooming); product managers are trained on substance (customer interviews, market research, hypothesis testing, outcome measurement). The former doesn't survive outside Scrum; the latter scales from IC to CPO. Successful transformations — Capital One is the clearest case — ripped out SAFe and rebuilt around genuine product management, mixing experienced hires with retrained staff.

For product owners trying to transition: drop agile cadences from your resume. Describe the customer problem you solved, what metrics moved, what constraint you navigated. Push back on features by asking "what do we expect when this ships?" — that question forces an outcomes conversation. If no one in your org demonstrates good product management, move laterally or leave.

The nature of product | Marty Cagan, Silicon Valley Product Group

Most companies calling themselves product-led are actually feature factories — the gap is wide enough that "product manager" shouldn't be the same title in both. At a feature team, the PM is a project manager: write requirements, herd stories into sprint planning, measure output. At an empowered product team, the PM owns valuable and viable — the two hardest risks — while engineers own feasible and designers own usable. Any of the four failing kills the product, so "the PM defines the what, not the how" is incoherent. Only 10–15% of commercial software companies are real product companies.

Steve Jobs' 1995 Lost Interview provides the best explanation Cagan has found for why good companies drift. As a company grows, revenue comes from sales and marketing rather than product innovation, so those functions get promoted and celebrated. Good product people leave. The company stops doing real discovery and defaults to low-risk A/B optimization — "the beginning of the end." A second disease Jobs named: executives believe an idea is 90% of the work, when iterating through discovery toward a winning solution is everything. Customers don't buy the problem, they buy the solution — so spending most of discovery time on problem validation is a mistake.

A PM trapped in a feature team can run a one-quarter experiment: reverse-engineer roadmap items into problems to solve, get direct access to customers and stakeholders, and learn discovery techniques. Teresa Torres's *Continuous Discovery Habits* and Jake Knapp's *Sprint* are the practical guides.

Three things a PM must never delegate: direct access to customers, engineers, and stakeholders. Structures that sever any of these — product owners, mediating product ops roles — destroy innovation even when well-intentioned. Scaling with process (SAFe, repackaged waterfall) is the dominant wrong answer; scaling with leaders who coach is the only path.

How to get better at influence

PM influence depends on seven moves, not formal authority. Align asks to the other party's goals — at Airbnb, pitching the Superhost badge as a guest-conversion win (not a host metric) broke a search-team deadlock. Treat trust as a battery (Tobi Lütke's model): charge it through reciprocity and follow-through before drawing on it. Bring people on your thought journey early rather than rushing decisions — Stephen Covey's slow is fast. A track record of hitting goals compounds influence; Amazon's Leaders are right, a lot captures this. Lead with hard data first — the blind-review redesign won support by citing double-blind studies from other fields. Executive sponsorship transfers credibility but over-reliance hollows out your own voice. Finally, likability is a force multiplier on all the others: Cialdini finds persuasion scales directly with being liked.

Humanizing product development | Adriel Frederick (Reddit, Lyft, Facebook)

Diverse teams are faster because they eliminate the research delay. When Facebook's growth team included a Black Trinidadian PM, an Israeli engineering manager, a Brazilian tech lead, and engineers from across the world, debates that would take two weeks of user panels resolved in 15 minutes. Adriel Frederick was the first Black PM at Facebook; his upbringing in Trinidad — 35% Indian, 35% African, 25% mixed, schools that scrambled income levels — directly shaped product design. Knowing that in developing markets one phone serves multiple people and SIM cards rotate, he shaped Facebook's phone-number registration accordingly, fueling global growth monoculture teams would have missed. The business case for diversity is speed, not obligation.

On growth: "10 friends in 14 days" worked not because the number was magic — Zuck picked it to end endless debate — but because a concrete goal galvanizes organizations. The real work was grinding on three things: find the product, enter it, find your friends. The marginal user framework: find a country with high traffic but terrible conversion, then watch the worst-case user (feature phone, EDGE, far from a data center) to surface every failure at once. Data shows how bad things are; only watching the person reveals what's wrong.

On algorithmic products: those who say feed data to the algorithm discover it cannot see snowstorms, competitor moves, new taxes, or long-term effects. At Lyft, a PhD-consultant pricing model had to be scrapped because no human could intervene to change prices. The lesson: deciding who makes which decisions — human versus machine, and what interface enables that — is a first-order product design problem.

Two skills matter most at senior levels: organization design (clearing the path for the team) and empathy — the discipline of taking your own shoes off before stepping into someone else's.

How to fire people with grace, work through fear, and nurture innovation | Matt Mochary (CEO coach)

Fear gives bad advice — Mochary's core claim, tested in hundreds of prediction contests with CEOs without a loss. When afraid, a founder's predictions become exaggerated; an outside observer sees clearly. The method: make one low-stakes bet against the fear-prediction. After one loss, "I perceive you to be in fear" is enough to unlock action. Anger is a cover for pain, not a base emotion; "I perceive you to be in anger" pierces it without triggering defensiveness.

Firing is the skill most managers avoid, believing they harm the person. The reframe: you are freeing someone not needed in that role. Protocol: warn them seconds before so the amygdala is not ambushed; invite emotions; make them feel heard by naming what you imagine they are thinking ("I think you're thinking: screw you, this is bullshit — is that close?"); then become their agent — outbound outreach to your network, not a passive reference.

For layoffs, assign dollar-savings targets rather than headcounts (headcounts cause managers to cut cheapest, not lowest-impact). Deliver news in 1-on-1s, never group email. Same afternoon: all-hands for the stay team. Within 48 hours: one listening session per remaining employee. Full execution yields measurably better output within two weeks; skipping the listening step delays that to two months.

To innovate inside a large company, spin up a separate C corp with a small founder-mentality team — failed YC founders are ideal — reporting directly to the CEO, bypassing product and engineering review chains. Wei Deng at Clipboard Health runs two parallel teams per new product, one engineering-led and one manual, racing each other.

The energy audit marks two calendar weeks hour-by-hour green or red, then eliminates, delegates, or redesigns the reds. Target: 80% green. Mochary credits it with reshaping Brex when Henrique and Pedro realigned by preference.

Alex Hardimen

Product at the New York Times is inseparable from journalism in a way that has no equivalent at a tech platform. Alex Hardiman, CPO at the NYT, spent a decade there (2000–2016) steering it from print-first to mobile-first and launching the subscription paywall in 2011 — when consultants predicted it might reach one million paid subscribers at best. She then joined Facebook and was pulled into leading news product after the 2016 election, working to reclassify public content by credibility on a platform whose binary "friends or public" model had no concept of journalistic quality. That contrast became foundational to her thinking.

The NYT's strategy targets an estimated 135 million people worldwide willing to pay for quality journalism-based products across news, games, cooking, sports, audio, and shopping. News is the "sun" in a solar system; Wordle, The Athletic, Wirecutter, and Cooking are planets sharing the same DNA of trusted expertise. The target is 15 million subscribers by 2027, up from roughly 9 million at time of recording.

The structural difference from Big Tech: NYT PMs work across the full stack — content, distribution, and product — and sit alongside editors in cross-functional missions. Home-screen algorithms are trained on editorial importance scores that journalists produce, not just engagement signals, so algorithmic reach extends editorial judgment rather than replacing it.

Wordle's integration illustrated execution risk: when the Roe v. Wade draft leaked, the pre-loaded word "fetus" was queued for the next day. Because migration was mid-stream, it couldn't be changed for all users. Transparent disclosure about the development reality defused the controversy.

At Facebook the goals were scale, engagement, and revenue. At the Times, business goals serve the mission: growing subscribers funds an informed democracy, but impact also includes a reported story that triggers a law change. That broader aperture gives NYT product managers a dimension of purpose purely commercial work doesn't provide.

How Figma builds product

PMs at Figma own the "why" — why this problem above all others — while design and engineering own the "how." OKRs were scrapped for "headlines" (falsifiable period-end claims), then reinstated as "commitments" once data-science matured. Planning runs at three altitudes: annual priorities, half-year roadmaps, quarterly adjustments. Design crits meet five times weekly using silent FigJam commenting before open discussion; product reviews map an "option space" of all directions and use an alignment widget so the loudest voice doesn't dictate. Teams split between two products (Figma Design, FigJam) and horizontal platform groups. Customer proximity is core: PMs text users for quick feedback; CEO Dylan Field personally monitors community sentiment.

How to be the best coach to product people | Petra Wille (Strong Product People)

Great product lead coaching runs on five ingredients. First: a written "compass" — an explicit definition of what a good PM looks like, separating hard-to-coach personality traits (curiosity, empathy — hire for these) from teachable skills. Wille's PMwheel divides PM competency into eight buckets: understanding problems, finding solutions, planning, executing, listening and learning, team dynamics, personal growth, and agile foundations. Second: pin each PM on the map and identify their "next bigger challenge" — write this list quarterly; the opportunity will come. Third: share your vision for them, which often expands their own. Fourth: a development plan the PM owns, with specific small steps. Fifth: follow-up — a weekly nudge beats quarterly bursts. Start with step four even before the compass exists; PMs usually know what they want to fix.

Poor storytelling is a career ceiling. Advancing past IC level requires rallying teams through narrative; teams can only partially compensate for a PM who can't. A three-month direction story should take roughly two weeks of low-intensity work to craft. Use natural, sensory language rather than jargon — "product discovery" may be so overused it needs replacing with "we need to learn something." Prepare every story in three lengths: 75-second pitch, six-minute planning brief, 80-minute all-hands. Place the team or the user as the hero and use the hero's journey structure. Wille points to Nancy Duarte's frameworks, Hans Rosling's data-driven TED Talks, and Sarah Kay's spoken word poetry as models for how language moves people.

Community is the cheapest people-development lever available. Many PMs have never heard of external product communities. Internal communities of practice reduce coaching burden on managers, improve retention through mastery, and cost a fraction of conferences or coaches. Healthy ones distribute ownership across circles of interest rather than one coordinator, and measure signal-to-noise — not raw engagement.

How to hit revenue targets in a recession | Sahil Mansuri (Bravado)

In a downturn, the go-to-market playbook inverts: cold outreach dies, retention becomes the business, and innovation beats optimization. Bravado's real-time quota data from 300,000 B2B salespeople showed it starkly — Q3 2022 saw 63% of reps miss quota (up from 46% in Q1), 76% of companies miss targets, and October alone hit 85% monthly miss rates.

On forecasting: set a conservative base (plan for a 10% revenue decline), then pre-establish milestones that trigger acceleration or deceleration as actuals come in. Founders' optimism bias kills them in downturns; pre-commitment removes in-quarter rationalization.

On compensation: the standard 50/50 base-to-variable split with new-business-only quota rewards the wrong outcomes. A rep who closes 15 deals at 150% quota but churns 10 earns far more than one who closes 12 and retains all 12 with upsells and referrals. Almost no SaaS companies tie comp to renewal rates, yet retention is now the metric investors care about. Fix this with renewal kickers and per-rep churn tracking against team baseline.

On retention as the primary motion: when enterprise sales cycles double (62 to 115 days) and cold response rates collapse, existing customers are the only reliable revenue. Move your best account executives into customer success. Use your cross-customer data to give clients insights unavailable elsewhere, shifting from tool to strategic advisor.

On closing new deals: warm intros over text, not email, beat cold approaches. Keep the introducer on the thread to hold the prospect accountable; persist for months. In-person customer events reduce churn and generate referrals simultaneously.

On innovation: when the existing model breaks, don't cut price 15% — change the rules entirely. Bravado launched a 100% commission fractional model pairing laid-off reps who prefer flexibility with companies wanting growth without headcount risk, turning a crashing recruiting business into its best revenue month.

What differentiates the highest-performing product teams | John Cutler (Amplitude, The Beautiful Mess)

High-performing product teams are not all alike — dysfunctional teams share the same anti-patterns, but successful teams can look radically different. Cutler (product evangelist at Amplitude, ~800 one-on-one leader sessions and hundreds of workshops) calls this the reverse Anna Karenina principle.

Four markers surface repeatedly, each achievable through different tactics. First, coherence between organizational structure and strategy — brilliant people in a structure-strategy mismatch fail regardless of skill. Second, strong opinions loosely held: a stubborn core belief (Bezos's premise that today's success was set in motion three years ago) combined with genuine openness elsewhere. Third, leadership coherence — actions matching words rather than espousing "empowered teams" while behaving otherwise. Fourth, contextual skills: what counts as the right experience varies by company type, not just domain. Culture sits underneath all four; values stated without the behaviors that instantiate them tell you almost nothing.

Most product advice is built for Silicon Valley startups optimized for scaling a pure-digital product. Companies undergoing digital transformation face structural inertia no framework install will dissolve. The prescription: create protected pods where a team can complete the full build-learn loop, and treat frameworks as learning aids rather than endpoints.

For individual PMs: skill equals knowledge multiplied by practice, mediated by environment and habit. The industry produces enormous knowledge volume but few loop reps. In slow-moving companies, more loops are available than people assume — write the one-pager anyway, document assumptions, get one metric in motion. Waiting for the environment to improve means arriving two years later with nothing to show.

Three book recommendations: How to Measure Anything (Hubbard) — measure only enough to reduce uncertainty for the next decision; Accelerate (Forsgren, Kim, Humble) — empirical model of software delivery performance; User Story Mapping (Patton) — deceptively simple, durable vehicle for teaching product thinking.

How Coda builds product

Planning should consume no more than 10% of the execution period — a rule that shaped Coda's evolution from two-week sprints to six-week cycles to the current "Quarterly Plus" model: quarterly OKR commitments plus a non-committed look at the next quarter. Annual Big Bets, drawn from *Good Strategy Bad Strategy*-style diagnosis of key challenges, anchor team-level planning. OKRs target 100% delivery, not Google's 70% stretch; teams goal on input metrics they control, not lagging outputs like activation.

Product and design are one joint org — a deliberate correction to the Google pattern where PMs dominated decisions by holding more context. Shared org forces shared context: designers write specs, PMs sketch in Figma. Teams are outcome-structured and kept small: one PM, one designer, two to four engineers.

Review runs through two forums. Catalyst meets three times weekly with named roles — Driver, Makers, Braintrust, Interested — and sends automated Slack reminders including the norm "Add lift, not drag." Design Huddle is weekly, open to any stage, with designers specifying the feedback type they need.

Product briefs move through six stages: Problem, Press Release, Approach, Details, Journey, Retrospective. Writing competing press-release variants early forces alignment on customer value before any solution is committed.

Everything lives in a Coda team hub — specs, tasks, Figma embeds, decision logs — eliminating link rot and preserving decision history for new hires. Three closing philosophies: turn ambiguity into clarity, question the familiar from first principles, and prepare with the rigor of a professional athlete.

How to foster innovation and big thinking | Eeke de Milliano (Retool, Stripe)

Process is variance-reducing: it lifts underperformers toward the mean but pulls top performers down to it. Treat it as a minimum viable floor — label templates optional and let high performers skip them.

Eeke de Milliano joined Stripe in 2013 as its first account executive, became one of its first PMs, built Connect and Radar, then headed product at Retool. Stripe's output came from three things: first-principles questioning from Patrick Collison down (citing "best practice" invited "why?"), a writing culture where muddled prose meant muddled thinking, and accurate one-way/two-way door discrimination. Pricing looks like a trapdoor but isn't — grandfather existing users, change for future ones. Titles are a real trapdoor, so Stripe stayed flat as long as possible.

Teams aren't blocked from innovation by ambition — they're blocked by fear of failure, urgency debt, and no permission to think big. Normalize failure publicly: retrospectives not postmortems, failure write-ups circulated beside ship announcements. Create permission structurally: Retool ends every team charter with "Think Bigger — with 20% more time, what isn't on this list?" The CEO's annual Crazy Ideas doc — 90% probably wrong, 10x–100x if right — yields three to eight shipped ideas per year; Retool Workflows came from it.

Retool launched three products in one year (Workflows, Mobile, Database) by starting each with one or two people, isolating the teams from the core org, and treating Retool as the VC — no resources until signals appeared.

The 70/20/10 portfolio: 70% core product and tech debt, 20% strategic-but-not-core, 10% bets. Build for the user who immediately gets it, not the potential abuser. Build the scooter before the car: a complete small slice beats a partial foundation. Map team strengths every six months and hire against the gaps.

Leading with empathy | Keith Yandell (DoorDash, Uber)

Generalists outperform specialists because they're not anchored to how things have been done — Tony Xu's logic for putting Keith Yandell (litigator by training) in charge of HR, marketing, customer support, legal, and BD at DoorDash. The playbook: admit the knowledge gap, hire the domain expert, get out of their way. Kofi Amoo-Gottfried (now CMO) and Tia Sherringham (now GC) joined because that transparency signaled autonomy.

Yandell maintains a "How to Work with Keith" document covering expectations for high performers, his own failure modes (he argues against ideas even when he agrees — a litigator reflex), and a commitment to help reports find their next job outside DoorDash. Forwarding recruiter pitches to direct reports became a retention tool: people whose manager invests in their career tend to stay, and blind references come back stronger.

DoorDash's WeDash program (four mandatory delivery shifts a year; Yandell does it monthly) enforces customer empathy and doubles as a culture screen — engineers unwilling to deliver McDonald's in their Tesla signal a fit problem. The last ten minutes of every one-on-one go to constructive feedback directed at Yandell, borrowed from Uber's T3 B3 system (three positives about the manager, three tough ones, enforced by Kalanick).

For executive impasses: steel-man the opposing position, name a tiebreaker before arguing starts, set a deadline. People arguing the other side often persuade themselves.

The Series D closed with weeks of runway after universal rejections. Every firm that later led a round had previously passed. Projections the company consistently hit built the trust that unlocked capital.

During the pandemic, volume dropped at lockdown then doubled — while all support centers closed and the app crashed every Friday. The defining moment: Xu cutting $100M in restaurant commissions mid-IPO prep despite internal pushback. "Doing the right thing is never the wrong thing."

The ultimate guide to OKRs | Christina Wodtke (Stanford)

OKRs are a vitamin, not a medicine — they supercharge companies that already have strategy, empowered teams, and psychological safety, but expose dysfunction in companies that don't. The framework delivers three things: focus (one big rock per quarter), alignment (no ambiguity about what matters), and a learning cycle (quarterly grading forces retrospection that compounds knowledge over time).

The atomic unit isn't the document — it's asking every Monday, "What am I doing this week to move toward our goals?" Objectives sit one level below strategy, translating a "strongly held hypothesis about how to win" into a quarterly aspiration. Three key results per objective, derived by asking "how would we know we succeeded?" — brainstorm every possible measurement for 10 minutes to surface non-obvious signals; aim for one hard number, one quality indicator, one revenue signal. Tasks in the key-result slot are the most common error; outcomes, not deliverables.

Cadence matters as much as the document. Monday commitments (confidence level, last week, next week, three P1s) and Friday celebrations are the twin bookends. Weekly check-ins should take 10 minutes — scan for what's off, don't narrate every row. Replace top-down approval chains with three peers reviewing in 24 hours. Grade target: roughly 70%, uncomfortable but not doomed. Retrospective quality matters more than numerical precision.

Common failure mode: skimming *Measure What Matters* and blaming OKRs when the shallow rollout fails. Pilot with your best multidisciplinary team — they'll adapt the system to your culture and return a working template.

On PM development: business acumen is more foundational than "product sense." Product sense is compressed experience — it can't be acquired without years on the job. Young PMs should focus on business models, target markets, and revenue dynamics. The PM role is to serve the business, not just the user.

Competing with giants: An inside look at how The Browser Company builds product | Josh Miller (CEO)

Building a great browser isn't about optimizing metrics — it's about optimizing feelings. Josh Miller argues Silicon Valley's obsession with graphs leaves everything human on the table. At Arc, teams ask how they want users to feel rather than what metric to move. The Peek feature was designed around "airy and effortless," not a conversion rate. Feelings drive growth: delight makes people tell friends, eliminating the need for a growth team. Arc's north-star is D5 — people who open Arc at least five days a week — capturing retention, engagement, and growth in one ungameable number. Cohort retention sits in the low-to-mid 30s and keeps climbing.

The team runs on "heartfelt intensity" — hiring people with something to prove — plus "assume you don't know," a beginner's mind that biases toward action. Senior ICs including Chrome's original creator and Tumblr's founding designer joined for the work of their careers: ambitious prompts, genuine trust, and public credit. Every hire is announced with personal storytelling; shipped features credit the maker, not the CEO. Building in public — filming board meetings, sharing scheduling crises — is radical trust-building for a product holding your most sensitive data.

The strategic logic: browsers are a commodity everyone uses, controlled by companies with perverse incentives. Google doesn't want immersive web experiences competing with search; Apple doesn't want them bypassing the App Store. That gap is the opening. Arc's long-term vision is to be the iPhone of the internet — not just a browser but an "internet computer," the interface for a world where all computing lives in the cloud. Hardware becomes empty shells; the real computer is on the web. The developer platform built on top is where value accumulates, the same way the App Store mattered more than any first-party iPhone app.

How Duolingo builds product

Duolingo runs product teams under a co-lead model — a PM and an Engineering lead jointly own the team, with explicit domain splits: PM owns discovery and roadmap decisions, Engineering owns implementation. Teams are either metric-based (the Retention team owns CURR; the In-App Purchases team owns IAP revenue) or feature-based where no clean metric exists (the Connections team is measured via employee dogfooding, Reddit/Twitter sentiment, and long-term holdout experiments run 90+ days because social features compound slowly).

OKRs run quarterly in a three-week cycle: teams draft in week one, area leads consolidate in week two, senior leadership reviews in the first week of the new quarter. Yearly company OKRs are set each October; named "key result facilitators" own tracking through the year.

Product review meets Tuesday and Thursday in 20-minute slots, with the CEO always present. New PMs watch ten sessions before shipping anything. Review stages escalate from one-pager (concept) to optional 1.5-pager (wireframes) to pixel-perfect spec; a separate quality review checks polish and performance before rollout.

Resource allocation follows a portfolio model: mature teams run roughly 50/50 between big bets and incremental improvements; early-stage teams (Duolingo Math at writing time) run 90/10 toward new features until product-market fit is established. Bugs get quarterly "grease weeks." Every product change runs as an A/B test — over 200 run simultaneously at any time.

Driving alignment and urgency within teams, work-life balance, and the changing PM landscape | Nikita Miller (The Knot, Trello)

Teams that obsess over outcomes — OKRs, briefs, alignment rituals — while neglecting output miss the signal that any of it is working. If you're not shipping frequently, the strategy machinery doesn't matter. The PM's job is to drive urgency: ask "what did you deliver this sprint?" then "what's the cycle time?" rather than narrating a roadmap. Competition is the useful reminder — companies that believed they had none often found several serious rivals shipping faster a few years later.

The most concrete tool Nikita Miller uses is a written roles-and-responsibilities contract across what she calls the "chair" — product, design, engineering, and data. Each function writes what it expects of itself and of the others; the group ratifies the result. Scrum Masters disappeared without anyone reassigning their work — sprint execution has migrated to engineering managers, decision velocity to PMs. When velocity breaks down, the contract surfaces where the gap actually sits. On data: analysts embedded inside one product zone spot patterns far faster than a shared central team and eliminate the cross-team negotiation tax.

To surface velocity problems without triggering defensiveness, ask only questions: "What shipped to production?" "What was the cycle time?" The questions let people see the gap themselves.

On remote work, documentation and async communication are table stakes, but hard problems need in-person time with people you've built trust with. Her pattern: periodic two-day offsites with a tight pre-agreed agenda when a decision is stalled. The unlock at one such offsite was giving the data person uninterrupted floor time to educate the room before debate resumed.

The PM landscape has mainstreamed — 2 million LinkedIn profiles, degree programs, career worksheets for grade-schoolers. Roles are converging: PMs trending more technical, designers more business-oriented, engineers more product-focused — the precondition for genuine collaboration rather than handoffs.

On work-life balance: reject "balance," adopt "optimization." Decide what you're optimizing for this quarter, accept you can't maximize everything simultaneously, and use tech's flexibility to sprint hard then recover. Her universal question — *What are you optimizing for?* — makes trade-offs explicit before decisions get made.

Building a culture of excellence | David Singleton (CTO of Stripe)

Stripe's engineering culture is built on the conviction that every engineer must think like a product manager. The company went five-plus years before hiring its first PM because the founding team co-created products alongside early users — Figma and Slack were embedded co-development partners for Stripe Billing, shown product weekly until they were "super, super happy."

The operating principle "be meticulous in your craft" is operationalized through friction logging: a team member adopts a specific user's persona, walks end-to-end through the product, and keeps a stream-of-consciousness record of friction. Singleton does this monthly for the Stripe onboarding flow. Error messages in the Stripe API link directly to the relevant documentation page — more code handles edge cases than the main happy path. That compounding meticulousness produced a measured 10.5% revenue lift for merchants who migrated to Stripe's hosted payment surfaces, an enormous outcome in a space where improvements typically measure in basis points.

"Walk the Store" extends friction logging company-wide: in Friday Fireside meetings, the team walks through critical product flows together, creating shared language and surfacing divergences early. Engineering managers do "engineer occasions" — clearing three to four days to join a team, pick up a feature, ship it to production, and write a friction log of the development experience itself. Singleton learned Ruby this way.

Reliability at 99.999% uptime with 16.4 deploys per day comes from automated test suites, selective test execution, and progressive rollout to production in roughly 45 minutes. After incidents, Stripe prioritizes root-cause fixes ahead of roadmap items.

On management: Singleton reviews the prior week every Sunday evening and lists what would make the coming week a success — a practice he has run for a decade. References consistently provide the strongest hiring signal — thousands of hours of prior experience versus eight interview hours.

How Miro builds product

Miro's product motto — "deliver customer value faster with high quality" — is operationalized through interlocking systems. The 700-person AMPED org (Analytics, Product Marketing, Product, Engineering, Design) runs in seven streams, with Product Marketing embedded from day one so positioning is built alongside features, not after.

Planning starts with a three-year "painted picture," feeds annual strategy built at an offsite (most recently Barcelona, 1,000+ org-wide questions), and resolves into a rolling six-month roadmap with 80% confidence in the current quarter. That flexibility let Miro reach an AI beta in eight weeks by deprioritizing everything else.

OKRs were simplified after three failures: duplicated objectives at every level, quarterly cycles consuming too much process time, and too many KRs killing focus. Now semi-annual OKRs exist at three levels only — Company, AMPED, Stream — capped at three to four objectives and two to three KRs each.

Quality is maintained through the "Mona Lisa principle" (only ship what you'd put your name on), reinforced by monthly design reviews that publicly rate every shipped item. Resources default to 60% innovation, 20% maintenance, 20% tech debt.

Frameworks for product differentiation, team building, and thinking from first principles | Ayo Omojola (Carbon Health, Cash App)

Product differentiation requires three conditions simultaneously: different from what exists, better in a way that matters to the end user, and in a domain worth caring about economically. Any two is insufficient. Cash App's edge from 2014–2017 was instant money movement — when asked "why not Venmo?" the answer was "try sending me a dollar I can spend right now." Only one app could do that. Venmo and Apple Cash eventually caught up, but by then Cash App had compounded across roughly ten dimensions: talent density, fraud infrastructure, design, and consumer-first trade-offs that irritated internal Square stakeholders. The product stayed deliberately firewalled from the rest of the organization.

The startup-within-a-startup worked because the team stayed small (around 11 people for the first year) and senior. Smallness reduces miscommunication; seniority creates trust. Head count had to be fought for, which forced real scale before real hiring — without that constraint, team size outgrows actual product potential.

Going deep means the person in the execution role must command every detail, not the manager. Building the Cash Card required visiting card manufacturing facilities and testing over 1,000 combinations of plastic, overlay, and laser-engraving settings before shipping. In regulated domains, an expert's first answer is their inherited belief, not the actual boundary — keep pushing until you reach a concrete, falsifiable constraint.

On hiring: deliberately recruit founders. Ayo runs roughly half-founder teams. Upsides are high output and zero tolerance for waste; the costs are surfacing dysfunction immediately and two-year attrition cycles. The tactic: meet target hires early, give them what they want now (intros, advice), and stay close so the relationship exists when timing aligns.

In healthcare, network access often determines outcomes more than product merit. Carbon Health avoided this by going direct-to-consumer. The discipline in any regulated market: identify exactly who controls the decision rather than warming a chain of reluctant introducers.

Storytelling with Nancy Duarte: How to craft compelling presentations and tell a story that sticks

The presenter is never the hero — the audience is. Nancy Duarte, whose firm shaped over 250,000 presentations for clients including Apple and Al Gore, builds her methodology around this inversion: the presenter is the mentor (Obi-Wan, not Luke), and real power sits with the audience, who choose whether to accept or reject an idea.

The core structural tool is what-is / what-could-be contrast, oscillating between present reality and a desired future state, landing on "new bliss." Duarte mapped this pattern across Dr. King's "I Have a Dream," Steve Jobs's iPhone launch, and presidential speeches — finding the same cadence everywhere. The structure works at any scale: a keynote, a Zoom call, or convincing a spouse to handle a chore. Rise-and-fall tension pulls audiences to long for a future they hadn't considered.

Three post-it tips: (1) make the listener the hero; (2) infuse with story using the what-is/what-could-be rhythm; (3) ask "can they see what I'm saying?" — the napkin sketch, whiteboard diagram, and single-idea-per-slide all serve this. Every slide should support one organizing "big idea": your point of view plus what's at stake if the audience does or doesn't adopt it.

For high-stakes presentations, Duarte runs a listening tour first, rough-cuts the message, tests it with a layer of leaders on purpose-ugly slides — message work, not design — then finalizes. The Pixar analogy: build narrative, script, and visuals in sequence before the presenter stands up.

Stage fright: nervous presenters typically have the best content. Pre-talk, find a quiet corner, breathe, and cue funny videos so laughter chemically resets fight-or-flight before walking on.

For change movements across many presentations, the *Illuminate* framework maps five acts — dream, leap, fight, climb, arrive — prescribing speeches, stories, ceremonies, and symbols at each phase to sustain followers' emotional fuel.

Moving fast and navigating uncertainty | Jeremy Henrickson (Rippling, Coinbase)

Velocity at scale comes from small autonomous teams, platform investment, and leaders who stay close to ground-level detail rather than delegating upward.

At Coinbase during the 2017 crypto boom, 40x usage growth was managed by anchoring every fast decision to a security constraint and holding a six-month direction even while reacting to daily crises — debate hard, commit to one answer, pivot deliberately.

At Rippling, the new-product model: one entrepreneurial engineer and one designer embed in the platform for months, then recruit 2–4 engineers with a zero-to-one mindset and ship from blank page to launch in six to nine months. Time-and-attendance shipped with four people in nine months, touching nothing else. Senior leaders review designs every two weeks directly, skipping layers.

The MVP rejection follows: designing only for simple cases bakes in architectural assumptions that compound for a year. Global payroll launched across six countries simultaneously; 80% is shared platform, 20% country-specific configuration. Adding a country becomes configuration work, not engineering rewrites. The rule: design for the most complex use case first, even if you don't ship it yet.

Decision-making tempo is the cultural differentiator. Deadlines are hard — missing the planning window means being bypassed, not waited on. PMs own their domain deeply enough to answer questions in real time. "Go and see" means personally reading tax law for each country before hiring tax experts; understanding the problem first is what makes correct specs possible.

On hiring: case studies are deliberately too large to fully solve, so the signal is how fast candidates restructure when an assumption changes. The top interview question is "what questions do you have for me?" — asked before the product discussion. Depth of questions reveals business understanding, and humility to keep discovering separates early-career PMs who grow from those who don't.

How to ask the right questions, project confidence, and win over skeptics | Paige Costello (Asana, Intercom, Intuit)

Winning trust as the youngest person in the room comes down to one thing: bring the insight. Know the customer, market, competitors, numbers, and product deeply enough to add what others don't have. Anne Raimondi's trust equation formalizes it: trust = (credibility + reliability + authenticity) / perceived self-interest. Credibility is where insight tips the scales; reliability is your say-do ratio; authenticity means being vulnerable rather than armored. Watch real customers use the product — dogfooding cultures breed navel-gazing and erode sensitivity.

Confidence is projected through openness, not assertion: scan the room, close meetings by asking "did I get all of that?" The reframe "how might the opposite be true?" breaks scarcity thinking when a packed schedule feels impossible.

At Asana, planning shifted to rolling 12-month cycles refreshed every six months — high confidence in the near half, lower in the far half. An "area" layer sits between pillar and team, each owning a specific customer segment with metrics and guardrails against optimizing one number at the expense of real value. The Double Diamond maps reviews onto a diverge-converge cycle: kickoff → direction selection → concept review → spec → full experience review. Approval chains cap at three reviewers, one blocker; meetings above ten people get restructured.

What holds new PMs back: needing to be the expert drives solo advocacy — arriving with a finished spec and merely tolerating questions. The fix is genuine curiosity and categorizing feedback as "must do / should do / consider."

Recurring advice: "always answer the question they should have asked" (from an Intuit exec); "think big, ship small"; the Conscious Leadership Group's above-the-line (open, curious) versus below-the-line (committed to being right) distinction. Career assessment runs on three axes: learning curve steepness, environment quality, and genuine interest in the problem. Passions are made, not found.

How Snowflake builds product

Snowflake's product discipline centers on a unified-platform tenet: when a capability ships, Snowflake absorbs integration complexity so customers never have to. Performance and simplicity are non-negotiables — products are held back until they meet the bar, and SVP Kleinerman acknowledges "simplicity makes our lives harder" as a cost the team accepts.

Alignment flows from 6–10 leadership-set "big boulders" each year (e.g., application development, cost optimization). Each product area writes an annual six-pager mapping to those boulders; co-founders Dageville and Cruanes, plus product and engineering SVPs, read and comment on every document. Competing priorities escalate to a BTCG roundtable. Quarterly plans focus on customer scenarios rather than feature metrics to prevent area-local optimization.

About 85% of ideas originate bottom-up from PMs spotting customer pain. Data science is embedded in each product area; Kleinerman tracks dashboards at 24–48-hour granularity and Slacks PMs directly on changes.

Product reviews use six-pagers shared a week ahead — problem statement, goals/non-goals, risks, FAQ — aimed at alignment, not pitching leadership.

Culture rejects organizational politics: no "what's best for my team," only Snowflake and customers. Hiring screens for "drivers not passengers"; interviews open with candidates presenting a product they shipped.

How Shopify builds product

Shopify rejects annual planning — roadmaps collapse by March — in favor of CEO Tobi Lütke setting six yearly themes written from the merchant's voice. One 2023 theme, "Shopify keeps me on the cutting edge," became the AI investment lens: stay on the bleeding edge so merchants don't have to. Themes cascade into six-month roadmaps aligned to twice-yearly Editions releases, with four six-week sprint cycles per half.

OKRs are rejected on the same logic: metric-per-team ownership produces local maxima — numbers rise while the product loses coherence. The Admin visual redesign shipped with zero metrics, judged only by whether it felt right. Checkout is the exception, since conversion rate maps directly to merchant revenue.

Reviews run through GSD (Get Shit Done): five phases (Proposal, Prototype, Build, Release, Results), async PM video walkthroughs, and two sign-off tiers — OK1 at director level, OK2 at senior leadership.

In 2018 Shopify had 10 GM-led divisions. Tobi collapsed these to two — Core and Merchant Services — so the org chart wouldn't show through the product. The 11 Core teams are structured by jobs-to-be-done, each expected to serve merchants from first sale to Supreme-scale with no separate Enterprise team.

The AAA framework (Aiming, Assembling, Achieving) separates strategy from coordination from execution; the aimer need not be the most senior person. The company mantra: make the best product, make money to fund that, never reverse those two.

The Magic Loop

Rapid career advancement comes from a five-step loop: do your current job well, ask your manager how you can help them, do what they ask, then ask how you can help in a way that also advances a personal growth goal, and repeat. The mechanism is social — helping a manager first activates reciprocity, making them willing to route growth-relevant work back to you.

Ethan Evans, who spent 15 years at Amazon growing from a team of six to over 800 while helping build Prime Video, Twitch, and Prime Gaming, credits this loop for every promotion he received; mentees now hold VP and EVP roles at Google and elsewhere.

Step one is a prerequisite: managers discount offers to help from employees who aren't keeping up, so verify your standing by asking directly. Step three builds trust; taking on operational or low-prestige work others avoid signals you're helping for the manager's benefit, not your own advancement. That trust funds step four, where you convert goodwill into career-directed assignments.

The advanced form evolves from asking what's needed, to proposing specific ideas ("I noticed X — would fixing it help?", always leaving room for the manager to redirect), to simply acting and informing. The end state is a partnership, not a boss-employee structure.

When the loop stalls — a disengaged or exploitative manager — seek peer feedback first to rule out self-assessment bias, attempt the loop explicitly, then exit if nothing changes. Stagnant companies with few advancement slots limit returns; the loop accelerates movement within opportunity but doesn't create opportunity where none exists.

What sets great teams apart | Lane Shackleton (CPO of Coda)

The core job of a product person is to turn ambiguity into clarity. Lane Shackleton, CPO of Coda, built his career on principles designed to operationalize that idea.

First: systems, not goals. Jerry Seinfeld rebuilt his comedy set not by targeting an hour of material but by writing every morning and performing every night. Teams with a standing system for weekly customer conversations build genuine product instincts; teams with a quarterly OKR to talk to ten customers don't. Second: cathedrals, not bricks. People need different facets of the broader purpose — mocks, metrics, narrative — not a single writeup. Third: learn by making, not talking. At YouTube the team debated skippable ads endlessly; boss Phil Farhi ended it: "test the extremes, start tomorrow." One experiment used a tiny skip button, another covered the entire player. Directional data emerged in weeks.

Coda's rituals extend these principles. Catalyst replaces standing review meetings with three weekly one-hour blocks the whole company keeps free; topics run simultaneously for groups with non-overlapping attendees, eliminating the single-threaded bottleneck. Tag-ups move project work out of manager one-on-ones — where the eng and design leads are absent — into small-group sessions with the full triad. Two-way writeups add a "done reading" button, a Dory table of upvoted questions, and a sentiment pulse; the pulse surfaces dissent from people who would never speak aloud in a meeting.

On planning: keep strategy separate from OKR-setting, and cap planning time at 10% of the execution period. For feedback calibration: Dharmesh Shah's four tags — FYI, suggestion, recommendation, plea — give shared language for how much weight any input carries.

Career growth diagnostic: not "do you feel like you're growing?" but "how many oh-shit moments — of feeling genuinely underqualified — have you had in the last year?"

Mastering product strategy and growing as a PM | Maggie Crowley (Toast, Drift, Tripadvisor)

Three traits separate great PMs from competent ones: ruthlessly simplifying (finding the single most important thing and staying with it through completion), following up on results (setting reminders weeks and months post-launch to report back — rare enough that managers notice), and carrying the water (doing whatever unglamorous work — QA, sales calls, customer implementation — falls in the gap, because PMs are accountable for outcomes, not just the spec).

Strategy is about 5% of the job; shipping more beats strategizing more. A useful strategy doc starts from company mission, maps the landscape (SWOT, risks, competitor positioning), gives an honest accounting of product state and tech debt, then works up to opportunity, what has to be true for success, a proposed solution in three bullets, and a sequenced plan. Writing it forces clarity and lets collaborators identify exactly where their disagreement sits rather than rejecting the conclusion.

Being "data-driven" is a red flag when it crowds out qualitative reasoning. Ten user conversations produce better insight than any dashboard, which tells you what but never why. Complementary principle: if something is obviously better, use judgment and ship it.

Product content online is useful as a toolkit but harmful when treated as a checklist. Frameworks get sanitized for publication; the nuance that makes them work in context disappears. PMs who mistake following the framework for creating impact lose touch with the actual job.

On breaking in: get a PM title by any means (lateral move, startup, rotational program) because hiring screens on title first, shipped work second. Deep tenure — staying long enough to see consequences of your decisions across multiple cycles — compounds faster than job-hopping.

Publishing about your work accelerates development because it forces you to process what you've learned in a way doing the work alone doesn't.

Crafting a compelling product vision | Ebi Atawodi (YouTube, Netflix, Uber)

A compelling product vision is a vivid picture of a specific future state — not a purpose statement, not a strategy. Uber's mission was "reliable transportation everywhere for everyone"; the vision was a city where continuous trips eliminate parking, freeing the 25% of urban land devoted to it. Good visions are lofty yet attainable, free of today's technical constraints, and anchored to a sharp user problem. A vision requiring annual revision was never real; durable ones last three or more years.

Three formats make vision concrete. First, a Mad Libs narrative: "Once upon a time… and then one day… and because of that… the world looked like X." Second, a working-backwards article — write the TechCrunch headline and subtitle for when the thing ships, mocked on the actual masthead (Atawodi wrote "Uber is replacing your Clipper card" before the payments infrastructure existed). Third, post-it sketches handed to a designer; lacking a designer is no excuse to skip.

Developing vision follows three stages: empathize, create, evangelize. Empathize means dog-fooding and cat-fooding (using competitors), plus maintaining a living "top 10 things you should know" document — ranked user, quality, and infrastructure problems, updated quarterly by every PM. A three-day offsite structures the work: insights, strategy, big rocks. Big rocks are three to five ordered bets, not a twenty-item laundry list.

Evangelizing moves in three concentric circles: core team first (multiple passes, doc open for comments), then adjacent stakeholders who contributed problems and need to see them reflected, then leadership as high as possible. Because stakeholders shaped the input, buy-in arrives before the formal pitch.

The underlying craft is clarity plus conviction. Clarity strips everything obscuring the core problem. Conviction means committing to one lane — if you cannot defend the pick when resources are stripped away, the uncertainty is the work.

Radical Candor: From theory to practice with author Kim Scott

Withholding criticism is not kindness — it is the most common leadership failure. Kim Scott calls this ruinous empathy: caring about feelings enough to prevent someone from learning what they need to know. She estimates 90% of mistakes land there. Radical Candor means caring personally and challenging directly simultaneously. The other failure modes are obnoxious aggression (challenge without care) and manipulative insincerity (neither).

The cost of ruinous empathy compounds. Scott carried an underperformer named Bob for ten months with vague praise while his work forced teammates to redo deliverables. When she finally fired him, he said: "Why didn't you tell me? I thought you cared about me." The counter-model: a Google boss who told Scott "when you say um every third word it makes you sound stupid" — calibrated to that relationship because gentler versions had bounced off.

Practical mechanics: give feedback immediately and synchronously — phone beats video because facial signals add noise. Content follows CORE: Context, Observation, Result, next stEp. Don't ask "do you have any feedback?" — ask "what could I do or stop doing that makes it easier to work with me?" Wait six seconds; almost nobody holds silence. If you disagree, voice the 10% you can accept and return with a respectful explanation.

On the "Elon proves aggression works" objection: Scott separates Steve Jobs — Tim Cook offered him half his liver — from Ray Dalio, whose Principles gives four pages out of 400 to caring about people, and whose Bridgewater emailed recordings of in-meeting pile-ons company-wide as feedback models.

Scott's follow-on book Radical Respect was triggered by a Black woman CEO explaining that her gentlest criticism got labeled "angry Black woman" — surfacing the invisible cost her colleague paid to appear cheerful, and Scott's own denial about bias she had both experienced and inflicted.

The essence of product management | Christian Idiodi (SVPG)

Product management is disliked at most companies not because the discipline is flawed but because most people have encountered only incompetent practitioners. The role is competency-based: a PM earns influence by knowing the customer, data, and business better than anyone else — the "Bob" every CEO already trusts, replicated across every team.

Every product team faces four risks: value, usability, feasibility, and viability. Value is most neglected. Roadmap-driven organizations hand PMs a feature list with value assumed, making the PM redundant.

Idiodi's preferred discovery method is the reference customer: someone who loves the product enough to stake their reputation recommending it. Six to eight references signals B2B product-market fit; fifteen to twenty-five for B2C. He built a staffing-tech product this way. A Starbucks call about eight hundred potentially undocumented bakery workers was the seed; he and two colleagues drove to construction sites and malls to find others with the problem, then ran hiring by hand — Excel, flyers, phone calls, on-site shadowing — learning that show-up rates hover around fifty percent, so candidate volume must run three times target. Nine months of manual discovery produced a product that booked $32 million in its first ninety days. Reference customers also write the marketing: Idiodi records exactly what they say and puts those words on the box.

On coaching: leaders reproduce what was done to them because most have never experienced good coaching. The key distinction — doing product management is the PM's job; getting better at it is the manager's. To accelerate trust, Idiodi embeds new hires with the most influential person in the company for a week; credibility transfers and that person becomes invested in the hire's success.

Promotions fail when titles precede preparation. Performing at the next level before the title changes means mistakes don't occur under full leverage.

The art and wisdom of changing teams | Heidi Helfand (author of Dynamic Reteaming)

Teams change constantly, and advice to keep them stable ignores that growth makes this impossible. Tuckman's forming-storming-norming-performing model omits a fifth stage: stagnating. The question isn't whether teams change — it's how to handle it.

Five patterns cover all team change. One-by-one: someone joins or leaves — help newcomers belong immediately and coach those who had no say. Grow and split: a team outgrows itself when meetings drag and work diverges; splitting trades those problems for new dependencies and resource gaps. Merging: the reverse, common in downsizing or acquisition; a "story of our team" exercise — each side builds a shared milestone timeline — creates common history before facing what's next. Isolation: spin a small team to the side, give it process freedom, report it to a senior decision-maker, shield it from interruption. Expertcity's isolated team built GoToMyPC, the pivot that saved the company; AppFolio's SecureDocs became an independent acquisition; the Chicken McNugget was rescued by a SWAT team reporting straight to a McDonald's executive. Switching: moving between teams for learning and fulfillment also builds knowledge redundancy — no single system owner — which matters when critical transactions run through your stack.

Three anti-patterns: percentage allocation (10% here, 20% there) fragments attention; silent arrivals and departures erode trust; spreading high performers to diffuse their energy destroys the source team without lifting others — AppFolio CTO Jon Walker confirmed it.

For transparent reteaming, Spotify and Procore posted draft structures on whiteboards — names, missions, open slots — so people could spot errors and express interest. The RIDE framework (Requester, Input-giver, Decider, Executor) keeps authority clear. Time-box participation; more people for longer displaces actual work.

There is no perfect org structure — only the best current approximation and workarounds for its failure modes.

Good Strategy, Bad Strategy | Richard Rumelt

Most organizations produce bad strategy by mistaking goals and ambitions for the real thing. Rumelt's definition: strategy is a design for overcoming a challenge, built from three elements — diagnosis, guiding policy, and coherent actions. Diagnosis is a causal account of what is hard, not a statement of desire. Coherent actions implement the policy without contradiction — a company targeting growth and return on equity with no mechanism has incoherent actions.

Bad strategy is recognizable: 17 priorities (violating the meaning of the word); performance goals dressed as strategy; diagnostic gaps — jumping from "U.S. 15-year-olds score low in math" to "send more people to college" with nothing between. The intelligence community's strategy for 17 agencies amounted to "they should coordinate better" — naming the symptom as the cure.

Power — exploitable asymmetry — is what makes a strategy more than a coin flip. Gerstner's IBM bet on trusted relationships with every large corporation on earth; network effects give Amazon and Google advantages compounding with scale and data. Those IBM relationships later became a liability when large enterprises were last to adopt cloud.

The Crux sharpens this: find the hardest part of the problem, the move you cannot skip. I.M. Pei resolved the Louvre entrance problem by making the new building transparent. SpaceX asked why a rocket can't land on its own thrust. Insight does not arrive on demand — it emerges from immersion in what makes the problem hard. History and biography train that muscle better than frameworks.

Practical prescription: call it an action agenda, not a strategy. Start from ambitions, identify which are addressable now, name what blocks progress, then specify what you will do. For startups the equivalent is a search posture: hold conviction and willingness to pivot simultaneously, because the initial hypothesis rarely survives contact with customers.

Lessons from Atlassian: Launching new products, getting buy-in, and staying ahead of the competition | Megan Cook (head of product, Jira)

Psychological safety erodes when teams split into distinct streams — visible by around 15 people, when feedback becomes "painfully polished" and ideas turn incremental. Megan Cook, head of product for Jira, traces this to a coach's framing: the opposite of play isn't work, it's fear. Her fixes: biweekly PM peer groups reviewing rough-draft work; biannual onsites where senior leaders open by narrating their own failures; a "$10 game" where managers and reports allocate attention across priorities to surface misalignment.

Atlassian found intentional in-person gatherings boost connection and productivity ~30%, lasting months — three or four times a year is enough. Remote discipline: block two deep-work stretches weekly across the leadership team, ban status-update meetings in favor of async tools, keep one-on-ones short with flex time for unexpected problems.

Getting buy-in is a journey, not a presentation. Partner with affected stakeholders early so they're advocates before the final meeting. Distinguish hypotheses from confirmed facts — naming what you don't know reads as more credible than false certainty. Open by stating what you need: a decision, feedback, or help stress-testing a hypothesis.

A two-year CSAT program on Jira usability: customer video sessions made friction visceral; a "shepherd" model let platform teams contribute one reviewer rather than bear full dev cost; starting with ~40 people built momentum before scaling.

Atlassian incubates new products through five gates: Wonder (one person), Explore (prototype + small team), Make (~12-person build), Impact (revenue self-sufficiency), Scale — kill-or-continue at each. Cook's failure: she scoped developer automation narrowly as a product feature; Atlassian later acquired a company to do what she could have built platform-wide. The standing question: where could this be in three to five years?

Fight Club — 30 minutes weekly with Cook, her engineering lead, and design lead, explicitly for conflict — keeps tensions from compounding.

My favorite decision-making frameworks

Consensus produces no ownership; the person who executes a decision should be the one who makes it. Gokul Rajaram's S.P.A.D.E. (Square) and Brian Armstrong's Coinbase sheet both enforce this — define the Setting, lay out alternatives, assign a single Decider accountable for execution. Shishir Mehrotra's Dory/Pulse equalizes voices in real-time meetings via anonymous voting and ranked-choice. Bain's RAPID clarifies roles without full process overhead. The Eisenhower urgent/important 2×2 handles prioritization. Reserve any framework for one-way-door decisions only — high-trust teams should run most calls on judgment alone.

How Netflix builds a culture of excellence | Elizabeth Stone (CTO)

High talent density is not a goal at Netflix — it is the prerequisite for everything else. Without it, the other pillars (radical candor, freedom and responsibility, minimal process) would be "dangerous," as Elizabeth Stone puts it. Reed Hastings built the company on the belief that the environment people thrive most in is defined by the quality of their colleagues, not perks or structure.

The operational expression of this is the keeper test: managers ask themselves, "If this person announced they were leaving today, would I fight to keep them?" If not, that conversation needs to happen now. Netflix has no performance review cycle — only an annual 360 for developmental feedback and a compensation review. Performance is supposed to surface in real-time, though Stone admits the annual 360 becomes a crutch when candor falters.

Stone's career accelerated (VP to CTO at four companies in two-to-three-year windows) through responsiveness as respect — treating reply speed and follow-through as non-negotiable even at senior levels — and the ability to translate between technical and non-technical stakeholders. Her economics PhD shaped how she thinks about incentives and unintended consequences; she sees economics as a flavor of data science that makes messy problems tractable.

On data structure: Netflix keeps data science, engineering, analytics, and consumer research in one centralized org rather than embedding them in business lines. The advantage is objectivity — the team's job is to be truth-tellers, not to validate decisions already made. Combining quant behavioral data with qualitative consumer insights in the same org produces what Stone calls a superpower in personalization and recommendations.

IC leveling was introduced two years ago — before that, all engineers were simply "senior engineer." Stone held a public postmortem on what didn't work, as a concrete example of the candor the culture demands of leaders.

Making Meta | Andrew 'Boz' Bosworth (CTO)

At Meta, the most valuable thing a leader does is keep information flowing. Boz's most-given career advice: lean on your manager more aggressively. Your job is not to prove you can do it alone — it is to get it done. A "no response required" status note with blockers flagged and a pre-drafted email for the manager to forward removes nearly all friction. Meta formalized this as HPMs (Highlight, People, Me) — weekly written updates every manager sent up the chain.

News Feed (2006): user outrage and doubled engagement happened simultaneously, making the conviction call easier. The lesson is not "ignore users" but to distinguish stated from revealed preferences and ask whether the details were wrong versus the direction. News Feed's ranking was the first consumer AI of its kind; the ads surface built on top became the most efficient monetizing surface in history outside Search.

Zuckerberg's management is "eye of Sauron" — whichever problem he picks gets pixel-level scrutiny; everything else gets delegation. His response to feedback: pressure-test it across subsequent meetings without attribution, triangulating before shifting.

Meta's 2022 downturn had two sources: COVID over-hiring and failure to explain the portfolio logic (AI plus Reality Labs) to markets that had never seen a downturn. The correction required org flattening, headcount cuts, and relentless communication of the rationale.

On hardware: Quest 3 beats Apple Vision Pro on hand tracking, field of view, display brightness, and motion blur in pass-through. AVP wins on stationary high-res video.

The deepest personal failure Boz names is a 2006 RPC encoding argument where identity threat triggered a meltdown in front of half the engineering team; Federman was right. The lasting lesson came from watching colleague Ami Vora meet strong disagreement with genuine curiosity — "Fascinating, tell me more" — and watch it dissolve walls instantly.

How to build deeper, more robust relationships | Carole Robin (Stanford GSB professor, “Touchy Feely”)

Interpersonal competence — not vision or strategy — determines whether people follow you long-term. Carole Robin taught this at Stanford GSB for 20 years in "Touchy Feely" (Interpersonal Dynamics), then founded Leaders in Tech to bring the curriculum to tech founders.

Relationships exist on a continuum from dysfunction to "exceptional," defined by six characteristics: each person is better known by the other, disclosures are trusted not to be weaponized, honesty is safe, conflict resolves productively, and both are committed to each other's growth.

The 15% rule: vulnerability is reciprocal — guarded signals invite guarded signals back. Step just outside your comfort zone, see whether the other person reciprocates, then settle into a new baseline before going 15% further. Feelings vocabulary matters: most people default to attributions. "You're being insensitive" is not a feeling; it crosses onto the other person's side of the net. Anger is almost always secondary — masking fear or hurt. Anger distances; fear and hurt connect. A CEO who replaced a blast after a missed deadline with "I'm scared I'm the only one worried about this" got faster action.

The three-realities model: every exchange has your intent/context, the observable behavior, and the impact on the other person. You only have access to two. Staying on your side of the net means speaking from behavior and feeling only: "when you did X, I felt Y, because Z." Artful inquiry uses what/when/where/how — never why, which invites defensiveness.

Address pinches (small irritants) before they compound into crunches. When repair is needed, ask "what did you hear me say?" — nine times in ten it differs from what was said. Inquiry before advice; unsolicited solutions widen power differentials and usually miss the problem. Failures are AFOGs (another fucking opportunity for growth) — the only useful question is what you learned.

Inside Canva: Coaches not managers, giving away your Legos, and running profitably | Cameron Adams (co-founder and CPO)

Canva reached $2.3B ARR, profitable for seven years, growing 60% year-over-year. That profitability traces to near-disaster: at the $100M round, a lead investor cut the valuation 50% two days before closing. Co-founders flew to Silicon Valley, found a replacement in a week, and made it permanent policy: never depend on outside capital for survival.

Canva has no managers — only coaches. Every employee is matched with a same-specialty coach; roughly 800 employees coach others, tracking skill development and when to level up. Reviews are 360-degree, every six months. The system grew from the founders' coaching experience and is introduced at onboarding alongside "giving away your Lego" — Molly Graham's concept that scaling means handing off identity-forming work to operate at higher leverage.

Canva PMs are connectors of ideas, teams, data, and constraints, not gatekeepers. The company is visually wired: strategy means mockups. Outside executives with preconceived processes tend to fail; those who observe first succeed. Canva avoided the title until year six.

The MVP took a year, treating delight as a prerequisite for word-of-mouth. Social media managers emerged as ICP through emotional intensity in testing. Onboarding was the unlock: chaining tiny steps until users said "I didn't know I could be a designer."

Growth hire Andre mapped jobs-to-be-done to search queries end-to-end — Google to landing page to template to magic moment. Internationalization began year three: 8 languages first year, 100+ by 2017. Brazil, India, and Indonesia are top-five markets.

Freemium was mission before strategy. Revenue evolved from per-element $1 sales to Canva Pro, then absorbed element sales into the subscription — two hockey sticks. AI runs on three pillars: proprietary models where Canva has data advantage, external partners (OpenAI, RunwayML), and a 170M-user app ecosystem for third-party distribution. Enterprise is next; 95% of Fortune 500 already use Canva organically.

General management, functional, and hybrid models: Which org design works best for top companies?

The choice between GM and functional org design hinges on two factors: how directly your North Star metric ties to revenue, and how critical a cohesive customer journey is to growth. Revenue-proximate metrics (Block, Coinbase, Amazon) suit GM models with P&L accountability per business unit. Ad-driven platforms (Meta, Google) suit functional models because engagement-focused North Stars sit steps away from revenue. Shopify's 2020 shift to functional unlocked organic cross-sell flywheels; Block reversed course in 2024 after Dorsey concluded its GM structure slowed engineering quality. Most companies land hybrid — GM-like "champions" inside functional orgs (Meta, Shopify) or centralized platform teams alongside GM units (Coinbase, Intuit). Talent reinforces structure: GM models create specialist career bottlenecks; functional models concentrate end-to-end ownership at the top.

Building a world-class data org | Jessica Lachs (VP of Analytics and Data Science at DoorDash)

Analytics at DoorDash is a business-impact function, not a service function — the job is answering not just "why" but "what do we do now." Jessica Lachs, who self-taught SQL and Python as DoorDash's first GM and built its data org across 10 years, argues centralized reporting beats embedding analysts in business units. A center-of-excellence model maintains a consistent hiring bar, enables career mobility across teams, and prevents six teams rebuilding the same churn model. The key mechanism: analytics pods carry the same goals as their partner teams, aligning incentives without reporting through them.

Lachs's metrics framework has three rules. First, goal on short-term proxies that predict long-term outcomes — retention can't be moved quickly; find the input levers instead. Second, keep metrics simple enough that people have intuition for them: DoorDash's weighted "merchant health score" was uninterpretable; three concrete inputs (photo coverage, accurate hours, first order within 7 days) worked better. Third, translate everything into a common currency — Gross Order Value — so marketing spend, logistics, and restaurant signing can be traded off across the whole business.

Averages hide fail states. "Never Delivered" orders are rare enough to vanish in average quality metrics but cause immediate churn and outsized cost: refund, repurchase, second dasher. DoorDash makes eradicating them a dedicated team goal. A related blind spot: login failures don't appear in order data because those users never reach the denominator.

A hackathon deep-dive on referrals found a bimodal distribution: genuine peer referrals had excellent payback, but a group posting codes online for discount-hunters dragged the average down. The fix — fraud caps, stricter rules — was found only because the team committed test fraud and ordered cupcakes.

"Ask Data AI" lets non-technical employees adapt SQL queries themselves, converting Office Hours support load into self-serve capability.

5 essential questions to craft a winning strategy | Roger Martin (author, advisor, speaker)

Every strategy must compel customers to hand over money they could spend elsewhere — the only real test. Roger Martin, former Rotman dean and P&G advisor, argues most companies fail because business schools replaced useful strategy with the "resource-based view of the firm," which tells firms to build resources without saying which ones or why. His alternative, the strategy choice cascade, links five questions: winning aspiration, where to play, how to win, what capabilities to build, and what management systems sustain them.

"How to win" is binary: differentiate so customers specifically demand you, or be the cost leader. No middle position is defensible. Vanguard reached $9 trillion holding the lowest-cost index fund position. Southwest entered Boston–Chicago at $200 against a $1,000 duopoly; competitors ceded share because replicating its single-aircraft fleet, point-to-point routes, and flexible labor model meant dismantling their business. Lego captured 80–90% of toy category growth most years because children define a store without it as "not a toy store."

Capabilities and management systems build and maintain the moat. Westlaw employs 1,500 lawyers to annotate every U.S. court case with a proprietary keyword system built over a century — replication requires 150,000 lawyers and 50 years seeding law schools, so no one tries. Four Seasons holds 10% staff turnover against an 80% industry norm through distinct recruiting and career development, sustaining the top luxury position for 35-plus years. The more multifaceted the capability stack, the more likely competitors pick a different where-to-play rather than attack directly.

For practitioners overwhelmed by scope: find the most painful gap between current and desired outcomes, use the cascade to change one choice, close it, then move to the next gap. Strategy is not innate — A.G. Lafley accumulated reps in the Navy at 25. Every great strategist Martin has met simply practiced more.

The ultimate guide to performance marketing | Timothy Davis (Shopify)

Paid search is table stakes for every company — Google has pushed organic results so far down the page that paid presence is mandatory. Beyond search, find channels with traction in analytics and amplify.

To enter a new platform, run a "signs of life" test: load your customer list, build Meta lookalikes at 1% similarity, put small budget behind it, and read click-through rate to diagnose targeting before blaming creative. Creative rarely transfers across platforms; each has distinct user behavior.

Platform sequencing: Google Search first (intent-driven), then Meta, then YouTube. LinkedIn costs roughly 3x other channels but enables job-title- and company-level targeting — Davis ran ads geo-fenced to Coca-Cola's LA office on two specific objections before a sales call, and the prospect acknowledged seeing them everywhere. Use LinkedIn only after Google and Meta are working, mainly for high-LTV deals.

Agencies apply cookie-cutter playbooks; in-house teams go deeper — weekly keyword reviews, negatives, ad copy tests, landing page iteration. Core discipline is signal versus noise: for a conversion campaign, clicks and conversion rate are the signal; reach and frequency are noise. An ops-cadence spreadsheet with explicit weekly/biweekly/monthly frequencies enforces this.

For attribution, time-decay multi-touch is preferred, but all models are biased toward claimed credit. Incrementality tests (Geo-Lift or Conversion Lift via platform partners) reveal true causal impact — eBay found brand paid search redundant given organic strength. Worth running only above $50K/month.

Hiring sequence: first, a data-oriented generalist — screen by throwing a data dump at candidates and watching whether they ask about campaign goal before optimizing. Second, in-house creative (outperforms agencies on tone and speed). Third, a data scientist for incrementality modeling. Two red quarters on the capacity calculator triggers the next hire.

AI is already embedded in smart bidding and copy recommendations; dynamic creative builders are the next edge.

Improve strategy, influence, and decision-making by understanding your brain | Evan LaPointe (founder of CORE Sciences)

Most workplace dysfunction is the gap between what neuroscience knows about human brains and what businesses do.

The brain has three motivational systems: safety (restore standing when threatened), reward (transactional gains), and purpose (activated when you understand your impact on real people). Most companies run on safety and reward alone. Activating purpose means answering "why does this matter to real people" through logical deduction from the role you play, not inspirational mission statements.

Personality differences shape strategy capacity. Openness is the key Big Five dimension: low-openness brains have abstract thinking wired to pain, so ambitious ideas feel physically aversive. High conscientiousness adds pressure toward execution and away from reframing. The fix is self-disclosure — naming your wiring lets teammates route around constraints rather than fight them.

Meetings fail by skipping priming. Teams launch into decisions without aligning on context or — critically — principles. If one person optimizes for speed and another for accuracy, the argument looks tactical but the disagreement is at the principle level. Priming can be three minutes in the calendar invite.

Influence has three speeds. Slow: let failure teach. Moderate: give someone a new frame and let them live with it (Challenger Sale). Fast: surface the underlying belief driving a behavior and challenge its logic; requires a habitat of sufficient trust.

Relationships have three components: ability, trust (three levels — simple delegation, "as good as I'd do it," "beyond my judgment"), and appeal. Appeal is the most powerful gate biologically: high ability and trust get neutralized when someone activates others' safety systems. Fix appeal first.

Brains operate in alpha (daydreaming), beta (execution), and gamma (deep reframing). Most teams run almost entirely in beta. Target 25% in alpha or gamma, structured as a quarterly offsite and a weekly protected block — giving calendar-invading distractions a natural home.

The things engineers are desperate for PMs to understand | Camille Fournier (author of "The Manager's Path," ex-CTO at Rent the Runway)

Engineers' frustrations with PMs cluster around four behaviors: hoarding credit (being the sole public face of work the engineering team built), dismissing technical details as unimportant, acting as telephone intermediaries who lose fidelity in translation, and monopolizing ideation — which drives engineers to over-engineer technology choices as a substitute creative outlet. PMs who invite engineers into product thinking get better decisions and less gratuitous churn.

On rewrites, the failure mode is consistent: engineers convince themselves a side-project replacement will free them from a painful legacy system, but underestimate how long migration takes and how much undocumented logic is buried in the old system. The better path is a staged uplift — isolate a well-bounded component, improve it in place, and avoid the fiction that feature development can pause for a year while rebuilding from scratch.

On the management transition, Fournier argues for waiting until technical mastery is genuinely internalized — roughly ten years of focused coding. Once hands-off, credibility comes from asking good questions and guiding decisions, not prescribing libraries. Senior engineers distrust leaders who issue specific technology directives from a distance; they respond to leaders who listen and have real feel for the problem.

On one-on-ones: peer and stakeholder one-on-ones scale badly. A dissatisfied stakeholder complaining privately never hears that five others are satisfied, so the complaint stays uncontextualized. Meeting load deserves the same audit as work priorities.

Platform engineering requires software engineers (not just SREs) and dedicated PMs, because platforms are products. A team without product ownership builds what seems technically interesting rather than what reduces cycle time or unblocks launches. Healthy platform teams surface solutions application teams already invented, assimilate them, and scale them out. Most of the work is migrations and stakeholder management — leaders who find that unglamorous will be unhappy in the role.

Thinking like a gardener not a builder, organizing teams like slime mold, the adjacent possible, and other unconventional product advice | Alex Komoroske (Stripe, Google)

Most builder effort caps at the value of the effort itself. The gardener mindset looks for dynamics that compound on their own — network effects, ecosystem loops — and invests only when traction appears. Plant cheap seeds, water the ones growing. One oak in ten beats the builder.

Organizational kayfabe names why large companies become dysfunctional without bad actors: a single asymmetry — you can't make your manager look wrong — propagates upward until leadership is orders of magnitude from ground truth. Give 70% of team effort to legible value creation, earning the 30% slack in which emergent work survives. Credibility keeps the acorn from being dug up.

LLMs are "magical duct tape," cheap distilled human intuition between humans and computing in cost structure. They upend the assumption that software is expensive to write and cheap to run. Taste — being distinctively different from what an LLM would produce on the same prompt — becomes the scarce resource. Compounding returns give you zero or 1,000, not 93 vs. 95.

The adjacent possible: the right move is always within arm's reach. Set a low-resolution North Star three to five years out, find the adjacent step with the steepest gradient toward it, take it, repeat. False precision is a comfort blanket, not information.

Strategy salons (nerd clubs): opt-in, "yes, and" communities seeded with four to six people who already want the conversation. Add diverse newcomers slowly. Redirect with "I wonder" framing. Post FOMO summaries after live sessions. Steer toward a predetermined output and it stops.

The Hallmark Card fallacy: rediscovering something said a million times means you're finally ready to hear it. Always rules beat sometimes rules for self-control. Happiness is reality minus expectations. Do things that give you energy and that you'd be proud of in front of people you respect.

Why no productivity hack will solve your overwhelm

Productivity overwhelm is not a time-management problem — it is an inner conflict problem, and no external tactic can fix it. Gustavo, a Director of Product, knew every productivity framework but kept dropping balls because one part of him wanted better time management, a second compulsively said yes to avoid disappointing anyone, and a third — exhausted by the other two — nudging him to quit. Coaching him to prioritize one big task per day did nothing; the root was internal conflict.

Internal Family Systems (IFS) names this dynamic with three principles: we are made of many parts, not a single fixed self; every part — including the inner critic and the procrastinator — carries positive intentions; and we each have a calmer Self capable of facilitating dialogue among parts.

The practical process runs five steps. Sense and name the competing parts — they surface as body sensations, images, or emotions. Validate each rather than shaming the "unproductive" ones, because shamed parts get louder. "Unblend" by interviewing each part one at a time: what is it protecting you from, and what does it fear would happen if it stopped? Facilitate dialogue with no winner — parts often share underlying goals. Rebecca, paralyzed by a COO offer as a new mom, found her ambitious, cautious, balance-seeking, and loyal parts all wanted her to feel fulfilled; once they talked, she negotiated a phased transition instead of staying stuck. Finally, synthesize the parts' input into a decision and communicate it back to the internal team.

Two payoffs follow: less internal thrash, and higher tolerance for external conflict — which transfers directly into leadership effectiveness.

Becoming more strategic, navigating difficult colleagues, harnessing founder mode, and more | Anneka Gupta (Chief Product Officer at Rubrik)

Being strategic means two things simultaneously: articulating a compelling "why" behind decisions, and championing hard changes that serve the long-term. Either alone fails — big ideas without clear rationale look like noise; clear rationale behind small ideas looks like incrementalism.

Anneka Gupta (CPO at Rubrik, Stanford GSB lecturer) built this after getting "not strategic enough" feedback twice. Two tactical fixes: synthesize what's been said before adding your view — "Here's what I've heard, are we agreed?" reads as strategic and surfaces disagreement early. Then apply "one click better": sharpen existing inputs through an outside-in customer lens rather than generating ideas from scratch.

On founder mode: understand the founder's actual objective before pushing back on their mechanism — the wrong solution to a real problem. Treat the founder as a lever for initiatives you think matter. As a product leader, go deep into business details to know where intervention is warranted, then get in early by asking teams to present strategy rather than rewriting their work.

Decision-making: commit at 70% confidence. Decisions generate high-fidelity information that hypotheticals never do; reward learning over outcomes to build a culture willing to take bets. Common failure: building the right product with no one ready to sell it — know the sales motion before building.

On difficult colleagues: map what they care about, connect it to what you need, and extract something learnable even when they're obstructive — it converts frustration into a generative frame.

Hard feedback: feel the reaction, wait, come back curious. When giving it, state care before content, frame as perception not verdict ("this is how you're being perceived"), and involve them in solutions. Ask about career goals first to calibrate what matters.

New PMs misunderstand the role as tools to learn. The core skill is driving clarity from ambiguity, consistently.

On being funny at work

Humor is a learnable skill, not a natural gift — and in business environments largely devoid of it, even modest ability to make people laugh confers outsized advantage. Laughter replaces cortisol with dopamine, oxytocin, and endorphins, priming audiences to learn, feel connection, and trust the speaker. Forbes ranks sense of humor fourth among leadership qualities.

The underlying mechanism is always surprise: assembling words to produce an outcome the listener didn't expect. Seven teachable strategies:

Nostalgia — contrast past and present to startle people into noticing how fast the world changed. Even when it doesn't land as funny, it reads as informative rather than a failed joke.

Exaggeration — only works when the scale is so outrageous nobody could take it literally. Steve Jobs's iPhone launch slide showed a janky iPod with a kitchen timer and got a big laugh.

One of these things is not like the other — list two expected items, then end with something emotionally accurate but categorically alien: "simplify workflows, shorten the sales cycle, and bore your family to death."

Definitions — distill a subject to its single sharpest attribute. Allstate's Mayhem character does this; so does calling a chair "an antigravitational device designed to keep your ass suspended in midair."

Specificity — "Jack Russell terrier named Ginger" is funnier than "dog." When it fails to produce a laugh, it reads as vivid detail, not a misfire.

Stating the obvious — name what everyone sees but nobody says. The refrigerator is the only household object where a plumber, a florist, and a hardware store all advertise simultaneously.

The unexpected — put the funniest word last. Revealing the punchline upfront kills it; let the audience arrive there themselves.

Scripts for difficult conversations: Giving hard feedback, navigating defensiveness, the three questions you should end every meeting with, more | Alisa Cohn (executive coach)

Avoiding difficult conversations doesn't protect people — it denies them feedback that could change their career. Executive coach Alisa Cohn's core argument: a leader's job is not to keep people happy but to drive results. Founders who prioritize morale over accountability end up with cliquey cultures and no traction. Work backwards from winning — a team that executes and hits milestones feels better than one built on avoidance.

On feedback, Cohn distinguishes observable-fact delivery from venting. The script structure: state the specific pattern ("what I've observed" or "what I'm hearing from peers"), anchor to shared expectations ("we both know"), close with a concrete ask — leave this meeting knowing what changes. When someone gets defensive, pause explicitly: "I can see the temperature has changed" — then restate that the conversation has to happen.

On no-promotion conversations, state the decision early, give specific reasoning, and always offer hope — what the person can build toward. On firing, the termination should never be a surprise: a 30-day "last chance" conversation with explicit consequence language makes the final conversation brief and defensible.

Three questions to close every meeting: what did we decide, who does what by when, and who else needs to know. Ask the first question around the room — six people in the same meeting give six different answers.

The founder prenup covers five areas: values (narrow to 3–5 core values and compare); vision of success (IPO vs. lifestyle business — misalignment surfaces painfully five years in); how each person handles conflict (ask your spouse, not just yourself); how you'll decide when you genuinely disagree; and what company culture looks like to each of you. A Personal Operating Manual — communication style, delegation preferences, pet peeves, what earns a gold star — serves the same function for any team.

An operator’s guide to product strategy | Chandra Janakiraman (CPO at VRChat, ex-Meta, Headspace, Zynga)

Product strategy is not a talent gene — it is a repeatable process. Strategy sits between mission/vision and roadmap: it forces choice to deploy scarce resources for maximum impact, comprising three strategic pillars, explicit non-focus areas, and the why behind both. The resonance metaphor: select the frequency that makes product and market vibrate at disproportionate amplitude.

The "small-S" playbook runs 8–12 weeks across five phases. Preparation (four weeks): a cross-functional working group (product, engineering, design, data) builds a master readout — behavioral data meta-analysis, UXR synthesis, leadership interviews (ask what leaders want before pitching — the "fruit story"), competitive stack charts, live user observation. The strategy sprint (one week): day one aligns the group's diagnosis; day two generates and clusters 10–15 problem areas, flips each into opportunity framing ("difficulty finding things" → "discovery"), and ranks all clusters on four criteria — expected impact, certainty of impact, clarity of levers, uniqueness of those levers to this team. Top three become pillars; the rest are explicitly out. Day three produces a "winning aspiration" via newspaper headlines (Meta example: moving consumer trust on Facebook). A design sprint follows with concept-car artifacts to make each pillar tangible. The PM writes a 3–4 page document; rollout is gatekeeper one-on-ones, then roadshows of 8–10 people.

The "big-S" process (up to six months, design-led) is future-backward: generate three distinct futures, build prototypes, test with users, push winners into live experiments. At VRChat both streams run in parallel and merge into one roadmap.

Zynga shows strategy encoded into operations: viral loops, paying to complete progressions (not skip), cross-promotional network — three pillars across every studio. Meta shows identical process, opposite outcomes: Oculus referrals thrived, Portal Memories did not. Strategy accumulates value only through execution; the discipline is honesty about what to double down on and what to cut.

Tobi Lutke

Shopify's founding insight was that 2004 e-commerce software was built for existing retailers porting complex business logic online — optimized for RFP checkboxes, not for a scared first-timer building something on a lunch break. Simplicity scales up; complexity doesn't. Every time Shopify makes something measurably simpler, more businesses survive that would have quit at a confusion point.

Shopify runs without OKRs because Goodhart's Law is the business equivalent of overfitting: 80% of product value is unquantifiable. Fun and delight are leading indicators — if metrics are down but the team is having more fun, the metrics will follow. Data-informed, not data-driven; pilots decide, instruments don't.

First-principles thinking is continuous re-derivation: treat every decision like a pure function rerun over updated state. When COVID flipped "are people allowed to leave the house?" the entire decision tree moved. Shopify went fully remote not as crisis response but because re-running the function produced a different stable output. The "Tobi Tornado" — killing a project fast and reconstituting the team as founders of the next version — applies the same logic to time compression. He craves contradiction, hunting for the divergent foundational assumption; if everyone agrees, he plays devil's advocate and becomes the end boss.

A 100-year frame forces you off tactics onto the positional game: territory, trust, merchant dependence, brand. Companies that over-tactic extract all the positional value they built and are left with nothing. Stripe and Shopify bet on each other when both were tiny; coordinate-always over decades beats any single defection.

Career advice: find the intersection of three things you know unusually well, treat yourself as a product, and identify where your talent stack makes you too good to ignore — not for financial optimization, but because curiosity compounds and enthusiasm is the best marketing.

Strategy Blocks: An operator’s guide to product strategy

Product strategy occupies the void between mission/vision and the roadmap, its job being to force a disciplined choice of where to deploy scarce resources. Chandra Janakiraman's Strategy Blocks framework splits this into two parallel tracks: small "s" (2-year, problem-solving) and big "S" (3/5/10-year, aspirational).

The 2-year track runs 8–12 weeks. Preparation (3–5 weeks) has a cross-functional working group produce behavioral meta-analyses, UX research syntheses, leadership interviews, and competitive analysis. The strategy sprint (1 week) distills that into 50–150 problem statements, clusters them into 10–15 themes, then scores each on four dimensions: expected impact, certainty of impact, clarity of levers, and uniqueness of levers. The top three by total score become the strategic pillars; a "2-year winning aspiration" headline is drafted on day 3. A design sprint produces illustrative concept-car-style artifacts per pillar — not shippable designs, but enough to make the strategy legible to leadership. The product lead then writes the document solo, followed by a gated rollout: 1:1s with gatekeepers, group stakeholder review, then team roadshows of 8–10 people.

The 3/5/10-year track, led by a senior design leader over up to six months, generates three distinct future scenarios, builds prototypes, tests them with users to identify resonant components, then feeds those winners into the active roadmap.

The two tracks run simultaneously, building a bridge from both ends. Execution need not wait for strategy; strategy sharpens with execution feedback, and value is proven only through results.

Why you’re so angry at work (and what to do about it)

Workplace anger signals an unmet emotional need — it is the alarm, not the fire. Executive coach Natalie Rothfels, drawing on Internal Family Systems and Buddhist philosophy, argues that professionals suppress or misdirect anger because they lack a framework for it, driving burnout, passive aggression, and escalating conflict.

Anger at work erupts when expertise is challenged, autonomy threatened, values compromised, or identity dismissed. It manifests in two physical patterns: hot anger (tight chest, racing heart, aggression outward) and cold anger (frozen limbs, withdrawal, self-criticism). Both are unconscious reactions that compound the original harm.

The four-step framework: (1) Recognize the alarm — notice bodily signals before speaking; (2) Do a U-turn — pivot inward with curiosity rather than building a case against the other person; (3) Identify the unmet need — from a set including autonomy, dignity, fairness, and belonging; (4) Choose a conscious response — speak from the need, not the anger. Tara Brach's RAIN tool (Recognize, Allow, Investigate, Nurture) operationalizes step two in the moment.

The case study tracks Etienne, a product leader whose peer Corey was added to his project. His surface resistance was operational arguments; the underlying need was for his contributions to remain visible and attributable. Once he named that, he had a different conversation with his manager about recognition — skipping the fight about Corey entirely.

Inside monday.com's transformation: radical transparency, impact over output, and their path to $1B ARR | Daniel Lereya (Chief Product and Technology Officer)

monday.com's path from $4M to $1B ARR rests on three operational shifts.

The first is radical transparency. Before going public, monday displayed live churn, signups, and revenue on dashboards visible to everyone, including job candidates. Advisors predicted it would demoralize teams in downturns; the opposite happened — people spotted problems and felt like partners. After IPO, PMs sign pre-arranged 10b5 stock-sale plans so insider-trading rules don't force secrecy.

The second is impact over output. When a competitor shipped 30 column types while monday had 5 (each taking four months), the team set a goal of 30 columns in one month and hit it in six weeks by defining shared infrastructure first, then running a hackathon where each engineer built one column in a day. This became standing practice: every team tracks a daily-numbers Slack update; PM success is measured by whether the metric moved. When AI Blocks had strong feedback but only thousands of accounts using it against 250,000 eligible, two weeks of legal work opened access to 98% — the quarter's highest-impact move was paperwork, not code.

The third is time-boxed delivery. Bottom-up planning is fear-driven: engineers surface every risk, scope compounds, and a three-week feature becomes two years of work. Hard external deadlines force teams to strip to core value. An enterprise-work-management alpha drew "premature" feedback, which Lereya called a win: actionable complaints beat "amazing product" praise — the latter just means you over-built.

At the strategic level, monday announced five new products simultaneously. Monday Sales CRM now outpaces early monday growth rates. Announcing five at once meant no single weak bet could kill the multi-product thesis.

On infrastructure: repeated board-performance crises became MondayDB, a three-year project converting a scaling liability into an enterprise-grade competitive edge. Don't just fix the problem — make the fix the strategic advantage.

How have I been complicit in creating the conditions I say I don't want? | Jerry Colonna (CEO of Reboot, executive coach, former VC)

Leaders fail not from lack of skill but from unexamined inner life — childhood baggage driving compulsive behavior, team dysfunction, and suffering success never resolves. Jerry Colonna, co-founder of Reboot, distills 27 years of executive coaching into an equation: practical skills + radical self-inquiry + shared experiences = enhanced leadership and greater resilience.

The central diagnostic question: "How have I been complicit in creating the conditions I say I don't want?" Complicit, not responsible — you're driving the getaway car. The distinction evokes agency without shame. Example: "I don't want to feel busy," yet you're unnerved when the calendar isn't packed.

Radical self-inquiry runs on four questions: the complicity question plus what am I not saying, what am I saying that isn't being heard, what's being said that I'm not hearing. Circles and peer groups let people answer honestly — entrepreneurs are socialized to perform wellness ("crushing it"), not tell the truth.

The core trap: Colonna grew up poor, found safety in his grandfather's lemon drops, and spent his 30s as a VC still feeling unsafe — money was meant to buy the feeling, not provide it. Buddhism's second noble truth: what we do to push away suffering increases it. The couch bought to prove self-worth amplifies the fear of losing it.

On teams: Jung — "Until you make the unconscious conscious, it will direct your life and you will call it fate." Teams replicate family-of-origin dynamics. A CEO who can't tolerate wrong decisions unconsciously hires people who never decide alone; refuse to do the inner work, and the organization becomes a manifestation of that unresolved past.

On growth mindset: fine as a concept, dangerous when the ego turns it into a fixed rule. Senge: "It's virtually impossible to challenge the assumptions that made you rich." Hold it loosely.

Five principles for successfully managing managers

Manager problems almost always trace to the skip lead — the manager's manager — because the skip hires, reviews, and coaches line managers. Every manager failure is at least indirectly a skip-lead failure.

The leap from line manager to skip lead is as steep as IC to manager, but rarely treated that way. Working through an abstraction layer without direct IC access is a fundamentally different job.

The "API endpoint experiment" forces the right questions: how would you run the org interfacing only through direct-report managers? It clarifies when to override, how to coach without doing the work, and what to delegate.

Two symmetrical failure modes: undermining — solving IC problems on the spot disempowers line managers and creates conflicting direction; covering — absorbing a struggling manager's work instead of surfacing underperformance up the chain.

Delegation uses Andy Grove's task-relevant maturity (TRM): match each project's scope, ambiguity, and risk to the right manager. Post-delegation review scales on uncertainty × impact — high on both demands close, ongoing oversight.

How to build a team that can “take a punch”: A playbook for building resilient, high-performing teams | Hilary Gridley (Head of Core Product, Whoop)

Product leadership destroys people who aren't in control of the voices in their head. Hilary Gridley's playbook for resilient teams has three interlocking ideas.

The core tactic is "counter-programming the narrative." When someone feels ego-injured by a colleague's perception, the instinct is to litigate — explain, defend, correct the record. That reads as defensive and amplifies the failure. Gridley's substitute: ask "what's one small action that demonstrates the opposite of what you're afraid this person thinks?" She illustrates with her own humiliation — laughing at WHOOP's CTO's serious ketamine-tracking proposal, then writing a note connecting it to sports-betting public health data. Counter-program, don't defend. The technique is borrowed from CBT's behavioral activation: act to feel better, not the reverse. Second-order effect: people who can absorb a punch become willing to take more swings — speaking up in meetings, tackling harder problems.

The second thread is radical downward transparency. Instead of sharing what leadership decided, Gridley shares verbatim what leaders said plus her interpretation of the mental model behind it. Her example: WHOOP's CEO pushing for things that "feel like the future" doesn't mean maximum scope — it means high-impact, low-cost touches that surface unique data magically. When the whole team holds that model, the ten-person approval chain disappears.

On habit formation, she treats AI adoption as behavioral design: consistency (one tiny task daily), friction reduction (vacation planning rather than deadline-critical work), and immediate emotional reward loops. She built a LSAT-style reasoning GPT called "Aristotle" that drills PM judgment through product scenarios, shrinking feedback loops that otherwise take months of on-the-job reps.

Unifying thesis: you are not the protagonist in someone else's company. Your job is to understand the CEO's vision well enough to operationalize it — then make the thousands of micro-decisions where your taste makes the product distinctly yours.

25 proven tactics to accelerate AI adoption at your company

The biggest barrier to AI adoption is organizational, not technological — vague mandates, procurement friction, and no guidance on where to start. Five levers work: explain the how concretely (Shopify named specific behaviors; Zapier gave staff a week off with a playbook); track and reward usage (Ramp publishes power-user counts by team; Intercom sees ~20% YoY productivity gains); cut red tape (Duolingo gave each employee $300 for AI tools); turn enthusiasts into teachers (Zapier's weekly demos draw 60+); target high-volume tasks first — Duolingo went from 100 to 150 courses in 12 months. Real adoption means demanding evals, not flashy demos.

How we restructured Airtable's entire org for AI | Howie Liu (co-founder and CEO)

Every software company must be "refounded" for AI — not patched. Model improvements keep implying novel UX patterns, so you have to continuously re-find product-market fit. The test: "If we were starting from scratch today, what would we build?" If your existing product gives no advantage, find a buyer and start fresh. Airtable's answer is that its no-code primitives — collaborative real-time data, view types, automation — give an AI agent a reliable DSL for building business apps rather than generating brittle raw code, solving the context-collapse problem that plagues vibe-coded apps at scale.

The reorg split EPD into two groups drawn from Kahneman: a fast-thinking AI platform team chartered to ship jaw-dropping capabilities weekly (the Cursor benchmark), and a slow-thinking group handling HyperDB-scale infrastructure that can't be prototyped in a week. They compound: fast-thinking drives top-of-funnel excitement; slow-thinking converts that adoption into durable enterprise growth.

For the IC-CEO shift, Howie cut standing one-on-ones and replaced them with urgency-driven meetings seeded by real alpha — a prototype, a competitor insight, LLM map-reduce across a year of sales transcripts. He claims Airtable's highest per-user inference cost: hundreds of dollars in compute buys "chief-of-staff quality" insight across the sales corpus — cheaper than consultants.

Team-skill prescription: everyone in PM, engineering, and design needs a minimum baseline in all three disciplines. Designers must understand tool-calling constraints; PMs must prototype rather than write PRDs; engineers must carry product taste. The collapse-the-role logic extends to sales (AEs must be SE-fluent) and marketing. Play — genuine curiosity, not box-checking — develops these instincts. He gave teams explicit permission to cancel a week of meetings to explore AI products.

On evals: start with vibes. Defining success criteria too early constrains discovery. Evals become useful once you've converged on which use cases work — then iterate empirically.

Introducing the GAIN framework for feedback: an evidence-based approach to giving feedback that people love, appreciate, and act on

Framing feedback around what someone stands to gain — rather than what needs to stop — produces more behavior change, stronger relationships, and less defensiveness. Executive coach Jack Cohen's GAIN framework does this in four moves: name the shared Goal, describe specific Actions and Impacts as observations rather than judgments, close with concrete Next actions (who does what by when).

The evidence base is broad. John Gottman predicted divorce with 93% accuracy from three minutes of conflict; stable couples named the "dream within the conflict" rather than just the complaint. A Tversky study found doctors given "90% survival rate" framing made optimal decisions 84% of the time versus 50% with the identical "10% mortality rate" — same information, opposite effect. Edmondson's research across 16 organizations showed aspirational framing far outperformed defensive framing in adoption. A Stanford randomized trial found that adding "I have high expectations and I know you can reach them" quadrupled revision rates and improved quality 26%.

The Actions and Impacts section carries a counterintuitive warning: flattering judgments ("you're a rockstar PM") damage growth mindset as surely as critical ones, shifting focus from impact to image. Dweck's research explains why — if brilliance caused success, it must explain failure too. Observations of concrete actions point people to what they can control. Acknowledging your own contribution preempts "but what about you?" resistance and reframes the conversation as shared problem-solving.

For Next Actions, a meta-analysis of 52 studies found adding "feel free to say no" doubled the yes rate. Frame new behaviors as time-bounded experiments to reduce commitment anxiety. Vague closings ("let's," "should") need resolving into who, what, and when.

Three AI applications close the framework: FeedbackGPT to translate raw reactions into GAIN-structured drafts; Voice Mode role-play to rehearse the dialogue; post-conversation transcript analysis to loop back into practice.

Slack founder: Mental models for building products people love ft. Stewart Butterfield

Product craft is not about removing friction — it is about eliminating the need to think. Butterfield, who built Flickr then Slack (sold to Salesforce for $27.7B), frames this around two lenses. Utility curves describe value as an S-curve against effort: a feature that doesn't cross a quality threshold produces nothing; the question is whether you're on the steep part or the plateau of diminishing returns. The bar also rises continuously — Bezos's "divine discontent" — so rarely-revisited flows like account creation silently decay.

The friction confusion: reducing clicks misdiagnoses most UX problems. High intent with a clear goal (Taylor Swift tickets on Ticketmaster) makes friction the right target. Low intent with fuzzy understanding (a new visitor) means the bottleneck is comprehension. Forcing a decision users don't understand burns cognitive ATP and makes them feel stupid. The rule: "Don't make me think."

Taste is learnable but undervalued. Slack's craft bets — magic-link auth, a rooster warning before @channel blasts 147 people, a tiered Do Not Disturb rollout — cost real engineering time but compounded as word-of-mouth fuel: growth came from someone at startup A joining startup B and insisting they switch.

The owner's delusion explains why products ship incomprehensible: builders have internalized context visitors lack, forgetting the user is a fraction of a second from bouncing.

Organizational bloat follows Parkinson's Law: managers hire because headcount signals advancement, generating hyper-realistic work-like activities — previewing the deck before the meeting that reviews the deck — indistinguishable from real work. Fixing this is leadership's responsibility. On pivots: cold rationality is required because persisting feels less humiliating than admitting failure; pivot only after exhausting all plausible ideas.

The mantra Butterfield had Slack employees chant aloud: "In the long run, the measure of our success will be the amount of value we create for customers."

A guide to difficult conversations, building high-trust teams, and designing a life you love | Rachel Lockett

Technical leaders who always provide answers train their teams to bring every problem to them. The fix is coaching: stay curious, ask questions, let people solve their own issues. Two skills carry this — level-three "global" listening (body language, tone, subtext beneath words) and the GROW model (Goal, Reality, Options, Way Forward), which unlocks solutions without prescribing one. Advise outright only when the issue is urgent or a specific action is needed.

Burnout tracks with role-to-strengths misalignment. The diagnostic: two weeks of nightly journaling — five energy-giving things, five draining things — then reading the pattern. Target 80% of work time in your gifts. Tell your manager what energizes you, hire around weaknesses, move laterally if needed. Navigating your career is nobody else's job.

Co-founder relationships fail from conflict at a 65% rate. The framework mirrors couples therapy: shared self-awareness first (Enneagram or a direct strengths conversation), then regular time outside the grind — weekly check-ins, quarterly in-person. One distressed PR duo found geographic distance had severed mutual support; recommitting to a rhythm repaired it. A co-founder choosing to leave counts as success — clarity over unexamined drift.

For difficult conversations, the goal is mutual understanding, not winning. Marshall Rosenberg's Nonviolent Communication: Observation (factual, photographable), Feeling (a real emotion — not "I feel like you..."), Need (universal unmet need), Request (small, actionable). Entering conflict asking "how am I complicit in creating conditions I claim I don't want?" — Jerry Colonna's framing — replaces blame with agency.

Operationally, a one-page plan columns vision/values, strategic intentions/KPIs, annual goals, and quarterly goals so every layer traces work to founding purpose. Alpine Investors' data shows portfolio companies running this rhythm generate higher returns. It requires quarterly "balcony time" — stepping outside execution to surface inconvenient truths about what is and isn't working.

10 contrarian leadership truths every leader needs to hear | Matt MacInnis (Rippling)

Extraordinary outcomes demand extraordinary efforts — not as inspiration but as a logical constraint. To reach the 99th percentile, you must run at the red line continuously — in any valuable market, competitors are always looking for the crack you leave open. Matt MacInnis, CPO and former COO at Rippling ($16B), argues that on a genuinely winning team, demanding the last ounce of oil is a gift, not a tax — the intensity becomes invigorating rather than crushing.

Several management principles follow. Deliberately understaff every project: overstaffing produces politics and work on low-priority items. Processes exist to lower beta (output volatility), but they suppress alpha (creative upside) — apply them surgically where you need reliability. Rippling's PQL ("product quality list," pronounced "pickle") is a lightweight factory-inspection checklist before every ship; when a forgotten feature flag caused Parker Conrad to hit a blank screen during an install, a new line was added that day.

On product market fit: you absolutely know it when you see it, and if you don't know, you don't have it. The "never quit" Silicon Valley doctrine serves VC incentives — they can't take their money back — not founder interests. MacInnis spent nine years at Inkling and could have called it around year four. Drug-receptor analogy: no marketing changes whether binding receptors exist; fate has already decided the outcome. Run the experiment — don't try to convince the body to develop receptors for the wrong drug.

On learning: you learn more from success than failure — join a winning team early. On feedback: withholding it is the most selfish act a leader can perform, optimizing for your own comfort at another's expense. Escalations from customers are gifts.

Books: Fred Kofman's *Conscious Business*, Donella Meadows's *Thinking in Systems*, Drucker's *The Effective Executive*.

The high-growth handbook: Molly Graham’s frameworks for leading through chaos, change, and scale

Scaling companies rip apart what you've built and hand it to someone new — that's the opportunity, not the problem. Molly Graham spent five years at Facebook (80 million to 1 billion users, 500 to 5,500 employees) and bootstrapped the Chan Zuckerberg Initiative from 30 to 250 people in year one.

The core metaphor is "giving away your Legos." Every time you master a domain, scope expands and your work gets handed to a new hire. The territorial reaction is what Graham calls Bob: an externalized monster whose job is to make you your worst self. Let Bob run without acting on him. Feelings under two weeks are Bob; anything longer deserves a real conversation. The career corollary is the J-curve: jumping into a role you're unqualified for produces a six-to-nine month fall, then a climb past what steady promotions reach. Graham moved from HR to running hardware at Facebook — the phone failed, her career didn't. Financial fear is worth heeding; fear of incompetence is a green light.

The Waterline Model diagnoses team problems at four depths: structural (goals, roles, expectations), dynamics (culture, decisions), interpersonal, and intrapersonal. Eighty percent are structural, but managers instinctively dive to the people layer. Rule: snorkel before you scuba.

On goals: Facebook ran on three for five years — MAUs, engagement, revenue. No company needs more than three. One must win in a fight. One goal, one owner; two owners equals none. Prioritization only counts when it hurts. Goals without a follow-up process change nothing.

From the founders: 80% of culture is the founder's personality, not values documents. Escalation is a tool — go up together when two equals are stuck. More than doubling headcount annually guarantees role duplication and drag. When politics cloud a decision, what's right for the business is the answer.

A child psychologist’s guide to working with difficult adults | Dr. Becky Kennedy

Humans of all ages need the same things, and bad behavior is always feelings overpowering skills — not a character flaw. Child psychologist Dr. Becky Kennedy applies that principle to corporate life, where most workplace dysfunction follows the same structure as a toddler meltdown.

Secure attachment — in families and on teams — is defined not by perfection but by willingness to repair: going back to say "I cut you off in that meeting and I'm sorry." Repair rebuilds trust faster than avoiding rupture.

The core operating error is collapsing behavior and identity. Treating lateness as laziness triggers defensiveness that makes the behavior undiscussable. Starting with "we're on the same team and something's going on" opens it. The Most Generous Interpretation (MGI): replace "this person is checked out" with "this person might not feel heard," then address that hypothesis.

Connection makes correction possible. Mindset matters more than words: agenda-free presence for 30 seconds reads entirely differently from transactional warmth deployed to extract compliance.

Sturdy leadership means holding your footing while validating someone else's experience — neither dismissing it ("stop overreacting") nor being overwhelmed by it ("everyone vote on the emergency landing"). People feel safe around leaders they can locate: clear position, doesn't dissolve under pressure.

A boundary is what you will do, not what you ask the other person to do. "Don't press the buttons or no dessert" is a request; "I'm standing between you and the buttons" is a boundary. Requests cede power to whoever you're asking.

Optimizing for short-term comfort wires fragility. The combination that works is "I believe you" (one foot in the hole with them) plus "I believe in you" (holding a more capable version than they can access right now). Withholding feedback steals the experience of capability — the thing that compounds over time.

How to debug a team that isn’t working: the Waterline Model

Blaming people for team dysfunction is usually wrong — most problems are structural. The Waterline Model, developed by Molly Graham from her time leading Google and Facebook teams, diagnoses issues at four ordered layers and mandates "snorkel before you scuba": check the shallowest layer first. Structure comes first — goals, roles, metrics. When Graham inherited a struggling marketing team, individuals couldn't state what they owned or what success meant; clarifying structure alone fixed most performance gaps. Dynamics is next — the implicit behavioral rules. A founder who reverses decisions after the fact teaches the team to optimize for safety over speed; the slowness isn't a performance problem, it's rational adaptation. Only after ruling out those two systemic layers should leaders address interpersonal tension or individual fit. Once the system is sound, clean calls on individuals — coach, change the role, or exit — are kinder than ambiguity.

Hard truths about building in the AI era | Keith Rabois (Khosla Ventures)

The team is the company — talent density is the variable that makes every other variable tractable. Rabois traces this to PayPal, where Thiel and Levchin built 12–17 "barrels" among 254 people, seeding 25 years of derivative companies.

The barrel/ammunition framework names the core scaling bottleneck: only a handful of people can take an initiative from inception to completion without supervision. At Loudcloud, the count was two. Hiring more ammunition without adding barrels just creates coordination drag. A barrel is tested by telling someone "get us over that hill" and firing and forgetting. Standard interview loops can't identify them; references are the workaround — DoorDash's Tony Xu does 20 per senior hire. The right question isn't "was this person a good employee" but "is this person capable of being world-class at X."

Consumer customer feedback is directionally dangerous: subconscious purchase decisions surface as rationalizations when consciously examined. Airbnb's signal wasn't surveys; it was 30 Craigslist listings of people already renting rooms. Enterprise is the exception.

AI collapses the product triad. The PM role as roadmap intermediary becomes incoherent when a year-long roadmap is obsolete within weeks. The skill that survives is CEO-level business acumen — knowing what to build and why. Indicator at portfolio companies: the heaviest AI token consumer is the CMO, shipping analytics and campaigns directly without deputy layers.

Winning companies share one early signal: operating tempo. Ramp shipped a card product in three months; the industry floor is nine. Speed compounds into the accumulating advantages Rabois demands founders articulate, even before they can demonstrate them.

Management: push hardest when things go well, support when they struggle. Talented people's morale drops when coasting. Criticize in public so the whole team understands an issue is being addressed — private feedback optimizes the individual at the expense of the system.

28 Ways to Grow Supply in a Marketplace 📈

Airbnb scaled from 100,000 listings in 2012 to 6 million not through any single lever but through many small wins compounding — most things tried didn't work, but enough did.

The highest-impact cheap tactics: showing a realistic earnings estimate on the host landing page was Airbnb's single biggest conversion driver — hiding it always caused growth to dip. Multiplying entry points throughout the product compounds this; being too aggressive here is nearly impossible. The host referral program became the most efficient channel of all, driving the largest share of attributable supply at the highest quality. Direct sales — calling, emailing, door-to-door — is heavy but essential for bootstrapping new behavior and remains evergreen for B2B; hand-holding early hosts builds loyalty and surfaces product insight. Piggybacking existing networks is the fastest cold-start move: Airbnb recruited from Craigslist; Etsy seeded from getcrafty.com and Crafster.org.

Mid-tier levers: converting demand-side users to supply ("pay for your trip by hosting"), performance marketing once LTV is known, acquiring competing companies (Airbnb targeted 60–80% migration rates when buying Crashpadder and Statthotel), and early meetups — whose impact Airbnb could never measure but, in retrospect, clearly mattered.

Structural strategies: reduce host costs and increase benefits explicitly (Host Guarantee and free photography were "reduce cost" moves); build single-player mode so the platform is useful before demand arrives (OpenTable did this with restaurant management software); establish trust early (Airbnb employees traveled free in exchange for reviews on new listings; Joe Gebbia's research found ten-plus reviews let reputation fully override stranger-distrust). Nate Blecharczyk calculated that 300 listings — 100 reviewed — was the critical mass where booking growth inflected. Airbnb's first-booking-within-30-days activation target proved more predictive than sign-up volume alone.

Focus beats breadth: pick a few tactics, double down on what shows promise, and never lose sight of the north star.

This Week #2: Tackling the chicken-and-egg problem, building a growth team from scratch, and addressing overlap with PM peers 🤔

Cold-start marketplaces resolve the chicken-and-egg problem through five tactics: single-player mode (OpenTable seeding restaurants before diners), paying early participants (Uber/Lyft driver salaries), employees as early supply (Airbnb founders hosting), geographic focus (Rover in Seattle), or niche focus (eBay with Beanie Babies). PM/TPM role tension: talk directly first, then escalate with concrete examples, reasoning from ideal roles not historical ones. Growth teams should target 2–10x levers early; once retention is healthy, top-of-funnel (referrals, SEO, paid) is where the wins are.

How to kickstart and scale a marketplace business

Every new marketplace faces the chicken-and-egg problem: supply won't join without demand, and vice versa. The near-universal solution, from interviews with 17 major marketplaces, is constraint — go small to get big. Location-based services constrain geographically: Uber targeted 30+ drivers per city for sub-15-minute ETAs; OpenTable needed 50-100 concentrated restaurants before diners found value. Non-location marketplaces constrain by category: Etsy launched with only three (vintage, craft supplies, handmade); Eventbrite focused on tech mixers. The exception is Thumbtack — deliberately broad from day one, because narrow focus would have cut usage from 8-12 household hires annually to once every few years. Product-market fit matters more than any growth tactic.

How to Kickstart and Scale a Marketplace Business – Part 2: Cracking the Chicken-and-Egg Problem 🐣 - Supply vs. Demand

In early-stage marketplaces, supply almost always deserves priority over demand: 14 of 17 companies studied (Airbnb, Eventbrite, Lyft, DoorDash, Etsy, Thumbtack, and others) put nearly all resources into supply and found it generated its own demand. The three exceptions — Rover, TaskRabbit, and Zillow — had naturally abundant or public supply, so demand became the constraint. Patreon discovered it wasn't a marketplace at all; dropping the "discovery" framing let it build a creator platform instead.

How to Kickstart and Scale a Marketplace Business – Part 3: Cracking the Chicken-and-Egg Problem 🐣 - Growing Initial Supply

Marketplaces stay supply-constrained not just at launch but throughout most of their history — strong demand always outpaces supply. The meta-finding across 17 companies: the median marketplace used only two supply-growth levers early on. Focus beats breadth.

Direct sales was the most common lever (~60% of companies): Airbnb sent local teams door-to-door, OpenTable carried software into restaurants, DoorDash pounded the pavement, Uber cold-called limo companies. Referrals ranked second, tied with Craigslist piggy-backing: at Uber, referrals drove one-third of early driver signups and those drivers were the highest quality; Lyft's ambassador program seeded each new city before launch. About a quarter subsidized supply — Uber guaranteed $40/hour, Lyft set income floors. Employees served as literal first supply at Rover, TaskRabbit, and DoorDash. Single-player mode was rare but decisive: OpenTable sold restaurant software first, demand second; Eventbrite's free-to-paid loop drove 34% of supply acquisition and converted 17% of free creators to paid within 12 months. Performance marketing and SEO rarely mattered at this stage.

How to Kickstart and Scale a Marketplace Business – Part 4: Cracking the Chicken-and-Egg Problem 🐣 - Growing Initial Demand (plus a Bonus!)

Word of mouth drove over half of early demand at the biggest marketplaces — Airbnb saw 50%+ on the guest side and 70%+ on the host side, Uber ran 50–60% WOM early on, and TaskRabbit credited 90%+ in its first years. For the other half, supply pulled demand directly: DoorDash restaurants put up window stickers unprompted, Etsy sellers marketed their own shops, and Patreon creators brought pre-built audiences.

SEO mattered for over 40% of marketplaces. Thumbtack built 80–90% of its growth on SEO after a co-founder met an expert in a bar who became a board member. GrubHub made it the top channel (30% of new users) by creating landing pages for restaurants with no web presence. Zillow only pivoted to SEO after watching a no-name competitor outgrow it on organic traffic.

Performance marketing worked for about a third: Airbnb bought home-stay keywords, Rover targeted kennel searches, and Breather ran a Twitter ad feeding a manual seven-step email sequence that personally onboarded thousands. PR drove Zillow's launch via controversial home-valuation data released during the 2008 recession. Referral programs contributed roughly a third of early Uber and Instacart demand. Direct sales ranked eighth for demand versus first for supply — Lyft went door-to-door at startups with cupcakes; Uber stationed street teams at Caltrain. Companies worked a median of three effective demand levers versus two on the supply side.

What didn't work: paid ads arrived years too early for Thumbtack, TaskRabbit, and Instacart; viral loops rarely materialized. The closing pattern: unscalable hustle bought runway — DoorDash's PDF-menu fliers, Breather's email-by-email onboarding, dashers running mall elevators — until the product could take over.

How To Know If You're Supply or Demand Constrained 🤹‍♂️ - Phase 2 of Kickstarting and Scaling a Marketplace Business

About 40% of marketplaces remain supply-constrained indefinitely — Uber targeted keeping surge pricing below 20–30% of trips as a proxy; Lyft found a 3-minute ETA was the inflection point where conversion and retention improved sharply. Only a handful (Rover, TaskRabbit) are durably demand-constrained, because supply volunteers itself when the value proposition is obvious. The interesting 40% see imbalance shift by geo or category: GrubHub tracked orders-per-restaurant to flag under-supplied markets; Thumbtack used Hire Rate (60% of searches returning 3+ quotes = healthy); Airbnb evolved from occupancy rate to a full econometrics model comparing incremental revenue per unit of added supply vs. demand.

Accelerating Growth at Scale 🔥 Phase 2 of Kickstarting and Scaling a Marketplace Business

At scale, the 16 growth levers available early collapse to 8, and performance marketing dominates — 70% of the 17 marketplaces surveyed (including Uber, Airbnb, DoorDash, Zillow) rely on it as their primary channel. Geographic expansion matters for 65%; conversion optimization and SEO each for 50%. Direct sales shrinks in importance as scalable channels outrun the cost of a sales org. Referrals drive 10–15% of new users at Uber and Airbnb. Loops remain critical for Uber (riders-to-drivers "R2D funnel") and OpenTable (restaurant-website diner acquisition). No silver bullets — just compounding lead bullets across channels.

Maintaining Quality 🏅 Phase 2 of Kickstarting and Scaling a Marketplace Business

Quality problems in marketplaces are almost entirely supply-side, and they never get solved — only managed, with increasing sophistication as scale grows. Ten tactics emerge from interviews across Airbnb, Uber, DoorDash, Rover, and a dozen others. The most common: publish clear standards with penalties (Uber warned drivers below 4.4; DoorDash lowers commissions for slow pickup). Manually onboard early supply — Airbnb ran a 12-point checklist for every new listing; Instacart physically bought one of every store item to build image catalogs. Reviews build trust. Subsidize the experience when transactions are small (Breather over-invested in room design; Instacart redelivered full orders at cost). Search ranking is cheap and powerful — Rover used "where does the dog sleep?" as its top predictive onboarding signal, boosted those sitters, and pulled ahead of competitors on LTV. Adding onboarding friction filters supply when quantity outstrips quality. Etsy accidentally set photography standards by only featuring well-shot Treasury items — sellers mimicked the aesthetic without being told.

What They'd Do Differently 🔮 Kickstarting and Scaling a Marketplace Business

Scaling before product-market fit is the most repeated mistake. TaskRabbit's Brian Rothenberg chased top-line growth while fill rate — tasks successfully matched — sat well under 100%; unmatched users never came back. Etsy devoted hundreds of engineers to infrastructure while three people owned search, which drove 100,000 daily item sales versus 700 for everything else. Recurring themes from Airbnb, Lyft, GrubHub, Rover, and Eventbrite: split supply and demand growth teams earlier; structure supply data from day one rather than retrofitting taxonomy later; maintain scrappy weekly A/B testing longer; build supplier empathy before imposing policy; and expand adjacencies (Rover into dog walking) before a clone (Wag) fills the gap. PMF mattered more than any growth lever.

This Week #12: Expanding your business internationally 🌏

Ignoring international expansion lets foreign competitors threaten your home market — HelloFresh ($4B) vs. Blue Apron ($50M) and Spotify vs. Rdio illustrate this. First-pass target markets almost always land on UK, Germany, France, Canada, and Australia. Execution runs four dimensions in order of difficulty: business model (regulatory fit, take rates), product (localization, payments), distribution (revive your v1 liquidity playbook; use divide-and-conquer — achieve $10K GMV/30 days in one city before splitting or expanding), and organization (document obsessively, avoid long-term process investment early).

Building a referrals program

Referral programs only work where word-of-mouth already exists — the incentive adds fuel, not fire. Airbnb's host referral program evolved from a $25 flat travel credit to a $100 cash award with LTV-personalized amounts, eventually becoming its single biggest attributable supply channel and most efficient paid channel. Fit requires large existing userbases, meaningful incentives (under-10% referral conversion is typical), and users who know peers in the same role. Pitch the program everywhere, model cannibalization conservatively, and make fraud costlier than the payoff.

How the biggest consumer apps got their first 1,000 users

Every major consumer app's early growth traces to one of seven strategies, and most used only one. The dominant pattern is direct, unscalable contact with target users — not ads, not virality.

Going offline meant showing up where users gathered: Tinder's team ran sororities at USC; DoorDash taped flyers to Stanford lampposts; Lyft dropped free ice cream at startup offices; Etsy recruited sellers at craft fairs. Going online meant infiltrating existing communities: Netflix's Corey Bridges spent months undercover in DVD forums before launch; Musical.ly gamed App Store search by stuffing the app name with keyword phrases; Drew Houston posted a Dropbox demo to Hacker News in 2007.

Personal networks seeded durable products. Yelp tapped PayPal alumni; LinkedIn seeded with Reid Hoffman's accomplished friends to make the network aspirational; Quora launched to college friends, then let invitations pull in startup contacts.

Manufactured scarcity drove word-of-mouth: Instagram pre-seeded with designers who had large Twitter followings, building demand before launch; Robinhood's waitlist showed how many people stood ahead of you; Superhuman used a two-question intake to build both list and product insight. Spotify gave each user five invites.

Influencers mattered: Om Malik's early Twitter post generated 250 signups when the network had under 600 users; Dorsey's day-one Instagram shares pushed it to #1 in camera apps. Press worked for Airbnb through election-themed cereals sent to bloggers, and for Slack through a launch blitz generating 8,000 invite requests day one. Product Hunt built as an email list first, shipped in five days, and emailed only contributors who had already shaped the design.

Evaluating a (marketplace) business idea

Most marketplaces fail for generic business reasons — weak PMF, wrong market, poor distribution — not marketplace mechanics, so evaluate the business first. Seven general criteria: product-market fit (high retention, 10x better than alternatives), large market ($1B+ opportunity), a "why now" trigger (tech inflection, regulation shift), distribution with high LTV/CAC, earned-secret founders, moat (network effects, switching costs), and unit economics. Then seven marketplace-specific tests: demand-side PMF (customers desperate for it), supply-side PMF (suppliers meaningfully benefit from underutilized assets), quality at scale (fill rate, 5-star rate), high frequency plus adequate AOV, on-platform retention via convenience or protection, non-monogamous demand (doctor-style one-match markets kill recurrence), and fragmentation preventing easy direct matching.

Magical growth loops

The most capital-efficient growth comes when existing users recruit new users, eliminating the need to acquire every user yourself. Four distinct loops achieve this: supply driving demand (DoorDash restaurants promote the app to customers; Cameo celebrities share their profiles; Substack writers bring their own audiences); demand driving supply (Airbnb guests become hosts; Faire retailers recruit their vendors); demand driving demand — either viral (Dropbox shared folders, Figma co-worker invites, WhatsApp) or passive (Zoom meeting links, PayPal money transfers, Superhuman email signatures); and supply driving supply (Typeform survey recipients become survey creators; Eventbrite attendees become hosts).

Choosing a take rate

Marketplace take rates are driven by three factors: demand generation, convenience provided, and competitive intensity — summed as Take Rate = Convenience + Demand − Competition. Platforms that don't bring new demand (Substack, Gumroad) cluster at 5–15%; marketplaces that do (Airbnb, Cameo) run 10–50%. Start at 10% (platform) or 20% (marketplace), then adjust for convenience and competition. Bill Gurley's counter-argument: in winner-take-all markets, a greedy rake creates friction and hands competitors an opening — lower fees can be the durable advantage.

Why marketplaces fail

Marketplaces require finding product-market fit twice — for demand and supply — and failing either kills the business.

Five failure modes dominate. No demand-side PMF: Neighborrow users praised tool-rental but never booked; Kitchensurfing tried advance-booking and on-demand and couldn't sustain either. No supply-side PMF: Prim's laundromat partners churned once they competed with its delivery; Exec couldn't retain errand runners because competent people avoid part-time work. Lack of liquidity: Threadflip's narrow brand focus left too few unique listings; Zaarly's request model gave buyers no certainty of supply. Bad unit economics: Shyp's flat pickup fee broke against package-size variance; Homejoy's $19 promo cleans trained customers not to repurchase. Scaling too fast: Handy cancelled thousands of bookings when demand outran supply; Laurel & Wolf's designer base collapsed after fulfilment failures.

Winning marketplaces deliver cheaper price, better product via exclusive supply (Airbnb, Cameo), or better experience via aggregation (Lyft) — while giving supply new income streams or incremental revenue on existing work.

How long it takes to find Product-Market Fit

Finding PMF takes longer than founders expect: averaging 1.5 years from launch across 12 marketplace companies — plus another year building v1. Marketplaces hold no timing advantage over other startups. Local-first companies (Lyft, DoorDash) often felt early signal in city one but needed city two to confirm durability. PMF typically showed as a sudden growth inflection, compounding retention, or a Series A close. Outliers: Caviar at 3 months, Thumbtack at 5 years — two years each building supply then demand, one year on revenue, before all three converged.

Kickstarting supply in a labor marketplace

Labor marketplaces are supply-constrained, so the early bottleneck is worker acquisition, not demand. Four channels dominate: job boards (Indeed/ZipRecruiter drove ~50% of early supply for Staffy; Thumbtack and Clipboard Health both relied on them); direct outreach (Instawork recruited cooks outside restaurant back doors; Incredible Health seeded nurses through a personal referral chain); word of mouth plus paid referrals ($300 per referred vet at Roo); and paid social, especially LinkedIn (Vettery's top performer). Secondary tactics: Tech Ladies built a community years before the product; Roo seeded hospital demand first to make supply activation credible; Hired concentrated on specific geos to build density before expanding.

How marketplaces win

Marketplaces win on the demand side through three routes: convenience (aggregating fragmented supply into a trustworthy, on-demand experience — Thumbtack, Rover), price (rare; Pinduoduo's group-buying is the purest case), or unlocking supply that had no prior market (Airbnb's private rooms, Cameo's celebrity videos). Uber/Lyft achieved all three but still have weak moats, showing quantity of advantages matters less than quality. On the supply side, the draw is always money — framed as life-changing income for people-as-supply, incremental revenue for businesses-as-supply.

Marketplace city expansion strategy

Expand to a second market as soon as your first market works — typically 6–12 months post-launch — but not before you have a repeatable playbook. Caviar waited for 100 average daily orders; Snackpass waited until 90%+ of new users came from word of mouth. In hyper-competitive categories (OpenTable in 1999, Caviar vs. last-mile rivals), the calculus flips: launch multiple cities simultaneously to block competitors.

Pick markets with a 3–5 attribute spreadsheet. Grubhub ranked population density, self-delivery restaurant count, Google search volume for delivery keywords, and a demographic score targeting young professionals. Instacart added unlaunched-city waitlists and startup-community presence. Snackpass picked across archetypes — college town, suburban, urban — to stress-test whether the model transferred.

Launch teams start tiny and sales-heavy: one or two reps sign 50 supply-side partners, hand off to a local hire, then move on. All eventually centralized demand growth while keeping supply acquisition local.

Playbooks share the same skeleton: anchor supply first with popular market-maker partners, gate demand activation until supply hits a threshold, then layer in SEO/SEM. Grubhub only bid on Adwords once a keyword had 4+ matching restaurants — the floor for CAC payback within six months.

Demand driving supply: The little-understood growth loop behind a surprising number of iconic billion-dollar companies

Airbnb, Eventbrite, Square, Lyft, DocuSign, SurveyMonkey, and GoFundMe all share a growth loop where demand-side users convert to supply-side — riders become drivers, attendees become event creators, buyers become merchants. The loop only merits investment if it's already happening organically. At Eventbrite, Roelof Botha spotted it during Series A diligence; 34% of creators first encountered the platform as attendees. Square's Keith Rabois measured roughly 1% of buyers converting to merchants — enough to spin a flywheel that carried Square to a ~$120B valuation. Typical conversion rates run 0.5–5%; Aalto (real estate) hit 11%. Because rates are small, the loop becomes powerful only at hundreds of thousands to millions of demand-side users. The main friction is awareness: an Eventbrite survey found ~50% of attendees didn't know anyone could sell tickets on the platform. Fix: add consistent branding and a supply-side CTA at high-traffic demand touchpoints — ticket confirmation pages, survey completion screens, listing pages, site headers. SurveyMonkey's "Done" page redirects respondents straight to survey creation. Eventbrite adding "Create an event" to its header was one of its highest-impact changes. Cohort analysis on short-latency windows (same week, 14-day, 21-day) provides leading indicators before the long conversion tail resolves. Segmentation — finding the behavioral tipping point where conversion probability jumps — lets teams concentrate effort on users who are most primed. Eventbrite ultimately tripled its demand-to-supply conversion rate through sustained iteration.

The Atomic Network

The smallest self-sustaining network a product needs — its "atomic network" — is the base unit from which all larger networks are built. Slack's was fewer than ten people; a credit card network required an entire city. The target is almost always smaller than founders expect: Uber's wasn't "San Francisco" but "5pm at Caltrain's 5th and King." Launch the simplest form, use unscalable boosts (PayPal's $5 referral, Dropbox's Hacker News video) to hit critical mass, then copy the playbook: Slack moved from startups to IBM; Facebook's campus launches built cross-school demand. Networked products always look like niche toys first — that's the signal, not the problem.

The inside story of Facebook Marketplace

Facebook Marketplace reached a billion monthly active users — ahead of eBay, Alibaba, and Craigslist — by solving a problem most Facebook employees didn't believe existed. Deb Liu first pitched commerce at her 2009 interview; nothing moved until 2015, partly because prior attempts (Beacon, Oodle) had failed, partly because organic buying and selling in Facebook groups was invisible to anyone not in mom groups or emerging markets where it was ubiquitous.

The team structured commerce groups with ranked feeds, then hit the twin liquidity problem: buyers couldn't find items, sellers had limited reach. They chose a dedicated Facebook tab over a standalone app to preserve existing social trust. React Native let them ship V1 in about two months.

Launch (October 4, 2016) hit a fraud-queue bug, cleared fast. Post-launch data surfaced two surprises: cars and rentals dominated immediately (real-name identity solved trust for high-stakes in-person deals), and users never searched — they scrolled, having been trained off the search bar. Adding a prominent search button permanently raised participation.

The most important marketplace metrics to track

Four metrics capture everything that matters in an early marketplace: fill rate (intentful sessions that convert), bookings (completed transactions per period), supply growth (activated supply reaching a first milestone, not raw counts), and GMV. Fill rate is the health metric — it bakes in supply quality, availability, and funnel conversion; Alex Taussig evaluates every new marketplace startup primarily on it. Bookings is the growth metric, stripping out price changes and outlier orders that distort GMV. Activated supply (e.g., a Lyft driver with one completed ride) protects against teams gaming raw numbers. "Liquidity" is deliberately avoided here — the term means something different to every practitioner, and fill rate is the more precise substitute.

How to build trust in a marketplace

Cracking the evolutionary distrust of strangers separates thriving marketplaces (Airbnb) from failed ones (Couchsurfing). Six mechanisms work. Reviews need texture: Shef added text and dish photos to cut decision paralysis; Good Dog replaced ratings with "Verified Owner Stories"; Peerspace forced mutual reviews before either party could transact again. Supply verification: GOAT screens sellers via social history, ML image authentication, and physical inspection; Thumbtack and Rover run background checks and badge verified providers. Social proof: Udemy made a handful of instructors earn $1M+ and publicized those outcomes to recruit more, accepting the dream-vs-reality cost. Perceived quality is independent of actual quality: Airbnb's free professional photography shifted conversion; Udemy priced courses at $199 then discounted to $10, using price as a trust signal. Safety nets remove worst-case fear: Airbnb's $1M Host Guarantee; Peerspace's custom insurance for liabilities traditional policies excluded. Delivering the core promise matters most — Thumbtack's trust came from liquidity producing real hires; Lyft's from making unknowables (driver identity, ETA, rating, cost) visible before the ride.

How to kickstart and scale a consumer business—Step 4: Find your early adopters by doing things that don't scale

Getting your first 1,000 users comes down to seven tactics — almost every consumer startup relied on just one or two of them.

The most common tactic (50%+ of startups) is going where your audience already gathers. Netflix seeded DVD forums before launch. Dropbox posted on Hacker News. Discord entered a Final Fantasy subreddit. Tinder's Alexa Mateen pitched USC sororities in person, then built a 15-university ambassador program. Etsy sent a team to craft fairs every weekend across the US and Canada.

For marketplaces, cold outreach to bootstrap supply dominates. DoorDash's Tony Xu went door-to-door to restaurants. Thumbtack emailed local service professionals with job-lead offers. Cameo used $10/month interns to DM celebrities. Behance emailed 100 admired designers and built their portfolios for free, solving the chicken-and-egg problem in one move.

Friends and colleagues account for 20% of cases — Yelp drew from PayPal alumni, LinkedIn from its founding team's contacts, Strava reached 1,000 from 100 direct friends via word of mouth in a year.

Influencers are underused but high-leverage. Instagram gave early access to designers with niche Twitter followings. Pinterest turned around via a "Pin It Forward" blogging chain. Spotify seeded journalists, musicians, and tech founders before launch.

Press works when there's a genuine story. Airbnb pitched small bloggers at the DNC first, then watched coverage escalate to NBC and CBS. Udemy's first 10,000 users came entirely from its press launch.

Viral content is rare but decisive — Duolingo's Luis von Ahn TED talk drove 300,000 beta signups; Calm's "Do Nothing for 2 Minutes" site collected 100,000 emails in two weeks.

Physical placement serves local products: DoorDash printed restaurant-branded flyers; Instacart used branded bags.

These tactics form concentric circles from the founder outward — friends are easiest, viral content hardest. Start close and migrate outward only when nothing sticks.

Developing a growth model + marketplace growth strategy | Dan Hockenmaier (Faire, Thumbtack, Reforge)

Marketplace businesses invert the usual growth curve: later cohorts have lower CAC and higher LTV because supply liquidity keeps improving, the opposite of SaaS or e-commerce.

Growth models are spreadsheets for opportunity assessment, not forecasting. For SaaS, three building blocks: acquisition, retention, monetization. For marketplaces, add supply-side dynamics and the hard-to-model supply-demand interaction. The most surprising output: growth is far more sensitive to retention than intuition suggests — retained customers refer, generate content, and fund paid acquisition. Half the model's value comes from the act of building it.

The four metrics that matter most: GMV (both sides transacting), unit economics (most early marketplaces lose per transaction — Instacart, Uber), liquidity (can buyers reliably get what they need? — Uber's magic threshold is a 4–5 minute wait), and share of wallet (depth over breadth: 10% more wallet share beats 10% new customers because it predicts retention and reduces multi-tenanting).

Demand is the only thing that ultimately matters — suppliers always say yes when you bring them customers. Acquire supply only insofar as it improves the demand experience; past the liquidity threshold, more supply adds nothing.

Vertical unbundling works only with high order value (Airbnb from Craigslist) or a self-contained network (blue-collar Workrise). The horizontal platform always wins on LTV through cross-sell; Thumbtack tracked hundreds of vertical competitors and outbid them all on SEM because cross-category upsell raised LTV.

Marketplace evolution tracks a rising commission curve: lead aggregators (Zillow, ~5%) → trust-generating platforms (Airbnb, Etsy) → logistics and underwriting owners (DoorDash, Faire). The endpoint is Opendoor: own the supply, exit the marketplace model. Marketplaces requiring supplier creativity (Etsy, Steam, Faire) stay marketplace-shaped; commodity experiences consolidate.

Running a marketplace is gardening, not construction — an action today can surface effects two months later. Tread lightly with core incentives.

LinkedIn’s product evolution and the art of building complex systems | Hari Srinivasan (LinkedIn)

LinkedIn's feed improved not by chasing engagement but by sharpening focus on two signals: in-network relationship content (which carries opportunity through people you know) and out-of-network knowledge and advice (what members consistently say they want). The algorithm was tuned toward those axes, not generic time-on-site.

The most consequential structural change came during COVID: skills-based hiring. When hospitality workers were laid off in 2020 while customer service roles went unfilled, the root cause was that recruiters searched by job title rather than capability. LinkedIn began translating profiles into skill graphs and surfacing candidates by what they can do. By the time of this conversation, 47% of recruiters explicitly filter by skills — a hospitality worker with 70% of the skills for customer service now surfaces in those searches.

Open to Work followed a similar arc: from a private recruiter-only signal, to the public green frame launched during COVID when unemployment stigma collapsed, to a newer internal-mobility variant (Open to Internal Work) still facing cultural friction inside companies.

Talent Solutions — hiring plus LinkedIn Learning — is LinkedIn's largest business. The two sides are framed as separate marketplaces (seekers-to-recruiters, learners-to-instructors) but deliberately interlinked: feed content can surface job suggestions, skill gaps can trigger course recommendations.

Running this complexity requires explicit decision infrastructure. RAPID assigns a single named decision-owner to every cross-team call. The five-day alignment rule escalates any unresolved blocker automatically. PM hiring at LinkedIn explicitly tests for systems thinking — modeling cause-and-effect across interconnected surfaces is treated as a rare capability distinct from optimization skill.

For job seekers, the practical advice: signal interest on company pages before roles open, populate skills sections with concrete evidence, and lean on industry-specific knowledge to differentiate — PM roles were running roughly 50% below year-prior hire rates at the time.

Inside Etsy’s product, growth, and marketplace evolution | Tim Holley (VP of Product)

Etsy's marketplace flywheel runs on a simple insight: GMS (gross merchandise sales) is the north star because it only moves when a buyer buys from a seller — almost every product decision maps back to facilitating that transaction, not serving one side in isolation.

When Josh Silverman took over in 2017, the most important shift was installing GMS as an unambiguous drumbeat and adding competitive benchmarking. A consensus-driven culture had produced thoughtful products slowly; the fix was KPI clarity plus a narrative repeated until teams internalized it. Etsy Studio, a craft-supply marketplace bridging Pinterest inspiration with materials, was shut down the same year when it couldn't justify resources against the refocused goal.

Covid's mask surge was Etsy's stress test. Overnight traffic hit Black Friday levels when the CDC mandated masks; sellers who made wedding dresses pivoted to fabric masks within days. Etsy called sellers to gauge capacity — unprecedented. Retaining new buyers pushed the team into 30–90 day retention cohorts rather than single-session A/B tests.

On buyer conversion, the biggest wins came from behavioral nudges: buyer review photos showing items in real hands and homes, scarcity signals ("only one left"), and a single cart line — "Etsy offsets carbon emissions from every delivery" — that produced outsized lift. About 80% of experiments fail, consistent with industry norms.

Supply quality is maintained through strict policy enforcement and a "production partners" framework: sellers can use manufacturers they have a direct relationship with, but anonymous drop-shipping fails the test. At 100 million items the challenge isn't volume — it's helping buyers choose within a search result.

Product teams run as five-legged squads: PM, engineering, design, research/analytics, and marketing. PMs aren't mini-CEOs — they choose among ideas generated collaboratively — but make the call when data is ambiguous and own the consequences.

Marketplace lessons from Uber, Airbnb, Bumble, and more | Ramesh Johari (Stanford professor, startup advisor)

Marketplaces sell the removal of friction — transaction costs — not goods. Hosts are Airbnb's customers; drivers are Uber's customers. Both sides depend on the platform equally.

No marketplace starts as a marketplace. Before scaled liquidity exists on both sides, the platform has nothing real to offer. The right early move is solving a narrower problem that doesn't need scale. UrbanSitter started with credit card payments for babysitters. oDesk started with screenshot-based proof of work to resolve remote-trust friction. Only after those hooks built liquidity did sophisticated matching become viable. The test: scaled liquidity on both sides? If only one, deepen it or use it to attract the other.

Early monetization choices constrain future options. oDesk's flat percentage-of-transaction fee worked when trust infrastructure was the core value; it became a disintermediation trap once long-term worker–employer relationships no longer needed the platform. Upwork had to rethink pricing post-merger.

Data science in marketplaces runs three stages: finding matches, making matches, learning from matches. The critical discipline is distinguishing causation from correlation. Predicting which freelancer historically got hired is not the same as ranking applicants to improve future match quality — the latter requires causal inference.

Measuring experiments by wins versus losses drives incremental, long-running bets. Learning has a real dollar cost — every control group forgoes revenue — but teams rarely account for it. Bayesian A/B testing lets failed experiments update priors and generate future value. Marketplace experiments also produce whac-a-mole effects: fixing new-supply experience degrades existing-supply experience and back, because the game is reallocating attention, not expanding the pie symmetrically.

Rating systems inflate over time via reciprocity and norming. Relative comparison questions ("did this exceed expectations?") outperform raw star scales. Averaging punishes new entrants — an early negative eBay rating cut revenue 8% and predicted platform exit. Prior-based smoothing corrects this.

The hierarchy of engagement | Sarah Tavel (Benchmark, Greylock, Pinterest)

Consumer products and marketplaces fail the same way: optimizing for vanity metrics — MAUs, GMV, downloads — rather than signals that predict whether users stay.

For consumer products, the Hierarchy of Engagement has three levels. Level one: the single "core action" that signals comprehension and drives return visits. Pinterest's was pinning, not following or clicking. YouTube's was subscribing, not watching — subscriptions lock in creators and viewers in ways views don't. NUX should funnel to that action. Level two: make the product get better with use and costly to leave. Pinterest's feed personalized from accumulated pins; Evernote held years of notes no one wanted to rebuild. Level three: self-perpetuation through loops. Pinterest's "Sarah pinned your pin" notification re-engaged dormant users effortlessly. Evernote had no such loop, relied on paid acquisition, and tapped out. Clubhouse had the mechanics but push-notification overload broke the flywheel as it scaled. Pure anonymity blocks level two — no persistent identity, no accruing benefits or mounting loss.

For marketplaces the hierarchy is focus, tip, dominate. Level one means constraining to a "thimble" and pursuing "happy GMV" over total GMV. DoorDash targeted suburbs where Grubhub was absent; Postmates spread across cities and categories. Level two is tipping: reach saturation until supply comes to you. REKKI enrolled London restaurants; their suppliers then sent CSVs of their own customers to onboard. Not all markets tip — concentrated supply and homogeneous supply (Mechanical Turk: more workers add nothing for buyers) block it. Level three is dominance. Grubhub data shows winner-take-most margins vastly exceed close two-player races. Etsy's warning: allowing mass-produced goods to chase GMV broke trust with handmade sellers and buyers and preceded a CEO change.

Market size matters less than current — structural momentum that pulls the company forward. Underestimated markets have no entrenched competitors and tip faster.

How to describe your business as an equation

You don't fully understand your business until you can express it as an equation that maps inputs to revenue and exposes where leverage actually sits. Eight tech models each have a distinct form. Bottom-up seat-based B2B SaaS (Figma, Slack): ARR = Visitors × trial conversion × paid conversion × seats × price, plus Expansion ARR from seat growth and upsells, minus churn. Usage-based SaaS (Datadog, AWS) swaps seats for consumption volume. Top-down sales-led SaaS (Salesforce, Snowflake): Leads × qualification rate × win rate × ACV. B2C subscriptions (Duolingo, Spotify): Traffic × trial conversion × paid conversion × average monthly revenue. Ad-supported B2C (TikTok): Active users × Impressions per user × CPM. Marketplaces: Transactions × AOV × Take rate. DTC: Transactions × AOV. For cost-heavy models, append contribution margin % — that drives valuation. Payback period (CAC ÷ annual contribution margin per customer) sits outside these equations but is the key acquisition-efficiency metric.

Business strategy with Hamilton Helmer (author of 7 Powers)

Sustainable competitive advantage requires both a benefit (lower cost or higher price) and a barrier that prevents replication. Operational excellence, fast-moving teams, and great UIs all fail this test — they must keep running but confer no durable edge. When Blockbuster built a DVD-by-mail site, it looked identical to Netflix's: if a competitor can copy the interface or hire a consulting firm to match your process, it isn't power.

For early-stage startups, three of the seven powers can be set aside: branding and process power only emerge in the stability phase, and resource power (patents protected by law) applies to a narrow class like pharmaceuticals. That leaves four, pursued in sequence: counter-positioning, scale economies, switching costs, and network economies.

Counter-positioning comes first: product-market fit usually means substituting for an incumbent (Amazon vs. brick-and-mortar, Google vs. Yahoo), and without it an incumbent simply extends their product and crushes you. Once scale matters, the other three become accessible.

Network effects are almost universally overstated. Helmer distinguishes network effects (a real but modest linkage) from network economies (a linkage material enough to produce permanently better margins). Uber and Lyft both have network effects; neither has network economies — the advantage isn't large enough to stop continuous competitive spending. Uber's durable edge is geographically scoped scale economy — why the China expansion was incoherent and Uber Eats leveraged the same local driver supply.

On AI: no eighth power has emerged. AI most likely plays the role electricity played for factories — broadly applicable technology that redesigns existing functions rather than creating a new category. The winners will be tertiary adopters redesigning accounting, HR, and R&D around it, not the model providers.

Business value reduces to three exhaustive factors: power, market size, and operational excellence. Everything else is a subcategory of one of these three.

How marketplaces win: Liquidity, growth levers, quality, and more | Benjamin Lauzier (Lyft, Thumbtack, Reforge)

Liquidity — the fraction of intentful demand that transacts — is the central health metric of any marketplace, and most teams fail to define it or build an actionable playbook around it. Benjamin Lauzier (VP Product at Thumbtack, driver-side lead at Lyft as employee #30) argues that pre-PMF founders should ignore marketplace dynamics entirely: nail one side, hack the other with existing supply pools. Thumbtack scraped Craigslist for contractors. Supply is the hard side 80–90% of the time.

Once past PMF, use a threshold-based market-health proxy over raw fill rate. At Lyft, the predictor was ETA: once the nearest driver was under two minutes away, conversion plateaued. Teams could then measure whether adding 100 drivers moved ETAs rather than noisy fill rate.

Marketplaces fail three ways: never building local density to kick the flywheel; neglecting supply's product needs while attrition mounts; and letting quality erode by lowering the supply bar. Over-filtering silently fragments supply — Thumbtack's "smoke machine" checkbox excluded 95% of DJs without users realizing it was a hard filter. Ranking adjustment fixed it without removing the preference.

Facing an Uber 30× its size, Lyft paid top-rated drivers $35 per mentor session to inspect and onboard new applicants. Peer credibility outperformed marketing: "drive Tuesday at 2 PM, text me" converted reliably. A "recruiter" role let idle drivers claim mid-funnel drop-offs for $20 per activation. Within 18 months Lyft matched Uber's footprint with a tenth of the resources.

Lyft's ground-loss traced to scope: when COVID eliminated rides, Uber's Eats and parcels business provided a survival floor Lyft lacked.

Supply control reliably backfires. When Thumbtack moved from leads to direct bookings — improving pro earnings ~20% — pros revolted because they valued the feel of the sale. Coaching tools and star-rating thresholds outperform top-down control; overt control also invites employment-classification liability.

Pricing your SaaS product

SaaS pricing is not primarily about the number — the most important decision is what you charge *for*. A flat monthly fee leaves revenue under the demand curve; a value metric gives infinite price points scaling with customer size. Companies using value metrics grow at double the rate with half the churn of flat-fee competitors.

Finding the right metric means identifying the ideal essence of value, then testing proxies customers trust. HubSpot charges on contacts rather than seats — unlimited users remove adoption friction while contact volume scales naturally. ProfitWell Retain charges on churn recovered. A good proxy must correlate with usage growth and support retention.

The second foundation is quantified customer segments: profiles by role and company size, with willingness to pay, LTV, CAC, and valued features. Patrick Campbell learned this when ProfitWell Metrics research revealed analytics products have terrible willingness to pay and retention — 18 months ahead of competitors, they shifted to freemium.

Work through value metric and segments before touching price points or freemium. Freemium is an acquisition model, not a pricing tier. Rapid-fire: discounts above 20% predict higher churn; case studies lift willingness to pay 10–15%; design adds ~20%; integrations improve both. Experiment with monetization every quarter.

Types of business models

Eight fundamental models cover nearly every business: sell a thing, rent a thing, take a cut (marketplaces like Stripe), subscription, usage-based (Twilio, AWS), sell a service, advertising, and percentage of assets. The real strategic lever is combining or replacing models — Peloton stacked subscription onto hardware, Uber swapped taxi-medallion rental for take rate, Robinhood replaced per-trade fees with selling order-flow data, Chime moved from penalty fees to Visa interchange. Every existing model has a plausible pivot path worth pressure-testing.

60 ideas to boost your growth

One-off growth events — "turbo boosts" in the Racecar Growth Framework — are neither scalable nor a signal of product-market fit, but nearly every successful company has used them to kick-start a growth engine. Seven types recur across the examples: viral video (Dropbox's YC demo on Hacker News, GoldieBlox's 3M-view launch); mini-product drops (Calm's precursor donothingfor2minutes.com pulled 2M visits and 100K emails in 10 days; Codecademy's Code Year challenge got 400K sign-ups); limited-time offers (Cards Against Humanity selling nothing for $5 on Black Friday); influencer partnerships (Spanx via Oprah's gift basket to her hairstylist; Reddit's first 1K users from a Paul Graham blog post); co-marketing; offline experiences (Hinge's DC launch party, Snapchat's mystery Spectacles vending machines); and picking fights (WePay dropping a block of ice at a PayPal conference lifted sign-ups 300%).

Prioritizing at startups

Pre-PMF B2B startups have one goal: make 10 customers very happy. Standard frameworks don't apply — you lack the data A/B testing requires (a 10% conversion lift needs 5,000+ users). Four traps: building vitamins not painkillers, chasing data over decisions, over-strategizing without shipping, and implementing feature requests instead of solving root pain. The SUSS framework — Segment narrowly, Understand pain through 100+ interviews, Solve it completely, Stay focused — directs all effort. Time allocation: ~80% on product value (blockers, retention features, differentiators), ~10% onboarding, ~10% delight. Best ideas come from founders building what they themselves need, then from conversations surfacing pain, excitement signals, and competitor-switching moments.

Differentiating your product

Michael Porter found only two paths to winning a market: operational effectiveness (doing the same things better) or differentiation (doing different things). Operational effectiveness alone fails long-term because competitors copy it and margins erode. Seven differentiation levers exist: cheapest price (Robinhood, GEICO), highest quality (Apple, Whole Foods), most convenient (Gopuff's 10-minute grocery), safest/most trusted (DuckDuckGo, Volvo), proprietary supply (Netflix exclusives, Airbnb homes), identity/mission appeal (Patagonia, Nike), and niche underserved markets (DoorDash in Tier-2 cities, Chime targeting $35K–$70K earners). Combining multiple levers compounds advantage — Apple stacks quality plus safety, Figma stacks quality plus convenience.

The art of building legendary brands | Arielle Jackson (Google, Square, Marketer in Residence at First Round Capital)

Brand is who people think you are — not your logo or color palette. Arielle Jackson (Gmail at Google, Square Stand launch) uses a three-part framework across 100+ early-stage companies: purpose, positioning, and personality.

Purpose is one sentence stating why you exist, independent of revenue. Google's "organize the world's information and make it universally accessible" is the benchmark — 100,000 employees could still recite it. Stripe's: "increase the GDP of the internet." Draft by listing cultural tensions in your space, then completing "we exist to…" It should serve as a conference intro and your about-page header.

Positioning is the space you occupy in a target customer's mind. If 10 employees give 10 different answers about what the company does, that's a positioning problem. The process runs in concentric circles: TAM → target audience (who you actively acquire over 18 months) → model persona — Eero's was a tech-savvy dad, suburban St. Louis, VP of sales, gaming teenage kids. Output: for [audience] who [need], [product] is a [category] that [benefit], unlike [alternative]. The benefit line is what a user would say to a friend — "turn your iPad into a point of sale" for Square Stand. The bar test: roleplay as your target user recommending the product; if it sounds natural, it passes. Jargon ("leverages," "empowers") fails.

Personality maps onto Jennifer Aaker's five dimensions: sincerity, excitement, competence, sophistication, ruggedness. Strong brands spike in two. Write five attributes as "we are X but not Y" — "playful but not silly" — tension that makes personality specific rather than synonyms.

On naming: suggestive names (Seesaw, Maven) do marketing work themselves; empty vessel names (Yahoo, Eero) require more spend. Neither kills a good company — Disney was just a last name. Seven criteria apply: trademark, domain, distinctiveness, timelessness, message reflection, pronunciation ease, visual appearance.

How to build a powerful marketing machine | Emily Kramer (Asana, Carta, MKT1)

Marketing breaks into two problems — fuel (content, messaging, positioning, copy) and engine (distribution channels, email infrastructure, paid, marketing ops) — and most startups build the wrong one first. SDR-driven companies run outbound before they have anything valuable to say; PLG companies generate traffic that bleeds through a broken funnel because no one built lifecycle email and conversion infrastructure.

The first marketing hire should be diagnosed against that split. Three archetypes exist: product marketer (understands product, audience, and market; owns positioning and copy; the default right answer for most B2B startups), content/community marketer (fuel-heavy; long-form and audience building), and growth/demand gen marketer (engine-heavy; channels, paid, ops). The ideal hire is "π-shaped" — expert in one of those three, proficient in a second, able to set strategy and hire contractors across all. Content + growth is the hardest π to find (opposite sides of the brain); product marketing paired with either of the others is more realistic.

Business model matters more than industry experience. A marketer who sold to HR teams at Salesforce has no muscle memory for building from scratch. Hire someone who has worked somewhere early enough to have set strategy themselves, and who has seen what high-quality marketing looks like.

On product-marketing collaboration, Asana's AOR list — a who-owns-what map beneath job titles — meant product always knew exactly which marketer to involve. The GACCS brief (Goals, Audience, Creative angle, Channels, Stakeholders) shared before any initiative launches replaces the back-and-forth that results from handing marketing a launch at the last minute.

Good marketing teams articulate their core steady-state work, their big bets (step-change growth candidates), and the broken foundations slowing them down. Red flag: activity goals ("publish 10 blog posts"). Quality signal: tracking conversion rates at every funnel stage, not just volume at the stage marketing owns.

How to get your marketing team to drive more impact

Most marketing busywork fails for predictable reasons: no clear goal, too broad an audience, undifferentiated creative, no distribution plan, or stakeholders excluded until too late. The fix is a one-page brief called GACCS (Goals, Audience, Creative, Channels, Stakeholders) written before work starts.

Goals link the project to a specific OKR ("increase free-to-paid conversion from x% to y%"). Audience means a precise persona, not "users." Creative requires a unique point of view — if nothing new is being said, don't make it. Channels forces a distribution plan upfront, plus "mileage" — repurposing one asset across formats. Stakeholders names a single DRI plus reviewers, shared before execution so feedback can still redirect the work.

If GACCS is hard to write, the foundation is missing: no shared OKRs, ICPs, or channel inventory. Fix those first. A companion AORs spreadsheet (one owner per area, no joint ownership) eliminates the "who do I Slack?" friction that slows cross-team work.

EOY Review

April Dunford: B2B companies lose 40% of deals to "no decision" — spreadsheets and interns — so competitive alternatives must include what buyers currently do, not just named rivals. Differentiated value emerges by mapping capabilities to outcomes, then finding who cares most.

Crystal Widjaja: measurements are not insights. "Power users book four times more" is an observation; an insight adds context that changes behavior — power users convert on free-shipping for high-GMV baskets, non-power users don't. Treating analytics as entertainment — interesting but never acted on — is the core failure.

Shishir Mehrotra's eigenquestions technique: force two questions about any problem and the structuring variables surface immediately. His PSHE framework (Problem, Solution, How, Execution) maps seniority — juniors execute handed plans; senior leaders find the right problems. The "trough of dissolution" is where scope stops predicting advancement and PSHE level starts.

Kristen Berman's three Bs: define a hyper-specific behavior ("2 workouts with 2 instructors within 7 days"), remove logistical and cognitive barriers (status quo bias, uncertainty aversion), then amplify immediate benefits — completion bias, social signals — because present bias means future benefits don't drive action.

Elena Verna: retention and habit loops must precede product-led acquisition; PLG only works when users have a one-to-many relationship. Make free whatever drives virality, habit loops, or the aha moment; gate only what would block the growth model.

Shreyas Doshi's LNO framework: tasks are leverage (10–100X return), neutral, or overhead. Perfectionists treat all three identically. Apply perfectionism to L tasks only; compress N and O deliberately.

Matt Mochary: companies that ran 2020 layoffs reported better absolute output within 60 days — more features, higher NPS — because each added person creates geometric coordination overhead. For hard conversations, warn first to prevent amygdala hijack, then name emotions and invite the person to speak before resolving.

How to price your product | Naomi Ionita (Menlo Ventures)

Most B2B startups make three pricing mistakes: monetizing too late, undercharging, and treating pricing as a one-time decision. Willingness to pay is proof of product-market fit — delaying it cheapens the product and kills the feedback loop. Freemium works, but paywall placement matters: anything needed to reach the "aha moment" belongs free; features whose value requires accumulated data or scale go in paid tiers ("day one vs. day 100"). At Invoice2go, restructuring tiers this way doubled the upgrade rate while raising the pro plan price 30%.

Evernote's $45/year flat price illustrates underpricing at scale. Surveys found "guilt" was the top upgrade reason — meaning the free tier was too generous. Avid users reported hundreds of dollars of value; casual users compared it to a pre-installed notepad. The fix is segmentation: different plans priced to each persona's actual willingness to pay.

To set initial prices: form a cross-functional committee, rank features as must-have / nice-to-have / unnecessary (or use a 100-point allocation), then apply Van Westendorp's four questions — price that implies poor quality, good deal, expensive-but-acceptable, and prohibitive. Plotting those curves identifies the viable range. Revisit every 6–12 months at product launches, not once and forgotten.

The returns are large. OpenView found roughly half of companies making a pricing change saw 25%+ ARR lift. ProfitWell's 500-company study showed a 1% monetization improvement delivers 4× the bottom-line impact of a 1% acquisition improvement. Envoy's founder 10×-ed his price mid-call; the prospect agreed without hesitation. Losing 20–30% of deals to price is healthy — losing none means the ceiling hasn't been found.

For PLG companies with a multiplayer expansion path (Figma is the model), growth over near-term revenue is rational. Pure usage-based pricing suits fewer than 10% of SaaS — most buyers want predictability, making hybrid tiers with quota-based escalators the practical default.

How to achieve hypergrowth in your business and career | Carilu Dietrich (Atlassian, Miro, Segment, 1Password)

Hypergrowth requires three things: a product people love enough to spread without being pushed, built-in viral mechanics (Miro's whiteboards, Atlassian's Confluence pages that pull in collaborators), and an org that can "ride the lightning" — companies compress a decade of growth stages into two or three years, so you must continuously hire 2–3x leaders who have already lived the next stage. ChatGPT is the extreme case; you cannot buy that velocity with paid channels.

Atlassian's model: redirect the sales budget into R&D, spend two to three times more on engineering than peers, run almost no outbound prospecting, use SEO/SEM as primary demand. For PLG, bundling kills velocity — Atlassian tested Jira + Confluence + HipChat as a landing bundle and it slowed deals; single-product, fast time-to-value wins. Sales should engage only after a usage threshold: Airtable and Miro wait for 20–40 paying users at an account before a rep touches it.

Big growth levers — moving upmarket, going global, adding a partner channel — fail when treated as marketing problems rather than company strategy. Product, customer success, and sales need aligned OKRs or they pull in opposite directions.

Career compounding follows the same logic. Dietrich's team grew 5→7 at Oracle in five years and 15→100 at Atlassian in four; company momentum is your momentum. She accelerated by riding industry waves (nonprofit → tech → B2B SaaS → dev tools → AI) and volunteering to cover departing department heads, gaining cross-functional depth that paid off for decades.

To evaluate a company: Rule of 40, net dollar retention (Snowflake at ~165–180%, nearly doubling revenue from existing customers), NPS, investor tier, market position, Glassdoor. Favorite hiring question for managers: "How many people have you fired, walk me through each" — reveals tenure, hard-cycle experience, and whether they hold compassion and accountability together.

How Notion builds product

Notion ran fewer than 15 PMs across 550 people and didn't hire its first PM until 50–60 engineers — a gap CPTO Michael Manapat now considers a mistake. Unlike Stripe, where lines like Radar and Capital are independent, every part of Notion is coupled: changing databases affects wikis, project management, and docs simultaneously, requiring central coordination most companies this size don't need.

Planning runs in half-year cycles with aligned two-week sprints company-wide, though teams vary in how strictly they use them. OKRs exist at company level; below that, much work is zero-to-one enough that key results reduce to "ship this."

Product reviews follow four check-ins: problem statement, directions, full solution, and ship candidate. The directions step shifted from async email threads to synchronous Figma walkthroughs because documenting exploration in writing consumed too much time. Co-founder feedback from Ivan Zhao and Simon Last now arrives at these fixed gates, not two days before launch.

Engineering, PM, design, data, UXR, and security all report to Manapat. The unified org surfaces crosstalk siloed functions miss. Hiring prioritizes joint EM–PM accountability — "end-to-end ownership" opposed to both Microsoft's waterfall (PMs hand tablets to engineering) and siloed accountability.

The product philosophy is "peeling an onion": flexible primitives power users can reconfigure, assembled into opinionated entry points for users who just want something that works.

M&A, competition, pricing, and investing | Julia Schottenstein (dbt Labs)

For any startup, there are only two or three acquirers that find what you're building genuinely strategic — so M&A is always about creating plan Bs, and the time to start is before you need one. Julia Schottenstein, dbt Labs' product lead who ran the acquisition of Transform, argues the playbook is to identify those likely buyers, inflict competitive pain in their strategic area, and stay friendly. Cutting off potential acquirers early kills the optionality founders need if the independent path fails.

The Transform acquisition illustrates the mechanic. Transform had built a strong semantic layer (defining business metrics so any query returns consistent data) but lacked distribution. dbt had distribution and community but was behind on the product. Transform made noise about solving hard technical problems, positioned as a friendly partner, and made it impossible for dbt Labs to ignore them — which made post-acquisition integration easier because the bridging work was already done.

dbt's success came from power through simplicity — inviting SQL-fluent analysts into workflows previously reserved for data engineers — combined with open source that reduced friction and created a flywheel: easy adoption, word-of-mouth, 20,000 weekly companies. Timing mattered: Snowflake going from $4B to $12B in 2019 created the chaos dbt brought structure to. The founders spent two years as Fishtown Analytics, a consulting firm, building dbt against live client pain.

On pricing: customers value dbt at 20–35% of their cloud data warehouse spend, yet dbt charges a small fraction of that by design. The first pricing change taught them to use relative-value framing in conversations and track conversion and churn.

On product philosophy: "worse is better" and "tech debt is a champagne problem" — ship the naive for-loop scheduler, get it in users' hands, scale when you have scale problems. Single-thread the team on one mission.

Velocity over everything: How Ramp became the fastest-growing SaaS startup of all time | Geoff Charles (VP of Product)

Ramp reached $100M ARR in under two years with fewer than 50 people in R&D by treating velocity as the primary organizing principle — not a byproduct of good work, but the explicit criterion for hiring, promotion, team structure, and decision-making.

The mechanism is small, single-threaded teams given one goal, tight timelines, and insulation from everything else. The accounts payable product (a Bill.com competitor) was built by three engineers, one designer, and one PM in three months and now processes billions annually. Protecting focus requires structural layers: a rotational production-engineering program absorbs bug escalations; product operators handle documentation and release so PMs are never pulled away.

Empowerment runs on "context over control." Alignment happens at the goal and hypothesis level, not the solution level. The PM's contract with a team is the strategy doc and roadmap; everything downstream is the team's call. Status updates are async — Charles has never scheduled a status meeting. Quality is held through hard controls: negative reviews route monthly to the responsible tech lead, PM, and designer; tickets-per-user gates new feature shipping if elevated.

Planning shrank from quarterly OKR cycles (consuming a full month in three) to a biannual one-pager. Strategy docs follow a fixed structure: goal, hypothesis, data, right-to-win rationale, metrics, initiatives. The right-to-win test is load-bearing — Ramp entered bill payments because money movement, liability processing, and accounting integrations already existed from the corporate card.

Support reports into product because every ticket is a product failure. Agents are incentivized to reduce volume, not close tickets: 400,000 users, under 30 agents.

Hiring weights hunger (leaving because things got too slow is the strongest signal) and depth of reasoning over domain experience. The clearest A-plus engineer marker: they set the pace, push back on specs, jump into customer channels unprompted, and own their product's quality without being managed toward it.

How the most successful B2B startups came up with their original idea

Only about 40% of successful B2B founders came up with their idea by personally living the pain. Interviews with founders of Gong, Figma, Amplitude, Segment, Slack, Retool, Linear, and Databricks reveal three reliable paths. A fundable idea needs three properties: a path to $100M+ revenue, existing solutions at least 4 points worse than yours on a 10-point scale (Kunal Shah's Delta-4 rule), and genuine excitement — Dylan Field's warning: "If you're three to four years in on an idea that you hate, you're going to burn out."

Past pain (~40%): Gong's co-founder built a system to surface what lived in salespeople's heads after CRM showed outcomes but never why deals failed. Retool's David Hsu rebuilt the same internal tools at every job until the building blocks became obvious. Linear emerged when founders at Coinbase, Airbnb, and Uber each hit the Jira ceiling and found no alternative.

Ponder and probe: Figma began when Dylan Field spotted WebGL and narrowed from drones to design after a Flipboard internship. Vanta's Christina Cacioppo restricted herself to security and collaboration, found every startup founder felt guilty about ignoring security, and started doing it for them manually. Notion drifted as a no-code builder for four years before users kept returning to its collaborative editor, revealing the docs/wiki wedge in 2017. Zip was the founders' sixth or seventh idea; a YC partner told them to find large markets with entrenched bad incumbents, pointing to procurement (now valued over $1B).

Present pull: Amplitude pivoted from a voice-texting app when YC peers demanded access to their internal analytics tool. Segment launched Analytics.js as a throwaway growth hack, watched it top Hacker News, and realized that was the product. Slack's IRC layer, built to patch gaps in the failed game Glitch, had become more relied upon than email before the pivot.

In all three paths, the signal is the same: pain reveals the opportunity matters, pull reveals you're solving it.

How to validate your B2B startup idea

Roughly 40% of successful B2B startups pivoted at least once — Retool started as UK Venmo, Amplitude as a voice texting app, Notion as a no-code website builder, Slack as a game. Distinguishing a real idea from "total crap" turns on four signals: people pay money (ideally strangers with no obligation), continued usage of a hacky prototype, strong emotion, and cold inbound interest.

Four validation paths map to different situations. The do-it-manually path works when you're unsure if the problem is real: Vanta's Christina Cacioppo manually wrote SOC 2 compliance reports for Segment and Front, then got an unsolicited inbound asking for the same service. Ramp's Eric Glyman produced "savings reports" by hand from 90 days of credit card data. The listening path is the default — talk to roughly 30 potential customers before building. Zip interviewed 75 CFOs in two weeks against 16 explicit criteria. Gusto's Tomer London found the threshold test was emotional intensity: polite interest means no, spontaneous cursing means yes. Amplitude stopped at 30 conversations, moved to building, and converted zero to paying customers. The prototype path suits founders with domain expertise: Gong ran 12 design partners through a beta; 11 bought when it ended. Hex used two. Across all paths, cold outbound is the purest signal — Retool's David Hsu and Zip's founders both avoided selling to friends because warmth contaminates the read. The fourth path is simply launching: Segment and Loom found users are better at discovering products than founders are at finding design partners.

Every prosumer collaboration product — Notion, Figma, Coda, Airtable — spent three to four years without visible traction. Notion ran out of money; Ivan Zhao borrowed from his mother and rebuilt from Kyoto. Median conversations before commitment: 30.

How to find and win your first 10 B2B customers

Early B2B customers are won through trust, not tactics — the sequence that works moves outward from highest to lowest trust, starting with the founder's personal network.

Personal networks outperform every other early channel because trust substitutes for proof. Gong's first dozen customers were all personal connections of founders or early employees. Census co-founder Boris Jabes called YC batchmates from a decade prior: "You can skip all the niceties of customer discovery." Gusto signed up fellow YC batchmates and a children's swimming camp. Coda's Shishir Mehrotra pushes back on avoiding friends: if they don't like the product they stop, so their feedback is honest.

Cold outbound works when surgical. Retool filtered Crunchbase by operational-heavy verticals — fintech and delivery rather than SaaS — then emailed the CTO and VP of Operations simultaneously, landing DoorDash, Rappi, and Brex. Figma's Dylan Field built a script ranking designers by Twitter influence and cold-DM'd down the list. Zip went entirely cold via LinkedIn to avoid polluting signal from friends buying out of obligation.

Investor networks pay off when mined directly. Vanta's Christina Cacioppo searched every YC Bookface post mentioning "compliance" over a decade and cold-emailed those founders. Amplitude avoided current YC batchmates ("not real companies") and reached ex-Zynga PMs through one investor introduction that then spread through a Facebook group. First Round Capital ran an SDR-equivalent program giving Sprig five targeted meetings per week.

Communities reward contribution before pitching. Snyk ran a freemium beta through dev meetups and open-source channels for nearly a year before any paid tier. Content drove waitlists for Front, Linear, and Hex. Press rarely converts — Slack got 8,000 day-one invite requests as an exception. Segment and Loom simply launched publicly.

None of these channels scale. That is precisely why they work at this stage.

A guide for finding product-market fit in B2B

B2B product-market fit is not a binary milestone but a continuous process of finding fit with larger and larger segments. Most founders never feel they've fully achieved it: Shishir Mehrotra ran YouTube at 100M DAUs without feeling PMF for their desired audience; Ali Ghodsi feared hitting a wall at $100M revenue; Rick Song frames each fundraise as demolishing the prior bar and raising the next.

Across 24 top B2B startups, the median time from idea to first feeling of PMF was two years; from working product to PMF, 9–18 months. Figma and Slack took 4+ years and were outliers.

The journey runs five steps. First, get one company to love the product. Figma's Dylan Field called a red alert over a single unpaying alpha user's bug. Retool live-paged the team on every in-app issue and shipped DoorDash's on-prem feature in 36 hours. Second, get paid five to six figures. Vanta's Christina Cacioppo closed $500k without sales experience — the problem sold itself. Third, expand to 3–10 paying customers. Gong converted 11 of 12 beta partners once they started charging; when one later turned it off, its CEO called demanding to know who authorized it. Zip and Persona both set the threshold at 10 live successful customers.

Fourth, notice the push-to-pull shift: inbound emails from strangers (Canva, Hex), prospects demanding features faster than you can ship (Segment, Figma). Census's Boris Jabes noticed his demeanor change socially before registering the sales shift. Fifth, sustain consistent growth. Databricks went $1M → $10M → $30M ARR before Ghodsi believed the business had wings.

If stuck: verify the problem is important and underserved and the solution meaningfully better, then reconsider the ICP before rethinking the idea. Amplitude spent a year building before talking to customers — the corrective is 50% of founding-team time on customers.

How to become a category pirate | Christopher Lochhead (author of Play Bigger, Niche Down, Category Pirates, more)

In technology markets, one company captures 76% of total category value — not market share, but market cap. Competing with a better product in an existing category means fighting for the remaining 24% by default. Category design is the alternative: define a new problem so compellingly that customers adopt your framing, making your solution the only logical answer.

The mechanism is problem obsession. Gojo Industries didn't improve bar soap — the founder reframed the problem as "how do I wash hands without water?" and invented Purell. Lomi named "smart home composting," framing food waste as environmental crisis and personal nuisance, turning months of outdoor composting into hours on a kitchen counter. Elisha Otis couldn't sell a "safety elevator" until he called it a "vertical railway," giving people mental scaffolding for why tall buildings were now possible.

This is "languaging" — strategic vocabulary that creates new thinking. Starbucks invented "venti" to charge $3 for a 10-cent cup. OpenAI coined "large language model" and owns the AI conceptual frame. Whoever names the category tends to own it.

Threads is the negative case. Despite Meta's brand and unmatched distribution, it cratered because it copied Twitter and called itself better. Red Bull Cola, Amazon Fire Phone, and Microsoft Stores failed identically. You cannot displace a known solution without reframing the problem.

Execution means one or two "lightning strikes" per year aimed at super consumers — the 8–10% of buyers who drive most profit and set industry norms — then word-of-mouth carries the category point of view. That POV must address the customer's problem, not the product's features: categories belong to customers, brands belong to companies.

"Product-market fit" inverts the logic. Categories make products, not the reverse. Design a market category for your product — don't fit your product into an existing one.

Hiring your early team

Engineers dominate early B2B hiring: over two-thirds of companies hired an engineer as employee #1, and 100% had at least one among the first three. The exceptions are instructive — Vanta hired a compliance expert because the founders could build but needed domain validation; Coda hired a recruiter who became COO; Snyk hired a PM because its founder saw developer UX as a product problem, not a technical one.

Customer success appears persistently in the first three hires, tied to getting one company to love your product. Over 40% had a designer co-founder or early hire. Subject-matter experts were common in regulated markets (Gusto, Zip, Vanta). By employee 10, sales is second, and in-house recruiters appear at quality-focused companies — Linear, Figma, Ramp, and Coda all built recruiting in-house to avoid agency incentives that favor speed over quality.

Finding people runs through four channels: former colleagues and friends (dominant), cold outreach via LinkedIn and GitHub, public job boards led by AngelList, and networks of existing hires. Closing requires a compelling vision, strong early team, 50–100 hours per hire, and "love bombing" — surprising the candidate at offer stage with the full interview panel.

Every founder did founder-led sales first. The handoff trigger is behavioral: calls feel scripted and boring, or inbound overwhelms. Most common crossover: $300–500k ARR. Retool stayed founder-led to $3M; Notion to $10–15M; Canva past $120M. Databricks tried full PLG in 2015, pulled reps from accounts, and revenue flatlined within two quarters.

The right first sales hire is a "hungry senior AE" — experienced enough to close autonomously, entrepreneurial enough to surface product insights rather than wait for a playbook. Senior enterprise veterans routinely struggle without structure. Both Gong and Amplitude used a part-time ex-VP of Sales as coach before the full-time hire, avoiding the IC-versus-VP dilemma.

How to get press for your product | Jason Feifer (editor in chief of Entrepreneur magazine)

Journalists serve their audience, not you — that's the premise behind all press strategy. Jason Feifer, editor-in-chief of Entrepreneur Magazine, structures the playbook in three steps: prep, find who to pitch, then pitch.

Prep. Know what you want press for — fundraising credibility, awareness, or partner discovery. Match publication to goal: a DC hot dog truck should target the Washingtonian food section, not Entrepreneur, whose readers can't buy local hot dogs. Learn a publication's mission by reading it. Entrepreneur teaches entrepreneurial thinking — a butter-dish maker's new hire is irrelevant, but her airport market-research trick (polling bored travelers instead of a $10,000 research firm) fits. Chase press only when you have a concrete reason; it's an unpredictable add-on, not a growth strategy.

Who to pitch. Don't email the editor-in-chief — they're not sourcing stories. Search the publication for your product category and find the writer on that beat. Prefer freelancers: they pitch editors to earn income; a targeted email reaches them at near-100% read rate versus a sliver for the EIC.

The pitch. Three paragraphs max, customized to show you've read their work. Open with the story mapped to the publication's mission, not your pitch. Canadian painter Meg O'Hara's cold email worked because she structured it as problem bullets and solution bullets, mirroring Jason's "Problem Solvers" format.

Two alternatives when direct coverage is hard: become a quotable expert on breaking news, or manufacture context. Fred Ruckel's cat toy wasn't press-worthy until he flagged Amazon-to-eBay arbitrage fraud as a small-business pattern — that became a 4,000-word feature. Zapier's annual fastest-growing-apps list earns coverage every year because Zapier creates the data.

Coverage may reach only 5–10k readers. Real value is often downstream: "As seen in" on your website, or a promoted post targeting the audience whose attention the coverage was for.

How to build a killer sales pitch

In B2B software, 40–60% of purchase processes end in "no decision" — not because buyers prefer the status quo, but because they cannot make a confident choice. Buyers fear a career-damaging mistake more than missing out. The cure is not a better feature demo; it is a pitch that teaches buyers how to decide.

April Dunford's two-part structure: a setup followed by a follow-through. The setup has three steps. First, a market insight — your distinct point of view on why your value matters (Help Scout's: "customer service is a growth driver, not a cost center"). Second, pros and cons of alternative approaches by category, not individual competitors — teaching buyers the whole market. Third, the perfect world: "Given those tradeoffs, what would an ideal solution look like — do you agree?" If yes, the sale reduces to proving you deliver it; if no, disqualify. The follow-through then demonstrates differentiated value — not a feature tour, but proof of what the prospect just agreed they need.

Postman illustrates it at scale: frame APIs as business risk, show how siloed toolchains cause it, define an API-first ideal, then prove Postman delivers it. The structure also aligns sales, marketing, and product on one consistent story.

What AI means for your product strategy | Paul Adams (CPO of Intercom)

AI is a meteor, not a hype cycle — go back to first principles about what your product does and ask whether AI can do it. Paul Adams, CPO of Intercom for 10+ years, argues that most B2B SaaS products with workflows or multimedia are directly in the disruption path. When ChatGPT launched in November 2022, Intercom scrapped its roadmap and rebuilt. Their AI-first chatbot Fin now resolves 50–70% of inbound queries for top customers without a human. The bigger challenge is organizational: managers struggle to conceptualize what their teams do when a bot is the first line of defense. Intercom created roles like conversation designers and distributed AI across every product team rather than siloing it — a mistake Adams calls "bolting it on," analogous to how Google sequestered mobile into a side team until it became everything.

Four frameworks Adams uses: Differentiation vs. table stakes — startups should skew toward differentiation early, then rebalance as they displace incumbents; Intercom swung 70/30 in both directions at different points. Swinging the pendulum — overcorrecting is unavoidable; recognize you've crossed a boundary and course-correct. Product-market-story fit — Rdio had a better product than Spotify in the same market and lost because the story was wrong; story matters as much as the product. Before/after moments — inflection points that reset assumptions and require re-learning from customers.

On pricing: compounding complexity (tiers, add-ons, tiered add-ons) leaves customers unable to understand their bills. Fight every incremental pricing layer.

On Jobs to Be Done: use it simply — what is the customer trying to do, how much energy around it, and what forces govern switching (attraction, anxiety, habit, push from status quo)?

On reference calls: "What feedback will I be giving this person in their first performance review?" — the referee cannot dodge it.

Lessons learned from a startup that didn't make it

Jake Fuentes shut down Cascade in 2023 after four years and $5.3M raised, naming four compounding mistakes. First, a frayed ICP: "nontechnical business analysts using Excel" was too broad — it said nothing about the actual problem, so the team drifted across logistics, HR, and retail simultaneously, scrambling every product decision. Second, horizontal products only win when one buyer type faces enough varied problems to want a single tool; Figma and Notion passed that test, Cascade's users did not. Third, Alteryx looked ripe for disruption but wasn't — community, embedded workflows, and power-user career investment made true switching costs far larger than the apparent installation cost. Fourth, early deals closed on founder relationships are not market signal; the car moved only while the founders pushed.

How to tell better stories | Matthew Dicks (Storyworthy)

Every good story is built around a single five-second moment — the instant of transformation or realization where someone used to think one thing and now thinks another. That moment determines both ending and beginning: start at the opposite of where you land. Three tests: it contains a change, it passes the dinner test (tell it like you'd tell a friend, slightly elevated), and it centers on you — telling another person's story strips away the vulnerability audiences connect to.

Stakes keep people listening. Five devices: the elephant (plant a worry at the start), the backpack (tell the audience your plan before executing so they share your hopes), the hourglass (slow down when tension peaks), breadcrumbs (hint at what's coming), and crystal balls (articulate a bad possible future). Spread stakes across the story; don't front-load. Surprise is separate: it lands best when it feels both inevitable and unexpected.

Four things overcome audience indifference in business: stakes, surprise, suspense, and humor. Nostalgia is the easiest technique — state how something used to work and the absurdity surfaces on its own. When you need a story on demand, don't match content to content; match theme and meaning, then snap it onto your subject. A scientist who told an apple-shopping story about giving his family choice — never mentioning tubes — generated more conference leads than four data-presenting colleagues combined.

The foundational practice is Homework for Life: each day, log the one moment worth telling in a single spreadsheet row. After twelve years the count rises from 1.8 moments per day to 7.6 — the lens sharpens, not the life. Past memories surface and behavioral patterns become visible only across years of entries.

Open every story with location and immediate action: one word of location activates imagination; action signals the movie has started.

What to do if your product isn’t taking off

When a product stalls, the failure traces to one of a few fixable variables: wrong users, wrong message, insufficient reach, or a wrong bet.

Start with user conversations. Look for pain (hatred of incumbents, strong emotion) and pull (unsolicited payment, continued use of a bad prototype). Localmind users loved the experience but the pain was trivial — it stayed a novelty.

If indifference persists, change the target audience before the product. Pinterest's real users were female bloggers; Retool's were CTOs. ICPs should be comically narrow — three or more narrowing characteristics — focus enables depth and keeps early adopters reachable.

If the audience is right but interest is low, reposition. April Dunford reframed a product from "Access killer" to "embeddable mobile database" without touching the code.

If positioning is clear but reach is thin, run unscalable kickstarts to find the first 1,000 users.

If none of this works, find the feature with organic pull and pivot fully. About 40% of successful B2B companies pivoted — Instagram cut everything except photos; Discord jettisoned its game.

PMF takes longer than expected — median two years in B2B, 6–18 months in consumer. If conviction has deepened, stay. If it has eroded, quitting is rational: Seth Godin notes winners quit fast and without guilt; Phil Knight argues knowing when to give up is genius.

How Gong builds product

Gong's product system runs on radical pod autonomy: once a pod owns a problem area, leadership barely touches design or iteration. Teams organize around customer outcomes (e.g. "sales forecasting"), not features or engineering specialization. 24 PMs sit across six groups, plus two cross-functional exceptions — a data platform group and a user journey group.

Planning follows a "W" shape annually: management sets three or four priorities, capacity flows to pods, a bottom-up straw-man gets leadership feedback, then a detailed plan is written. The quarterly version trims to an "M" — a 20-plus-page document. Gong rejects Scrum (artificial deadlines undercut real trade-offs) and dropped OKRs after finding alignment costs outweighed benefits. Bug prioritization belongs to engineers, not PMs, on the bet that engineers have enough context and most bugs take less time to fix than to triage.

The core product claim is that every feature is co-developed with design partners from near-zero fidelity. For the Forecast product, a half-dozen partners got access to software in an "embarrassing state" for months, iterating weekly. The sales engagement product ran 20–30 design partners in parallel. Features stay in "limited availability" — sometimes with hundreds of customers — until they demonstrably deliver value before going GA.

Hiring centers on an ambiguous real-life assignment: an actual customer problem with recorded call snippets, solved and then defended before a panel. Gong will hire weak-solution-correct-reasoning over great-solution-poor-collaboration.

Inside TikTok: Culture, strategy, monetization, and more | Ray Cao (Global Head of Monetization Product Strategy and Operations)

TikTok runs on a principle called "context, no control": give people complete business context, then expect them to think as owners rather than just execute their job description. Ray Cao, who scaled TikTok's global monetization from a two-person team to hundreds across LA, New York, and Singapore, contrasts this with Google's siloed structure where people protect their own scope. The antidote is treating yourself as a business owner, then collaborating reactively when you need to act.

The two hiring signals TikTok screens for: curiosity and discipline. Cao built the team fast and wrong in early 2020 — 100 hires in six months under pressure — and learned that wrong-caliber people slow you down more than understaffing does.

Three structural moves keep sales, product, and engineering close: quarterly 180-person doc-reading sessions (Amazon-style memos, no slides), PMs and engineers joining advertiser immersion trips to feel client friction directly, and annual team reshufflings to match market need rather than preserve org chart stability.

TikTok differs from Meta and Google as an ad platform in one structural way: it operates on a content graph, not a friend graph or intent graph. Broad early targeting outperforms narrow retargeting; the platform needs volume to learn. Cao's benchmark: at least 10 creative variants per week. The machine tests and distributes — the advertiser's job is generating creative surface area, not optimizing a single ad. Give it a minimum of one month. The "TikTok made me buy it" trend (billions of views organically) is what led TikTok to build TikTok Shop — product following observed user behavior, not internal roadmap logic.

On globalization: the algorithm still requires local humans to seed relevant content. Japan's top content is food and consumer electronics, not dance. TikTok Shop launched in Southeast Asia before the US.

Kunal Shah on winning in India, second-order thinking, the philosophy of startups, and more

Products achieve irreversible adoption only when users rate them at least four points higher in efficiency than the alternative on a 1–10 scale — the Delta 4 framework. Below that threshold behavior is reversible, failure tolerance is zero, and word-of-mouth never ignites. Above it — Uber over cabs, LLMs over prior search — adoption locks in and CAC approaches zero.

India's market defies Silicon Valley orthodoxy. DAUs come cheaply — lowest global data costs plus high smartphone penetration — but ARPU is capped by ~$2,500 per-capita income. Meta likely earns $3–4 per Indian user annually; Netflix and Spotify hit the same ceiling. The cause: no Indian has ever been paid an hourly wage, so time has no unit price, and many Indian languages lack a word for "efficiency." Low-trust markets concentrate trust in conglomerates like Tata (salt, cars, jewelry), making brand more powerful than product. The "focus on one thing" orthodoxy fails when ARPU is this thin. CRED targets only 25 million high-income families where time has value.

The prevalence of Indian-born CEOs at Microsoft, Alphabet, Adobe, and IBM stems from IIT filtering, immigrant hunger, and math culture — but Shah's sharper explanation is dharma: sustaining a founder's values rather than imposing new identity. Tim Cook preserving Jobs is the example. Effective leaders oscillate between Krishna (high values, low obedience — opportunistic) and Rama (high values, high obedience — scaling), discouraging the ego plays that erode companies.

Wealth is stored energy, not a zero-sum ledger; concentration is physics and the total is infinitely expandable. Curiosity compounds information asymmetry: the most durable species (sharks, crocodiles) slash metabolism at will, convert hunts at high rates, and adapt across environments. Second-order thinking — mapping butterfly effects — is the strongest predictor of success, built through strategy games and "why" questions early in life.

A framework for finding product-market fit | Todd Jackson (First Round Capital)

Product-market fit for B2B companies is a four-level progression taking four to six years — 60–70% of startups stall at the first two.

Extreme PMF requires demand, customer satisfaction, and delivery efficiency simultaneously. Efficiency is the element founders drop — WeWork and Casper had the first two but never achieved viability. The test: each successive customer should be easier to acquire and serve.

Level 1 (nascent, 0–500K ARR, under 10 people): find three to five paying customers with an urgent problem and deliver a satisfying solution. Efficiency is irrelevant — Vanta's Christina Cacioppo manually completed SOC 2 audits for Segment, Front, and Figma. Warning signs: customers wouldn't miss the product, or each needs a different core feature (the "consulting business" trap).

Level 2 (developing, 500K–5M ARR): scale from five to 25 customers by adding demand generation. Ironclad broke through by repositioning from "AI legal assistant" to CLM — a category buyers already had budgets for. Regretted churn above 20% signals stall. Jack Altman: "Most founders do a 10% pivot when they need a 200% pivot."

The four pivot levers are Persona, Problem, Promise, and Product. Vanta changed all four. Lattice kept the persona (HR heads) and changed the other three. Plaid kept the product code and flipped the rest.

Level 3 (strong, 5M–25M ARR, ~Series B): one scalable demand channel, 10%+ inbound from referrals, burn multiple below 3, NRR above 110%. Efficiency must now be actively managed.

Level 4 (extreme, 25M+ ARR): gross margin above 80%, burn multiple below 1, NRR above 120%. Growth requires TAM expansion — each new product must find its own PMF from scratch.

On customer discovery: listen for wow statements or demonstrated pull (the word "interesting" is a polite no), confirm an existing budget, and use fair/expensive/prohibitively-expensive price questions to bracket real willingness to pay.

How should you monetize your AI features?

Most companies default to bundling AI into existing plans (59% of 44 incumbents studied), but direct monetization is usually the better long-term choice: gen AI carries real variable costs — compute, storage, compliance — that indirect retention gains rarely cover.

The decision follows two axes: how broadly is the feature used (70% threshold) and how strongly do users want to pay for it. High breadth + high willingness = bundle with a price increase. Narrow but high-value = add-on. Genuinely different problem = standalone. Standalone works only for near-full LLM products; add-ons suit features valued by a subset (Notion AI: $10/user); bundled-with-increase suits core features, distributed across tiers as Canva does.

On price: nearly all companies use per-user monthly fees ranging $4–$30. GitHub Copilot at $19 is 4.75× its base price, supported by a reported 55% coding speed gain; Microsoft Copilot at $30 exceeds the M365 subscription, citing 70% productivity gains.

The one pricing-model outlier is Intercom's Fin, charging per resolved conversation — outcome-aligned pricing likely to spread as the application layer matures.

Lessons in product leadership and AI strategy from Glean, Google, Amazon, and Slack | Tamar Yehoshua (Product at Glean, ex-Google and Slack)

Product-market fit covers a multitude of operational sins: chaotic companies with high executive turnover still post extraordinary growth, while well-run organizations sometimes flatline. What determines success is whether people want the product, distribution works, and capital exists to get there. Professional management becomes critical only past 5,000–10,000 employees.

Tamar Yehoshua — Glean president, ex-CPO at Slack through its 10x revenue growth and IPO, earlier a Google Search lead and Amazon/A9 VP — argues career compounding comes from following exceptional people, not planning around domains or financial upside. In quarterly meetings Bezos polled every executive before speaking, remembered architecture details across quarters, and held consistent principles — customer primacy, no unexplained icons — so the organization knew what he'd push on. When she cited a 10x competitor headcount advantage, he replied: "That is your advantage."

Stewart Butterfield wrote Slack's master plan in 2014 as four boxes: product love, Slack Connect, Slack Platform, magic AI stuff. Annual work changed; the boxes never did. His evaluation tool was real prototyping on real data — he once asked the team to put every interface element behind one button, not to ship it, but to reveal what actually needed to exist.

On AI: enterprise users tolerate non-determinism in personal ChatGPT use but expect determinism at work — a product education problem distinct from the technical one. One PM fed an entire Discord community into Gemini's expanded context window to surface feature sentiment no human would parse. Yehoshua's own Glean prompt aggregates Launch Cal dates, open Jira tickets, and beta-customer Slack threads into a feature-confidence view. The strategic rule for AI builders: never differentiate on compensating for today's LLM weaknesses — that advantage disappears as models improve. Execution-heavy PMs face the most displacement; those who specify what to build are better positioned than any other function.

Building Wiz: the fastest-growing startup in history | Raaz Herzberg (CMO and VP Product Strategy)

Wiz hit $100M ARR in 18 months — faster than any software company — but started with the wrong idea. The team pitched network security in 10–15 daily calls; meetings ended with "sounds interesting." First PM Raaz Herzberg realized she couldn't describe what they were building and said so. That triggered a five-hour founder meeting pivoting to cloud security, where CEO Assaf Rappaport had run Microsoft's cloud security division. Prospects asked price, demanded POV timelines, connected them to technical teams. A Fortune 10 company returned a questionnaire within 24 hours to begin a POV on an unfinished product. Real pull: customers doing annoying work to get your product.

Founders closed several million ARR before hiring sales: if the team can't close end-to-end, hiring won't fix it.

At two and a half years Wiz won every POV but lost deals where prospects had already signed competitors. The CEO asked Herzberg — who had never used "pipeline" in business — to become CMO. At RSA she replaced the dark booth with a Wizard of Oz theme: pink, blue, actors. Foot traffic was five times the prior year.

Crossing from product to marketing, she found product messaging can be fuzzy because people fill gaps in the room — marketing can't; messages degrade at every handoff. Insiders say CNAPP; buyers Google "cloud security solution." Her "dummy explanation" rule: write as if readers know nothing about your company or jargon, and don't change messaging because you're tired of it — customers are still absorbing what you posted months ago.

CMOs fail because the role demands founding-team trust (one bad ad breaks it) and domain knowledge outside hires rarely carry.

Wiz declined Google's reported $23B offer. Only 15–20% of infrastructure has moved to cloud; the company intends to be the defining player when the rest follows.

Seth Godin's best tactics for building remarkable products, strategies, brands and more

AI will stop being a feature the same way electricity stopped being one — what matters is the specific promise a company makes and whether it keeps it. That frame runs through everything here.

Good taste is knowing what others want just before they do. High standards mean relentlessly raising the spec in service of users, not refusing to ship out of perfectionism — perfectionism is hiding. Quality is meeting spec; once met, you're done. Godin learned this as a 24-year-old PM at Spinnaker Software, building games around Ray Bradbury and Arthur C. Clarke novels: the product exists to delight the user, not please the boss.

A brand is a promise, not a logo. Nike has a brand; Hyatt has a logo — you know what a Nike hotel would feel like, but a Hyatt sneaker tells you nothing. Claude earned a brand by showing humility when it doesn't know something; ChatGPT over-promises and under-delivers. Airlines have no real loyalty — only points, which is bribery.

The four choices that determine your product's future, from *This Is Strategy*: (1) Choose your customer — the smallest viable audience determines everything downstream. The Humane Pin failed by targeting Apple-quality buyers before it was Apple-quality. (2) Choose your competition — entering Walmart's arena means accepting a price war. (3) Choose your validation source — agreeing on whose taste you're matching changes every subsequent meeting. (4) Choose your distribution — when Steam arrived, everything about game products changed.

Tension, not stress, is the engine of innovation: the gap between a promise and whether it holds. Remarkable means worth making a remark about; Google's two-button homepage made recommending it raise the recommender's status.

On Jaguar: they re-logoed, not rebranded, discarding earned equity for buzz. On the Cybertruck: pickup trucks are the largest US vehicle category, waiting to be taken — Tesla chose divisiveness over conversion. Leadership means painting the picture of where they want to go, not where you do.

Superhuman's secret to success: Ignoring most customer feedback, manually onboarding every new user, obsessing over every detail, and positioning around a single attribute: speed | Rahul Vohra (CEO)

True virality is word of mouth — unmeasured, unmechanized — not viral mechanics. LinkedIn's Elliot Shmukler showed Vohra that no product sustains a viral factor above one for long; even Facebook peaked at 0.7 for a year. Build something worth talking about.

Superhuman's positioning bet is speed. After interviewing hundreds of customers who universally complained email was slow, Vohra confirmed the position was unclaimed — almost no software had been sold on speed since Chrome. The "cocktail party test" — watching users pitch to friends — yielded: "Dude, you have to use it, it's really fucking fast."

Product-market fit is measurable using the Sean Ellis question ("how disappointed would you be without this product?"). The engine: ignore not-disappointed users; don't over-serve the very-disappointed. Focus on somewhat-disappointed users whose main benefit resonates but who have one specific friction. Spend half the roadmap doubling down on what fans love, half removing that friction. The algorithm writes the roadmap.

Manual one-to-one onboarding ran until ~20 people were doing it. Payoff: engagement, retention, and PMF scores above benchmarks, plus a superfan cohort that bootstrapped word of mouth. It stopped when self-service became the bottleneck.

Pricing used Van Westendorp. The anchor is "starts to feel expensive but you buy it after thinking about ROI" — not the bargain price. Median answer: $30/month, backward-computing to 300,000 subscribers for $100M ARR.

Org design was the hidden velocity lever. Tracking task switches via Slack revealed Vohra spent 6–7% of his week on product, design, and marketing. Hiring a president cut his reports from eight to two; that share rose to 60–70%.

Decisions use Single Decisive Reason (SDR), from Reid Hoffman: name one reason that alone justifies the call. Multiple weak reasons rarely compound into a strong one, and every feature carries the opportunity cost of what wasn't built.

How Intercom rose from the ashes by betting everything on AI | Eoghan McCabe (founder and CEO)

Intercom was heading toward zero net-new ARR in late 2022 when GPT-3.5 launched. Six weeks later the AI team had a working prototype of Fin, the customer-service agent that became the company's entire future.

McCabe had stepped away as CEO in 2020 due to illness, and the company drifted into late-stage bloat: unfocused strategy, pricing so complex it became a Twitter meme. When he returned he describes having "nothing to lose" — which made an all-in bet possible for a company doing hundreds of millions ARR.

The turnaround had four planks. First, cost cuts and picking a single lane — customer service — even while $80M in ARR came from other areas. Second, a $100M internal AI investment. Third, a culture reset: McCabe rewrote values around resilience and shareholder primacy, hard-coded a formula that automatically cut employees below threshold on goals and values alike. Roughly 40% of staff turned over across two to three years; there were board letters and a failed soft coup. Survivors produced a 98–99% management approval rating. Fourth, outcome-based pricing at $0.99 per resolved ticket — B2B companies were paying $20–30 for a human-resolved ticket. Fin initially cost $1.20 per resolution; McCabe priced to value and let model costs fall.

Results: Fin grew from $1M to $12M ARR in year one, now exceeds $100M ARR growing above 300%. Intercom overall moved from low-single-digit growth to the 15th percentile among ~120 public software companies.

On jobs and the future: CX roles largely automate away, SDR-style sales shrinks, but human trust and creative craft persist and command a premium in an AI-abundant world. McCabe's advice to incumbents — hire actual AI scientists, empower younger engineers who use AI natively, and go all in or stay out. Half-measures lose to startups that treat AI as their native operating environment.

How to find the perfect name

Great product names don't describe — they signal a new idea. Descriptive names like Infoseek flatten into noise; invented names like Azure, Pentium, and Navan earn attention because humans are drawn to the unexpected. Invented names also cost less to build into brands than existing words, contrary to common belief. The Lexicon process generates 1,000+ candidates before shortlisting, uses sound-symbolism deliberately (the "-ium" suffix made Pentium feel elemental), and runs a Diamond Framework — "what does winning look like / what do we need to say?" — to align strategy before generating names. Large brainstorming sessions consistently underdeliver; small teams outperform them.

“Dumbest idea I’ve heard” to $100M ARR: Inside the rise of Gamma | Grant Lee (CEO)

Gamma reached $100M ARR and a $2B valuation in just over two years with ~30 people, profitable throughout — despite an investor calling the pitch "the worst idea I have ever heard" and hanging up mid-call.

The critical inflection was a near-death rebuild. After winning Product Hunt's product of the month in August 2022, signups plateaued and word-of-mouth never ignited. The team spent three months rebuilding onboarding so every user experienced AI generation in the first 30 seconds. Relaunching in March 2023, signups went from hundreds to 20,000 per day organically. Product Hunt wins are vanity metrics; the real test is whether the product grows without spend.

Growth from 10M to 100M ARR came from three buckets: word-of-mouth (over 50% of new signups throughout), influencer marketing, and paid performance. Grant personally onboarded every early creator on a call, letting them tell Gamma's story in their own voice. The right approach is thousands of micro-influencers with real audience fit — educators who save time on slides, not million-follower celebrities handed a script — at $200–2,000 per post. 90% of reach comes from under 10% of creators, so cast wide and iterate monthly. LinkedIn converts at 4–5x other platforms. Gamma open-sourced its brand kit at Brand.gamma.app so creators don't rebuild assets.

Gamma runs 20+ models simultaneously — outlines, first drafts, visual review, image selection — constantly testing quality against inference cost. This orchestration layer is what makes the "wrapper" durable.

Pricing launched at ~$20/month in May 2023 after Van Westendorp surveys, anchored to ChatGPT's price. Gamma hit $1M ARR and turned profitable within months. All 10 original employees remain five years later. Every people manager also does IC work. Roughly a quarter of the team is product designers — because new AI surface areas require getting the experience foundation right before scaling.

Jeetu Patel

AI is a megatrend, not a hype cycle — the test is whether a layperson can immediately grasp what the technology does. Web3 failed; ChatGPT passed in seconds. Jeetu Patel, CPO and president at Cisco (90,000 employees, 30,000 in his org), credits AI with making his role possible: taking over with no networking background required a three-month crash course otherwise impossible.

Cisco's transformation rested on three moves. First, refusing to hedge once AI commitment was declared — large companies experiment plenty; what they fail to do is double down when experiments work. Second, dismantling the GM-fiefdom incentive that fragmented a $40B company into silos of $40M businesses, replacing it with a platform model: loosely coupled, tightly integrated. Third, abandoning zero-sum competitive thinking, because customer success flows back regardless of which partner sits beside you in the stack.

Cisco's infrastructure role: GPUs are useless unconnected. Cisco networks them across racks, clusters, and data centers 800 km apart into a single coherent training cluster. Three things threaten AI progress: infrastructure constraints, a trust deficit (hallucination is fine in poetry, fatal in predictable systems), and a data gap as human-generated internet content runs thin.

On leading at scale, he inverts standard advice. "Praise public, criticize private" is wrong: establish trust privately so direct public debate becomes possible — posturing crowds out problem-solving. Own the story yourself; delegation introduces packet loss across layers, and what reaches the front line through seven relays bears little resemblance to what was said at the top.

His six-factor company framework, stack-ranked: timing (most important, least controllable), market (bad markets beat great teams), team, product, brand, distribution. All six required; ranking determines where to focus. Career advice: pick hard problems, seek compounding platforms over mere roles, and remember stamina beats intellect because hunger cannot be taught.

Father of the iPod and iPhone on building taste, judgment, and creativity in the AI era | Tony Fadell

Genuine 1.0 products can't be designed by data — the data doesn't exist yet. When Apple debated physical versus virtual keyboard for the iPhone, evidence was roughly even. Steve Jobs made an opinion-based call and dismissed dissenters. New categories need one or two "taste makers" with informed judgment and authority to take the heat. User studies on genuinely novel products produce rationalization, not signal.

Finding the right problem starts with chronic pain tolerable only because the technology to fix it didn't exist. Nest's $249 thermostat cost five to six times more than competitors but saved $800–$1,200 a year — on-device AI learned occupant schedules and replaced the VCR-programming interface nobody used. iPod was portable mass storage plus lithium-polymer batteries plus MP3s converging. iPhone added multi-touch, wifi, and coming 3G. Formula: habituated pain plus newly available technology.

Three generations is the pattern: make the product, fix it on real feedback, fix the business (margins, distribution, ecosystem). Early iPod only sold to Mac users — under 1% of the market. Jobs blocked Windows for two generations; the counter-argument that won was that requiring a Mac raised the effective price from $349 to $3,000, killing mass adoption. The Windows-plus-iTunes third-generation iPod saved Apple from near-bankruptcy.

Marketing is not downstream of product definition — it is part of it. A consumer absorbs three or four features before it becomes noise, so knowing which three forces early cuts. OpenAI's positioning confusion was the consequence of shipping a technology demo before doing product thinking.

On AI: code generated without architectural discipline accumulates technical debt. Humans should lock the architecture and delegate sub-functions to agents. The next phone will still need a screen — displays remain the only viable high-bandwidth output short of BCI. Voice should become primary input, but consumer trust will take many iterations.

What to ask your users about Product-Market Fit

PMF assessment triangulates three signals: affection (Sean Ellis's "very disappointed" question, 40% threshold), relative utility vs. alternatives, and behavioral evidence of actual use. With 100+ users, survey all three. With fewer, interview 6–12: walk through their last key task, compare alternatives, and ask why they tried your product. The signal is clear articulation of superior value plus genuine enthusiasm, not hedged "potential." Avoid NPS and asking users what they want.

A playbook for fundraising

Fundraising is a trust-building exercise: each VC partner makes 1–3 early-stage investments per year and stakes their career on each bet, so the whole process is about convincing them you will build something huge.

Series A is the product-market-fit round — a growing, paying (or highly engaged) user base in a defensible market. SaaS companies cite $1M ARR, but growth rate often matters more. Prepare for at least four weeks: a short teaser deck (3–7 slides), a full pitch deck (12–15 slides), and a 2–3 year forecast. Narrative beats data — Eoghan McCabe credits Intercom's fundraising success to linking the company to broad societal shifts. Before going wide, "harden the pitch": test it at 70% completion with 4–6 friendly founders or investors to stress-test the thesis and collect warm intros. Enter with at least 8 months of runway; below 6 months, bridge first. Plan 8–16 weeks from first outreach to term sheet.

Target 50–60 funds with specific partners identified. The strongest intro comes from a portfolio founder inside that fund. Associates are useful scouts, but always reach a partner.

Compress all first meetings into a 2–3 week window so no fund races ahead on diligence. A fund that gets far ahead creates a market of one buyer. Y Combinator estimates 30 pitches per term sheet; 90%+ rejection is normal. Five funds still engaged after diligence is a success.

Once a partner wins over their partnership, they become your co-conspirator. When a term sheet arrives, don't reveal which fund offered or at what valuation. Term sheets typically last 1–2 weeks; signing triggers another 3–6 weeks of legal diligence before money lands.

When to hire your first product manager

Someone always does PM work — shaping, shipping, and synchronizing — whether or not the role exists. Hire a dedicated PM when three symptoms appear: the founder becomes a bottleneck (Ryan Hoover's Product Hunt regret); teams misalign on priorities (how PM was invented at Microsoft); or post-PMF a focused problem needs full attention, as Twitter and Lyft each did. Delay until PMF; most companies hire their first PM at employee 15–100.

Winning at early-stage hiring

Founders who hire consistently well do five things that struggling founders don't: articulate a vivid, ambitious vision (Siqi Chen at Runway credits the product demo as their single biggest draw); set the bar for early hires even higher, because A++ teams compound through referrals and social proof; budget ~100 hours per hire and court passive candidates for up to six months; mine the founder's personal network first; and close candidates the way you close investors — Stytch surprises finalists with a Zoom of the full interview panel, Vori delivers a custom economic-outcomes deck per candidate.

Why now

A strong "why now" raises your odds but isn't required — SpaceX, Airbnb, Pinterest, and DoorDash had no compelling timing thesis and won through execution and 10x product improvement. Where timing does matter, it works through two mechanisms: enabling a better product (WebGL unlocked Figma; smartphones with GPS unlocked Uber) or opening an untapped market (COVID for Zoom; e-commerce growth for Shopify). Five sources drive this: new technology ubiquity, newly available APIs/data, regulatory shifts, behavior change, and belief change. Timing can emerge post-founding: Netflix rode broadband, PayPal pivoted from PalmPilots to web payments. A great tailwind doesn't guarantee survival — Quibi had COVID and died anyway.

Lessons from 140+ angel investments

Angel investing rewards broad participation over precise picking. After 140+ investments over five years (12 unicorns), only a third of the biggest winners were rated "Great" conviction at entry — meaning selective picking would have missed two-thirds of the best outcomes. AngelList data confirms this: early-stage investors do better indexing every credible deal than trying to select the few winners. Power laws make this stark: 70% of total paper gains came from a single company, 80% from four.

Access matters more than judgment. Over two-thirds of the biggest winners came from "hot" rounds, and 80%+ of deals co-led by top-tier funds tracked toward big outcomes. Deal flow itself evolves — early investments are mostly friends, but the majority eventually comes from other investors. Great deals circulate as currency: share your best opportunities, reciprocal flow follows. Top-tier lead is a strong signal worth following for most bets; reserve ~30% for low-signal companies with unique insight.

Getting onto cap tables means becoming someone founders want there. Four paths: develop expertise they need (growth, ops, hiring); build an audience that amplifies their story; accumulate signal so your name on the cap table means something; earn founder NPS through genuine service — helping find one engineer puts you in the top 0.01% of investors by perceived value-add.

When evaluating: severity of pain, market size ($1B+ revenue potential), evidence of PMF, founder execution speed and unique insight, unfair advantage, business model quality, overall trajectory. Brett Berson's framing: assess the next 18 months concretely — team, early customer love, market pull — not the 18-year TAM story.

Entry is more accessible than it appears: first checks can be $3K–$5K via AngelList syndicates. Losses materialize before wins; plan to invest in at least 30 companies before a realistic shot at a rocket ship.

What is a good growth rate

Early-stage B2B investors ignore month-over-month percentages on a small base and instead track time to $1M ARR — reaching it in 12 months is good, 9 months is great. Pre-launch is safer than slow post-launch traction; bad data is worse than no data. If ARR ramp is slow, strong retention and cohort curves (U-shaped NRR) can substitute. For B2C, revenue growth is irrelevant early — what matters is DAU/MAU above 50%, D30 retention above 50%, and organic K-factor. Once past $1M ARR, the standard benchmark is triple-triple-double-double-double (3×, 3×, 2×, 2×, 2×). By funding round: Series A targets 3× annual growth, Series B and C similar. These are heuristics, not mandates; retention, CAC payback, and unit economics always run alongside the growth rate.

How to validate your startup idea

Real validation means finding a problem so acute that customers reach for your solution before you've finished building it. Todd Jackson (First Round Capital) interviewed eight founders to surface what genuine market pull looks like.

Three paths to an idea map to company type. Market-first: Christina Cacioppo of Vanta interviewed founders about their days until she heard SOC 2 was blocking revenue, built a spreadsheet MVP, and knew she had something when strangers she'd never spoken to started calling her unprompted. Experience-first: Kevin Tan watched Yale students queue at G-Heav after Uber Eats launched, built a fax-order website over Thanksgiving break, and had 90% of campus using it within six months; a social-points gifting layer drove expansion to Brown with almost no budget. Problem-first: Edith Harbaugh had watched releases fail at multiple startups, knew Facebook and Dropbox had internal feature-flag systems, and built LaunchDarkly to sell that externally. VCs called it a $5/month developer tool; she got one customer per month for eight months, funded on revenue, and eventually mailed skeptical VCs a copy of the check stub from her first six-figure contract.

Two cases show how the idea itself can emerge mid-validation. Kurtis Lin (Pinwheel) was building HSA plumbing when payroll integrations proved the real bottleneck; pitching direct-deposit switching to prospects, he kept getting interrupted — "Wait, you can do what?" Ryan Petersen (Flexport) posted fake product screenshots with Google ads; 300 companies signed up for a nonexistent product, including Saudi Aramco, which reset his market-size assumptions.

Seven traps kill validations: asking hypotheticals instead of past-behavior questions; settling for polite reactions rather than genuine pull; too few conversations (B2B needs 50+); building vitamins not painkillers; ignoring go-to-market; pivoting too early; and pivoting too late. PMF is unmistakable — if you're asking whether you have it, you don't.

When and how to invest in new acquisition channels | Adam Grenier (Uber, MasterClass)

New acquisition channels succeed roughly 5% of the time; the rest are resource sinks chased for heat rather than fit. Grenier's three-filter framework determines when to invest. First, channel-customer-goal overlap — does the channel's native format match what your customers need and what your business is trying to achieve? Spotify fits Clubhouse (audio, discovery, artist intimacy); it doesn't fit Paparazzi (photo-only). B2B companies routinely ignore this and chase consumer channels that can't reach their buyers. Second, channel DNA — where is the platform in its growth curve, and how does it monetize? Early-curve platforms are risky because any working tactic can vanish as the product evolves; Facebook's viral notification loop that fueled Zynga disappeared when Facebook killed it. If your business model supports the channel's monetization strategy, you gain leverage: Grenier got HotelTonight into Facebook's mobile-ads alpha by positioning them as the non-gaming case study Facebook needed to broaden its advertiser base. Third, company DNA — do you have risk appetite for true first-mover chaos (no tracking, no playbooks), and have you extracted solid volume from Google and Facebook already? Without that foundation, channel exploration is a distraction. Don't let experiments run past a quarter; directional signal arrives within a month.

The Growth CMO concept explains why world-class traditional CMOs keep failing inside product-led companies. Traditional marketing plans in campaigns and long roadmaps; product-led companies need agile iteration applied to everything — brand, landing pages, pricing, copy. The fix is learning product development methodology; Grenier recommends *Hacking Marketing* as a starting point.

On burnout versus depression: exhaustion leaves non-work motivation intact — depression kills it across the board. You cancel the improv class, skip the post-class drinks. The earliest workplace signal is declining adaptability: people shift from "let's try it with caution" to "why bother" and start minimizing challenge rather than embracing it.

Building Substack | Sachin Monga (Substack, Facebook)

Subscription economics — not advertising — is what finally makes the internet work for writers. Sachin Monga reached that conclusion at Facebook, where ad-driven incentives forced products to maximize time-on-site, and tested the opposite model with his startup Cocoon before Substack acquired the company.

Monga joined as head of product when Substack had zero PMs and roughly 15 engineers. His first months were spent not on roadmapping but closing the communication gap between CEO Chris Best — who had carried the five-year vision largely alone — and a team now too large to absorb it organically. Weekly bookend syncs with Best became the trust mechanism. He structured teams around customer types (writers, readers, growth) rather than product surfaces, a discipline he credits with unusual staying power compared to Facebook's quarterly reorgs.

The Recommendations feature illustrates how the writer-control principle drives product decisions. The obvious mechanism — algorithmically inserting "you might like" units — was rejected because it hands control of the publication to Substack. Instead, writers nominate peers; picks appear in the subscribe flow; recommended writers get notified of every subscriber sent their way, creating a goodwill viral loop. Best doubted the multi-step opt-in would generate real adoption. It drove millions of subscriptions across tens of thousands of writers. By 2022, over one in three new subscriptions and roughly one in ten paid subscriptions come from the Substack network. Lenny Rachitsky reports 70% of his own subscriber growth from this single feature.

On big-company versus startup PM work: Facebook trade-offs are often permanent — doing A forecloses B forever. At a startup, the main variable is time — A now means B later — and sequencing matters because each capability unlocks the next. Any process you get right becomes obsolete as soon as the company grows through it.

For new writers: start without high stakes, don't fear charging, take breaks — paying subscribers are more forgiving than writers assume. A thousand at $10/month is a living, and the actual market is almost always larger than it looks from inside.

Startup to exit: Lessons from a first-time founder

Good startup ideas come from honest customer research, not pattern recognition. Suril Kantaria co-founded Savvy — a fintech payment card for employee health insurance — through YC, raised $2.5M backed by Marc Andreessen, and sold to Take Command Health in 2022 after three years and tens of millions in payment volume.

Six lessons. Savvy spent months building a health debit card vision before unbiased Mom Test interviews showed nobody needed it; they scrapped everything and started over. VC funding is a drug: Demo Day FOMO pushed a raise before conviction existed, locking them into a story that delayed the pivot. Retool's David Hsu stayed at co-founders only for a full year post-seed to preserve pivot speed. Headcount looks like progress from outside but isn't — PMF is the only metric that counts.

Building in stealth cost Savvy an estimated six months of signal. Startups win through execution, not secrecy; inbounds from public building — customers, hires, investors — compound faster than any first-mover advantage stealth might protect.

The most dangerous outcome is steady sales without acceleration. Savvy's large early customers weren't desperate enough and churned once product cracks showed, after consuming disproportionate ops and engineering time. Pre-PMF, overwhelming pull or clear rejection both give signal; middle-ground push-selling just burns runway.

When conviction in the market collapsed, options were shutdown, restart, or sale. Savvy ran acquisition as a sales funnel: 50 targets ranked by capitalization, VC/PE backing as a speed proxy, and accessible C-suite. Early acquirer interest should never be dismissed — they turned Gusto down cold and later found their "strategy had shifted." Approach targets as potential partners, not buyers. Once a term sheet is signed, compress timelines: a bear-market close nearly collapsed before Kantaria pushed the CFO past a bank lockout at 10 p.m. to wire funds that night.

How to activate your investor network

When founders fail to get value from investors, it's almost always the founder's fault — investors lack context and won't proactively help. Sam Corcos of Levels sent 2,626 personal investor requests over 3.3 years; 1,151 converted into delivered value.

Network construction comes first. Optimize for eigenvector centrality, not degree centrality: one connector into each distinct dense network beats deep ties within one cluster. Granovetter's 1973 "Strength of Weak Ties" backs this. Levels' first 15 alphabetical angels spanned 15 domains — a basketball player, a YouTube creator, a Fortune 500 sales leader, an insurance founder. Set participation expectations before taking money; let investors self-select out.

Good asks are targeted (only relevant investors), time-bounded (ideally under a minute), and specific ("your two best engineers," not "engineers you know"). Keep a Google Sheet of investor backgrounds and scan it per need. Levels reached Jeff Jordan and Sander Daniels for marketplace advice, Katie Biber Chen for legal guidance, and investors to review growth-hire interview recordings.

Women on the cap table outperformed by 10x on allocation-to-conversions; early post-IPO employees delivered outsized candidate pipelines on small checks. Closing the loop with a thank-you after delivery raises future responsiveness — almost no founder does it.

Gustaf Alstromer

Almost every startup failure traces back to one cause: not talking to customers. Gustaf Alströmer, YC Group Partner and former head of Airbnb's growth team, makes this the center of every office hour. External validation — investor interest, press, YC acceptance — gets confused with product-market fit, which only customers can confirm. YC tells every new batch that almost none of them have it.

Fear of rejection explains why founders avoid users, but the realistic outcome isn't hatred — it's indifference; people forget they signed up. Reaching the ~10% who are early adopters requires 25 to 50 conversations, not five.

YC Office Hours take two forms. Individual sessions open with one question: "what's holding you back from moving faster?" — not strategy, not updates. Group sessions run biweekly goal check-ins across six or seven companies, creating peer accountability and revealing that every startup is equally broken. Many alumni continue this informally for years.

Strategy discussions before product-market fit signal confused priorities — early companies execute one thing at a time; strategy assumes a scale they don't have. Successful founders share four traits: determination that outlasts adversity; technical depth to value engineering (offering a co-founder 10% reveals the opposite); continuous customer contact over internal conviction; and storytelling that attracts talent and capital.

On climate tech: the energy transition moves trillions and governments and corporations have committed to it. The IRA locked in US incentives; corporations now show up as paying customers because they fear being the Toyota to Tesla. YC has funded 130+ climate companies including Pachama, Seabound (shipboard carbon capture), and Heart Aerospace (electric planes). Standard software skills transfer directly; domain expertise is complementary.

The best predictor of a breakout company is whether each office hour brings new problems rather than reruns of the same conversation from two weeks prior.

Navigating comms and PR | Lulu Cheng Meservey (Substack, Activision Blizzard)

Comms works best when you map your message to what people already care about — "cultural erogenous zones." Converting someone's worldview is a huge lift; fitting into what already lights them up is light. If there's no natural fit, they're not your audience.

Ideas stick when they minimize cognitive load: short enough for a second-grader, free of clichés, carrying a concrete image. Mitt Romney's "binders of women" went viral because you could picture the binder and meme it. The failure mode is inside jokes — an esoteric metaphor where no one else has the context.

For startups, comms conservatism is the mistake — errors of omission are harder to learn from than errors of commission. Build distribution in concentric circles: crystallize the message at your own desk, then spread to co-founders → employees → investors → power users → influencers. Skipping a circle is fatal; if employees aren't believers, every outer ring follows their lead. Disgruntled early employees, credible and in your same networks, can actively destroy the company.

The physics formula for underdogs: pressure = force / surface area. Same effort against a smaller, sharper target produces more impact. Balaji Srinivasan's *The Network State* skipped traditional book tours, concentrated on true believers, and hit number one on Amazon through organic evangelism. NYX cosmetics spent $0 on advertising and sold out ordinary black eyeliner by pouring everything into a small influencer roster.

Going direct matters offensively and defensively. Defense: if institutions won't stand up for you, only you can — prime the audience before a crisis, not during one. Offense: no reporter understands your mission as well as you do. Match channel to the spokesperson's natural voice (Elon Musk for short-form, Brian Armstrong for long essays). LinkedIn is underused — huge time-on-site, 95% filler. Consistency beats viral chasing.

Taxi mafias, cash vaults, and 100% MoM growth: The story behind Southeast Asia's biggest startup | Kevin Aluwi (Gojek)

Gojek hit 100%+ month-on-month growth for 16–18 months after its 2015 Indonesia launch — a pace Sequoia called the fastest they'd ever seen globally — built on $2M against a rival with $250M. The edge came from solving genuinely broken daily problems for a young, adoption-hungry population, and treating brand as a survival tool.

The company's foundation was a uniquely Indonesian phenomenon: the ojek, the motorcycle taxi. Branding drivers with green jackets and helmets wasn't just visual recall — commuters stuck in Jakarta traffic watched those riders whiz past with passengers and packages, making the value proposition self-demonstrating on the street. Car-centric competitors entering motorcycle ride-hailing missed this entirely.

Super-app expansion exposed the limits of that frame. A mobile top-up product relevant to 95%+ of users was unknown to 60–70% of customers despite sitting on the homepage. Users organized the app around "the driver," so driver-adjacent services (rides, packages, food) cross-sold cleanly — but anything outside broke down. When Gojek added massages, users asked if the driver would provide them. The promised super-app benefits of lower CAC and higher retention require reselling each new category, erasing the supposed advantage.

Operations looked nothing like a standard tech playbook. No digital payment rails existed, so Gojek built physical cash booths where drivers queued to collect earnings. When fraudulent driver apps spread by offering auto-accept features Gojek had disabled, they copied the fake apps' features rather than invest in security. When motorcycle-taxi mafias attacked drivers with bricks and machetes, Gojek hired private security firms to patrol hotspots, buying driver loyalty money couldn't match.

Aluwi drove for Gojek himself — one ride involved lugging a customer's laundry to a detour stop, directly informing later decisions on wait fees and multi-stop routing. His highest-impact governance change: making the decider explicit for every product decision.

Your startup idea probably isn’t venture-scale

Most startup ideas cannot reach $100M ARR and a $1B+ valuation in ten years — the threshold VCs need for fund economics — and the mistake is taking venture capital without qualifying. Beloved products like Trello, Basecamp, and DuckDuckGo generate tens of millions annually but are never venture-scale. VC money brings board seats, 2–3× annual growth expectations, dilution on misses, and forced exits within a decade. Three diagnostics: Is TAM over $5B? Is the pain a 9–10 (Hunter Walk's LUV — Large, Urgent, Valuable)? Does capital unlock step-change growth, not just sustain operations? If not, bootstrap, use Calm Fund or TinySeed, or try revenue-based financing. Patrick Campbell bootstrapped ProfitWell to $200M ARR — then concluded taking VC earlier would have unlocked a $1B exit.

Building minimum lovable products, stories from WeWork and Airbnb, and thriving as a PM | Jiaona Zhang (Webflow, WeWork, Airbnb, Dropbox)

The most common PM failure is solution-attachment — arriving with a specific thing to build before understanding user problems. The role is influence without authority; the job is editing possibilities, not calling shots.

Airbnb Plus illustrates this. Facing quality concerns, the team went solution-first ("inspect inventory") instead of naming the real problem: guests couldn't tell what they'd get. Unit economics never worked; the right answer — guest reviews, free and high-signal — went unused. To push back on a committed founder: understand what they want to achieve, then return with options that serve the same goal.

Minimum lovable product supersedes MVP when users have alternatives. The distinction is five things done excellently over fifteen done adequately. "Lovable" means targeted pixie-dust moments — keyboard shortcuts, pre-populated templates — exceeding expectations in specific spots. When MLP is out of reach, the ecosystem can carry it, with a clear-eyed read of where you are.

Roadmaps should tell a narrative — themes and reasoning, not a RICE spreadsheet. When a data point changes your read on users, shift the whole story, not just a score.

OKRs: the failure mode is sandbagging from fear of red. All-green means targets were too safe. Commit to ambitious north stars, break them into quarterly milestones with go/no-go gates — exit at one quarter instead of two years.

WeWork: don't over-hire ahead of proven milestones. The edge was operational inventory management; a large tech org adding futuristic features was a category error.

First 90 days: run 40–50 conversations across levels and functions. Trust is a bank; deposit before you withdraw — pushing for change too early is the most common new-leader mistake.

Across Dropbox, Airbnb, WeWork, and Webflow, the same pattern: chasing adjacent spaces instead of doubling down on why users love you. Know your core; extend from strength.

Reflections on a movement | Eric Ries (creator of the Lean Startup methodology)

Lean Startup went from insurgent to obvious so fast that critics moved straight from dismissing it to calling it overhyped, never pausing to understand it. Eric Ries — who coined MVP, pivot, vanity metrics, and the sticky/viral growth engines — describes that arc as what winning looks like: new founders take the vocabulary as given, the way continuous deployment now seems obvious, when people were yelling at him for proposing it in 2008.

Persistent misconceptions frustrate him: lean does not mean cheap, experimentation does not foreclose vision, and MVP is not about low quality — it is the most efficient test of whatever hypothesis you are betting on. At IMVU an anchored-avatar hack made customers call the product "more advanced than The Sims" because teleportation beat watching a character walk. Quality is what the customer believes. Cut the feature list in half, then in half again; most teams overshoot by one to two orders of magnitude.

Vision evolves through building. Ries found a whiteboard from an early company meeting and could not recognize his own handwriting as his own position — memory rewrites itself to erase pivots. One foot stays anchored in what was learned; one foot moves. Asking whether to pivot means you already know; product-market fit leaves no time for the question.

AI is a management technology that will collapse the summarization hierarchy making middle management necessary, but outputs inherit whatever misaligned values the deploying organization has.

His current preoccupation is governance: embedding mission into legal structure before IPO rails make change impossible. Foundation-controlled companies in Denmark empirically outperform matched public peers. Standard Delaware incorporation may legally require selling to Philip Morris if they outbid the market. Tools exist — public benefit corp status, mission pledges, LTSPV instruments — but lawyers advise "later," and later never arrives.

When they hired their first PM

Most startups wait two to three years and 10–15 engineers before hiring their first PM — but more than half hired before finding product-market fit. No founder regretted it; Notion regretted waiting, Stripe was glad it did by hiring product-minded engineers.

The trigger is whether the founder wants PM work, is good at it, and what they'd accomplish freed from it. At Lyft, handing off Zimride let founders conceive a new ridesharing product. The dominant mistake is hiring a PM to discover direction — Gusto's co-founder: bring one in once strategy is clear, not to find it.

Trust is the second failure mode. Kenneth Berger of Slack: "trust is given" beats "trust is earned" — low-trust environments fire over inevitable early mistakes; high-trust ones coach through them.

Internal transfers dominated: Notion, Looker, Figma, Robinhood, and Segment all promoted from within. Looker's first PM was an analyst writing post-meeting notes on what prospects needed; Robinhood's was an iOS engineer who knew the product cold. In B2C, a PM often disempowers design and engineering — the case is strongest in B2B, where relationships are multi-stakeholder.

Billion dollar failures, and billion dollar success | Tom Conrad (Quibi, Pandora, Pets.com, Snap, Zero)

Every company is a math equation converting investment into returns — no product iteration rescues a broken one. Tom Conrad, who took Pandora from 10 employees to 80 million users as CTO, used this lens to explain why Quibi failed despite $2 billion raised and 70 shows in 18 months.

Quibi's equation required landing in the App Store top 10 from day one and staying permanently — a near-impossibility Conrad admits he should have flagged harder. The content library needed $6–10 billion to reach scale. Covid then killed the planned live-studio daily content meant to differentiate from YouTube, forcing hosts to shoot from garages and producing something indistinguishable from the competition.

Pets.com followed the same pattern: three overfunded pet e-commerce sites entered a broadcast advertising arms race that destroyed all three. The model was never unsound — Chewy later reached $9 billion — but 80% of the US was still on dial-up.

Pandora's growth came from the inverse: zero paid acquisition, growing on word-of-mouth seeded by routing all support to all@pandora.com, where users might get a reply from the engineer who built the feature they complained about. The strategic error was betting on displacing terrestrial radio ($30B) while dismissing on-demand streaming as smaller — leaving Pandora locked into a statutory licensing regime the labels hated while Spotify negotiated direct deals.

At Zero, a fasting and metabolic health app (~1M monthly users, ~$10M ARR), Conrad models LTV against CAC as CEO and resisted pandemic-era pressure to chase paid acquisition — vindicated as markets shifted toward capital efficiency.

His contrarian view: the industry has too many founders. Many people who raise a $2M seed would achieve more — financially, culturally, by peer recognition — by joining a team with sound fundamentals rather than grinding an equation that was never going to win.

Jason Fried challenges your thinking on fundraising, goals, growth, and more

Bootstrapping is not the contrarian path — it is the default for nearly every business on earth. The venture-backed unicorn route is the statistical outlier; 37signals makes this visible by running against Silicon Valley narrative. After 24 consecutive profitable years the company produces double-digit millions in annual profit with 75 employees and 100,000+ customers — while Asana runs 1,600 people, Slack 2,500, Monday 1,500.

The mechanism: no enterprise clients, one Basecamp codebase with a $299 price ceiling for unlimited users, and every feature built by exactly two people (one designer, one programmer) capped at six weeks. That ceiling is an appetite, not an estimate — a fixed budget that forces the simplest effective solution. Work that stays on the "uphill" side at deadline dies; work that crests the hill executes downhill. After each cycle, a two-week cool-down replaces back-to-back sprints. Full methodology is free at basecamp.com/shapeup.

Fried rejects OKRs, revenue targets, and growth goals. The only financial test: did we make more than we spent? Decision-making runs on gut and intuition — he argues human judgment is irreducibly amorphous even when dressed as data. Designers are evaluated mid-critique with "if you had two more days, what would you do?" — not to score ideas but to sense whether instincts are loaded and active.

The new ONCE product line (starting with a relaunched Campfire chat tool) challenges the subscription-forever model: pay once, download the source code, self-host, receive 1.x updates free. It targets commodity software where dozens of alternatives exist but all charge recurring luxury prices. A planned book will make the case for gut and intuition as underrated business tools. The animating principle throughout: independence above all — no investors means no permission required, and nobody ever went broke making a profit.

Adding a work trial to your interview process

Interviews are a poor proxy for job performance — charm and presentation dominate over real ability. Work trials fix this by having candidates complete tasks representative of the actual role. Six companies that practice this extensively (Linear, Automattic, 37signals, Gumroad, Auth0, PostHog) report that trial performance frequently surprises relative to interview impressions.

Implementations span a wide spectrum. Gumroad goes furthest: candidates become paid contractors for 4–6 weeks at $150–$200/hour, working on real roadmap items in the live codebase. Linear runs 3–5 day in-codebase trials with a kickoff, candidate-led check-in, and final presentation; some projects ship. 37signals gives designers 7 days, requires working HTML/CSS rather than Figma ("we don't ship Figma"), keeps prompts vague to reveal prioritization, and pays $1,500. PostHog's SuperDays run 1–3 days with a task deliberately too large to finish — revealing prioritization judgment — plus coffee chats. Automattic allots 40 hours measured in hours not days (to avoid penalizing constrained schedules), using synthetic projects and a dedicated Trial Lead. Auth0 formerly reused real engineering problems across seniority levels; post-Okta, they switched to 90-minute CodeSignal tests, explicitly trading signal for a wider candidate funnel.

All six pay: Linear $600/day, 37signals $1,500/week, Automattic $25/hour, PostHog at the candidate's day rate.

Evaluation is consistently blind individual feedback, then group debrief, one final decider. Both Linear and PostHog treat anything short of a strong yes as a no; PostHog holds that a majority of soft yeses is still a no without a passionate advocate.

Candidate buy-in is rarely the obstacle expected. Payment helps; framing them as two-way — candidates see real tools and culture beforehand — handles the rest.

Key lessons: reserve trials for candidates you'd already lean toward hiring; keep tasks open-ended to surface judgment; accept longer hiring timelines — Automattic averages over 70 days from first contact to offer.

Zigging vs. zagging: How HubSpot built a $30B company | Dharmesh Shah (co-founder/CTO)

HubSpot's $30B outcome rests on "high conviction, low consensus" bets most advisors argued against. The clearest: ignoring the rule to do one thing well. In year one, HubSpot built SEO, blogging, CMS, and web analytics simultaneously because SMBs didn't lack tools — they lacked the ability to wire them together. Internal discipline: if any category ranked top-three in its market, they'd over-invested. The value proposition was integration, not depth.

The SMB bet itself was contrarian. Enterprise imposes long sales cycles and customers who dictate roadmaps; consumer outcomes are bimodal. SMBs offer consumer-scale feedback loops with measurable ARR. "Reverse gravity" — better margins from larger customers pulling software companies upmarket — requires active resistance; staying SMB means competing only with people disciplined enough to do the same.

Dharmesh has had zero direct reports across 7,000 employees, agreed at founding. Management was a skill he didn't want years becoming "passively okay" at. He gets scale's upside — big bets, long horizons — without the overhead.

Culture is a second product, built for the team the way the customer product is built for buyers. It gets quarterly NPS surveys, bugs triaged at all-hands, and iteration rather than preservation — the founder's job is not to preserve culture any more than to freeze a product's codebase. Post-IPO, every employee became a designated financial insider — no legal cap on that number — keeping full financials transparent at 7,000 people.

The "flashtag" system resolves the founder megaphone problem: #FYI (no response expected), #suggestion, #recommendation (response appreciated), #plea (deepest conviction, still not a mandate). HubSpot has no mandates.

On AI: the internet gave distribution at scale; AI gives cognition at scale — moving from imperative interfaces (step-by-step clicks) to declarative ones (describe the outcome, software finds the path). Dharmesh built GrowthBot on this idea in 2017; it failed because the models weren't ready. ChatSpot is the same bet with technology that now works.

Lessons from 1,000+ YC startups: Resilience, tar pit ideas, pivoting, more | Dalton Caldwell (Y Combinator, Managing Director)

Founders die from losing hope, not running out of money. The real cause of death is a co-founder fight, exhaustion of ideas, and a quiet decision to stop while cash remains. Irrational persistence is the antidote: Airbnb should have shut down before YC; the two objectively worst companies in Caldwell's Winter-17 group — a VR headset startup and a UK Venmo clone — pivoted to become Brex and Retool.

The rule for continuing: do you still enjoy this and love your customers? If yes, stay. If you dread the work and resent your co-founder, quitting is fine.

A good pivot moves toward expertise you already have. Brex went to fintech because the founders had built fintech in Brazil; Retool to internal tools because they'd built dashboards running Cashew; Segment's founders discovered that the data-routing layer under their analytics product was the actual product. The pivot signal: you've run out of good growth ideas. When the best remaining move is "maybe pay influencers," change the idea.

Tarpit ideas trap founders because they generate genuine enthusiasm — social coordination apps, music discovery, location layers — while failing consistently since the 1990s. Validation is easy; traction never comes.

The early move is 20–30% of your calendar in in-person meetings. Stripe's Patrick Collison drove to customer offices to install Stripe himself; Zip ran hundreds of LinkedIn conversations before writing code. Pre-sell before you build. A/B testing with no users is actively harmful — that advice is calibrated for companies with scale.

On investor rejections: investors take a few bets per year and need the deal to feel like the one. On TAM: it barely matters at pre-seed — Razorpay's market was negligible in 2015 India. On founder type: breakout founders share no personality, only an internal conviction strong enough to warp reality around them.

This will make you a better decision-maker | Annie Duke (author of “Thinking in Bets” and “Quit,” former pro poker player)

Converting implicit intuitions into explicit structures is the core move behind every decision improvement. Intuition is sometimes right, sometimes wrong — you can only find out which when it is on paper.

The highest-leverage group change is separating discover, discuss, and decide. Opinions should be collected independently before any meeting — written prompts, forced rankings, private forms — so the loudest voice doesn't collapse the true spread of views. Meetings are only for discussing disagreements. Alignment is a false goal: people genuinely don't agree, and expecting it makes meetings coercive. Reflecting opinions back without endorsing them means everyone feels heard even when the decision goes against them.

There is no such thing as a long feedback loop — only a choice to keep it long. At First Round Capital, Duke introduced explicit pre-investment forecasts (Series A probability, 1–7 ratings on market, team, product). Over five years this revealed which partners are predictive about which dimensions, and whether factors they pound the table about correlate with outcomes. Sometimes the confident intuition is right; sometimes it predicts nothing. Without explicit records you cannot tell.

Pre-mortems are only useful paired with kill criteria. Imagining failure rarely changes the plan. What it produces is specific signals with attached actions: prospect only asks about price → kill; can't get a decision-maker in the room → offer executive alignment, and if they decline, kill. The value is eliminating in-the-moment rationalization.

By the time you are thinking about quitting, you have usually passed the right moment. Sunk cost, identity, and the need for certainty all conspire to extend bad bets. Stewart Butterfield shut down Glitch in 2012 — critically acclaimed, $6M in the bank — once math showed customer acquisition would never reach venture scale. The internal tool his team had built became Slack. Continuing Glitch would have foreclosed it.

The art of the pivot, part 2: How, why, and when to pivot

Nearly two-thirds of failed founders attempted a pivot before shutting down — pivoting is no guarantee. The case for it is opportunity cost: more pivots mean more shots at product-market fit, and it's easier to be lucky with six rolls of the dice than one.

Two categories matter and should not be conflated. Ideation pivots happen before meaningful traction — Brex went from VR to business banking, YouTube from dating site to video platform, both in under three months. Hard pivots happen when a live product with real users changes direction, keeping one element: Instagram stripped everything except photos, Slack doubled down on its internal IRC tool, Loom on screen recording.

Timing: ideation pivots cluster within three months of launch; hard pivots cluster around the one-year mark, rarely beyond two.

Two signals justify a pivot: persistent lukewarm engagement (low retention, plateau — Systrom watched users bounce off Burbn, Loom made $600 in seven months), and realizing the market is smaller than assumed (Slack's Glitch had burned $17M with 45 people and would never scale).

Four paths cover most successful pivots. Go all-in on one feature pulling disproportionate use — Yelp found users spontaneously writing multiple reviews, Pinterest users building collections inside a shopping app, Instagram users sharing photos inside a check-in app. Commercialize something built internally — Hugging Face's weekend BERT port went viral on GitHub, Segment's Analytics.js topped Hacker News on launch day. Chase an adjacent bigger market — Box followed enterprise use cases, Twitch noticed gaming streams were the only content Emmett Shear enjoyed. Brainstorm entirely fresh: Twitter and Lyft came from hackathons, Brex from noticing YC companies couldn't get credit cards.

The stay-or-pivot test: "Knowing everything you know now, do you have more or less conviction?" Eroded conviction means quit. Increased conviction despite failure means keep going.

Lessons from a two-time unicorn builder, 50-time startup advisor, and 20-time company board member | Uri Levine (co-founder of Waze)

Fall in love with the problem, not the solution — Uri Levine's core principle after co-founding 10 companies (including Waze, sold to Google for over $1 billion) and advising 50+ startups. The problem anchors pivots, makes your story compelling, and pulls customers into wanting you to succeed. "AI crowd-sourced navigation" made eyes glaze over; "help you avoid traffic jams" made people care. Validate with 20 strangers — if they correct your description, follow that.

Product-market fit has one metric: retention. Waze took four years; Microsoft five; Netflix ten; ChatGPT was seven years old before anyone heard of it. Iterate fast, preserve shots at the basket. Once PMF is found the product freezes and the company shifts to growth and business model — never both at once. High-frequency products chase growth first (word-of-mouth compounds daily); low-frequency products must crack the business model first.

On fundraising: investors decide before you sit down. Put your strongest point on the first slide — it's displayed longest — and repeat it on the thank-you slide. Expect 100 nos before one yes; VCs invest in 1–2 of 200 companies seen per year. Don't argue rejections. Tell detailed stories — specificity creates believability even when the story is constructed.

On hiring and firing: failed-startup founders knew the team was wrong within the first month but did nothing. Top performers leave when hard calls aren't made. Rule: 30-day calendar reminder for every hire — ask "would I hire this person knowing what I know now?" Yes earns them more equity; no means fire immediately. The same logic applies to any life decision.

To understand users, watch people who churned, not your loyal base. Most product builders are innovators; winning the market means winning the early majority — people afraid to look stupid. Simplicity is the only bridge.

Hard-won lessons building 0 to 1 inside Atlassian | Tanguy Crusson (Head of Jira Product Discovery)

Large companies fail at zero-to-one not from lack of resources but from inability to simulate scarcity. Tanguy Crusson spent a decade at Atlassian — 300,000 customers, flat decision-making, direct analyst access — and still ran 50/50 before Jira Product Discovery.

Three case studies expose the failure modes. HipChat died from two errors: the team assumed Atlassian's developer-community playbook would transfer to business users, and when Slack surged they fell into competitive myopia, fast-following features instead of doubling down on devtool integrations. Crusson now watches competitors quarterly, then replays user interviews to re-anchor. The rewrite into Stride — rebuilding the platform mid-service — was the core mistake: rebuilding mid-flight kills you; delay takeoff instead. Statuspage taught that acquisitions are a people problem — founders who join lose decision authority while OKRs, three-year roadmaps, and craft rituals arrive simultaneously. Plan for slowdown before re-acceleration.

Jira Product Discovery succeeded because Atlassian created Point A, an internal incubator with four stages: Wonder, Explore, Make, Impact. Gates required six-pager reviews with founders present. The shared vocabulary protected the team — "we're in Explore" told stakeholders not to question architecture yet. Crusson acted as both product and engineering lead, hired contractors to bypass headcount rules, and operated from France where the timezone gap kept Atlassian processes out.

The Lighthouse Users Program staged early customers at 10, then 100, then 1,000, with qualitative milestones dominating the first two phases. Engineers attended every Zoom — a team that knows ten customers by name builds differently than one reading CSAT scores. Crusson published weekly Atlas updates mixing data, demos, and three-minute user-clip videos: no one wants to stop a high-speed train. Five months to first alpha customer, six months of alpha, a year of beta, then GA. JPD reached 8,000 customers and became one of Atlassian's fastest-growing products.

The social radar: Y Combinator’s secret weapon | Jessica Livingston (co-founder of Y Combinator, author, podcast host)

Picking the right founders matters more than picking the right ideas — Jessica Livingston's edge at Y Combinator was reading founders better than her co-founders because she wasn't distracted by the technology. While Paul Graham, Robert Morris, and Trevor Blackwell interrogated the product in 10-minute interviews, Livingston watched the room. Her nickname "Social Radar" stuck because everyone turned to her afterward.

The signals she tracked: co-founders who interrupted each other or held lopsided equity splits were flags. Defensiveness was always a bad sign — the best founders treat hard questions like a tennis match, staying open to what users actually want (PayPal's pivot from PalmPilot to web is her example). She prized earnestness — authentic connection to the problem — and flagged its absence: 45-year-old men building a teen fashion app were passed because they clearly chased a trend rather than cared. Scrappiness mattered too: Airbnb was funded despite YC hating the air-mattress idea because Joe Gebbia pulled out hand-glue-gunned Obama O's and Cap'n McCain's cereal boxes, signaling they'd do whatever it took. She championed GOAT's founders over the team's skepticism; they pivoted from a failing group-dinner concept to build a major sneaker resale business.

YC's batch structure was an accident of inexperience: the co-founders ran a summer cohort to learn angel investing, then kept the model when they saw the community effect. That first batch included Sam Altman, Emmett Shear, and the Reddit founders.

Livingston scored 36/36 on the "Reading the Mind in the Eyes" test — eye-region photos where you name the emotion. Her technique: ask what those eyes are trying to tell you.

She hosts The Social Radars podcast. The Parker Conrad episode stands out: she surfaced a paid smear campaign that drove him from Zenefits — never publicly corrected by any reporter.

Summary: Lessons from working with 600+ YC startups | Gustaf Alströmer (Y Combinator, Airbnb)

Startups fail almost exclusively because founders stop talking to users. Gustaf Alströmer (YC Group Partner, Airbnb growth founder) identifies not-finding-PMF as the root cause of nearly every failure he has seen across 600+ companies, driven by fear of rejection and confusion between investor praise and actual customer validation. Successful founders are determined, technically capable, and make weekly progress — YC's clearest leading indicator. Pre-PMF, execution beats strategy; speed and quality are a false trade-off when you genuinely know what customers want.

Raising a seed round 101

Seed funding makes sense only if you want massive scale, can accept 10–20% dilution, and have concrete reasons outside capital beats bootstrapping. Jason Fried warns that raising too early trains founders to spend rather than earn — revenue-making needs practice from day one.

Before pitching, show three things: full commitment (no day job), proof of work (Gusto's Tomer London recommends 10–100+ buyer conversations targeting 40%+ "when can I get this?" responses), and a written thesis or early paying customers. Conviction across all three is the most predictive variable.

Target 24–36 months of runway with a 25% buffer. Typical rounds are $2–4M at ~$20M post-money, selling ~15% equity; Carta puts the median seed-to-Series-A gap at 23 months. Reach Series A under 40% dilution — partner quality matters more than price.

The tactical core is manufactured FOMO: compress all investor meetings into two to three weeks, open a SAFE early to bank angel checks and build social proof, then use that momentum with larger funds. Warm intros from portfolio founders carry the most weight; intros from investors who passed send a negative signal. Never name which investors gave term sheets — describe them abstractly. Verbal commits are not term sheets; a round isn't closed until wires clear.

A raise going well feels like pull: quick follow-ups, specific numbers, clear next steps. Confusion means no — ask "are you interested?" rather than burning time.

SAFEs dominate at seed (fast, no lawyers). The only terms that matter are post-money valuation and board composition — resist director seats, offer observer seats. Choose investors as you'd hire someone you can never fire. Announce only if it addresses a current top-three challenge.

Which companies accelerate PM careers most

Data from 160 million career histories reveals which companies most accelerate PM careers across six metrics: post-departure promotions, speed to leadership, CPO/Head-of-Product rate, first-PM-at-a-startup rate, and founder rate. Revolut and N26 — both European neobanks — top the overall ranking. Fintech accounts for five of the top ten. Palantir stands out on founder rate at ~25% of alumni; Plaid follows at ~20%. eBay outranks every FAANG company. FAANG underperforms because their PMs learn company-specific operating skills that don't transfer. Notable mentions: Ramp, Notion, Discord, and Intercom.

Pulling back the curtain on the magic of Y Combinator

YC's edge lies in founder selection, not in repeating a winning company profile — the top companies of each era look nothing like the next. Across 4,939 companies (2005–2024), YC beats benchmarks decisively: 45% reach Series A vs. 33% average; 4–5% become unicorns vs. 2.5%; over 50% survive past 10 years; 25–50% of mature batches get acquired vs. 15% broadly.

Returns concentrate sharply: Airbnb, DoorDash, Coinbase, and Instacart — all backed in the Paul Graham era — account for 84% of $300B+ in realized market cap. Future exits will come from B2B and fintech: Stripe ($70B), Rippling, Brex, and Scale AI lead the private candidates.

Recent batches have shifted hard: 75%+ are now B2B, reflecting a view that large consumer apps are already built. Geography is concentrated — 70%+ of companies are U.S.-founded, and 99% of returns come from the U.S. Solo founders have declined to under 10% of recent cohorts. The current bet is B2B AI, led by Engineering/Product/Design (25%), Infrastructure (12%), and Sales (8%).

Notion’s lost years, its near collapse during Covid, staying small to move fast, the joy and suffering of building horizontal, more | Ivan Zhao (CEO and co-founder)

Notion's first four years failed because Ivan Zhao built what he wanted rather than what people needed. The original product was a developer tool letting anyone create software — almost nobody cared. The pivot was conceptual: hide the vision inside a form factor people already use. They call it "sugar-coated broccoli." The sugar is a productivity suite; the broccoli is a no-code platform where users construct tools from composable Lego blocks. That realization took two years. A third year was lost rebuilding on a stable technical foundation after betting on Web Components over React. The company shrank to two people, moved to Japan, coded 18-hour days, and survived on money borrowed from Ivan's mother.

During COVID, growth nearly killed the product. Notion ran on a single PostgreSQL instance; as remote-work adoption spiked, they had weeks before hitting the ceiling. All feature work stopped while every engineer sharded Postgres.

The lean-team philosophy — no salesperson until $10M ARR, first PM at 50 people — rests on a belief that abstractions outperform headcount. Notion tracks revenue-per-employee rather than size. The internal metaphor: a small bus corners faster.

The biggest self-inflicted mistake was hard-coding a sprints feature to compete in project management rather than building sprints from Lego primitives. It fit poorly with the rest of Notion and took over a year to reverse. The principle: build the Lego way and the system works for you; hard-code and it works against you.

On horizontal growth, a B2C2B flywheel drives scale — personal note-takers bring Notion into their companies. Calendar followed docs as the next billion-user wedge; email is next. AI sharpens the advantage: bundled knowledge in one place is exactly the substrate LLM reasoning works best on, and coding agents can now assemble Lego blocks into custom vertical software without users writing code.

Inside the expert network training every frontier AI model | Garrett Lord (Handshake CEO)

Handshake — a LinkedIn-for-college-students platform at $200M ARR — launched a human data business for AI labs in January 2025, hit $50M ARR in four months, and is on pace to exceed $100M in year one.

The business rests on one structural insight: pre-training gains have plateaued because labs exhausted the publicly available internet corpus. Improvements now come almost entirely from post-training — supervised fine-tuning, RLHF, preference ranking, and trajectory data (screen recordings with narrated step-by-step tool use). The generalist crowd-labor market Scale AI pioneered no longer serves the frontier; current models already outperform generalists, so labs need domain experts who can break the models and supply ground-truth corrections with step-by-step reasoning traces.

Handshake's moat is audience access: 500,000 PhDs and 3 million master's students across 1,600 partner universities, acquired at zero marginal cost, with brand trust that drives high conversion. Competitors spend tens of millions a month on LinkedIn outreach and performance ads to find physics PhDs who have never heard of them; Handshake messages people it already knows. PhDs earn $100–$200/hour finding model failures, verifying reasoning steps, supplying rubrics for non-verifiable domains like educational design, and producing multimodal audio data.

Quality, volume, and speed are what labs care about — in that order. Handshake built its own post-training team (including a Meta post-training hire), rents GPUs to evaluate data quality internally, and sells pre-built data packages to multiple labs simultaneously.

Organizationally, Lord ran Handshake AI as a fully separate entity inside the company: dedicated engineering, design, finance, and recruiting; separate all-hands; 80%+ of his own time; metrics cadence from day one; equity tied to milestones. The core business ran under the existing executive team. Once a hypothesis showed model-quality gains, resources flooded that pipeline while others were dropped — the same iteration loop the labs use.

$46B of hard truths from Ben Horowitz: Why founders fail and why you need to run toward fear (a16z co-founder)

The single most destructive thing a leader does is hesitate — and hesitation is always caused by both options being bad. Leadership adds value only at exactly those moments. Horowitz went public with Loudcloud at 18 months old on $2 million of trailing revenue in March 2001 not because it was a good idea but because bankruptcy was worse. Most competitors hesitated and went bankrupt.

Running toward fear is the required muscle. He can coach the specific conversation — reframing "your CTO is an asshole" into actionable feedback about cross-functional effectiveness — but the willingness to act when both options are awful must be built through experience. Jensen Huang developed it over decades; instant-hit founders typically haven't.

Founder failure follows a consistent pattern: expensive mistakes erode confidence; confidence loss produces hesitation; hesitation creates a power vacuum; senior executives fight to fill it; the company goes political. a16z responds structurally — giving new CEOs a Fortune 500-level network from day one and peer-level board relationships.

Hire for strength, not absence of weakness. A CEO who doesn't know sales cannot develop a PhD into a sales leader. The Adam Neumann reinvestment: WeWork built the most recognizable commercial real estate brand in history; the failure was inexperience and no one willing to tell him the truth.

On AI: current valuations are supported by real revenue at unprecedented growth rates; the dot-com bubble was defined by unit economics that never worked. The "thin wrapper around a foundation model" dismissal repeats the 1980s "thin wrapper around a RDBMS" mistake that would have ruled out Salesforce. Cursor has built 14 proprietary models on top of the foundation layer. Application-layer opportunities are large and stickier than assumed. Foundation model competition requires raising at least $2 billion on scientific credibility alone — fewer than ten founders globally qualify.

Why experts writing AI evals is creating the fastest-growing companies in history | Brendan Foody (CEO of Mercor)

The primary bottleneck to improving AI models is not compute or data — it is the ability to measure what good looks like. Evals (rubrics and verifiers defining success for a capability) are the rate-limiting input to post-training and reinforcement learning. Once a lab has an eval, RL climbs it rapidly: Olympiad Math and SWE-bench were saturated quickly once researchers focused on them. Every capability gap — contract redlining, medical diagnosis, long-horizon tasks — is blocked not by architecture but by the absence of an expert-written eval.

Mercor, founded by Brendan Foody at 19, pivoted in mid-2024 to supplying expert labor to AI labs for post-training. They grew from $1M to $400M ARR in 16 months by recognizing that labs had moved past cheap generalist annotators and need credentialed professionals: Goldman bankers, FAANG engineers, radiologists, Emmy-winning screenwriters, Harvard Lampoon writers (to improve model humor). Median pay is $95/hour; specialists reach $500/hour. In any batch of 100 experts, the top 10% drive most of the model improvement — Mercor's moat is identifying and delivering that top decile within 24 hours of a customer request.

The market thesis: the economy is becoming an eval-environment machine, with humans defining success criteria across every domain AI cannot yet handle. Foody disputes near-term AGI timelines — progress requires high-quality post-training data, not more pre-training scale — and expects a decade-plus runway.

On careers: jobs with elastic demand (software, product, consulting) expand as AI multiplies productivity. The durable skill is leveraging AI within your domain; Mercor's interviews ask candidates to build a product in one hour using any AI tools.

Mercor ran zero sales or marketing for 18 months, growing on pull. Three core values: can-do attitude (setting absurd targets and hitting them), high standards (former Scale head of growth as employee two), and output-oriented intensity.

The secret to better AI prototypes: Why Tinder's CPO starts with JSON, not design | Ravi Mehta (product advisor, previously EIR at Reforge)

Startups have lower latency than big companies, not higher velocity. Large companies out-execute on volume; a startup can go from idea to validated learning in a day. The implication: drop A/B experiments without sample size; shift to conviction-based decisions, execute, then pivot fast.

Ravi Mehta's product strategy stack separates five concepts teams conflate: company mission, company strategy, product strategy, roadmap, and goals — in that order, goals last. Most teams start from a goal ("grow retention 20%") and build backward; without a destination first, hitting a metric can undermine the strategy. Tripadvisor teams optimized per-visit bookings, which kept sending users back to Google rather than building the direct relationship the strategy required.

Goal type should match what the team knows. Teams that don't understand what moves a metric should set understanding goals, not outcome goals — then dependency goals (do we have the tools?), execution goals, and finally strategic goals. Committing to a retention target at the understanding stage is spaghetti-throwing.

Strategy documents should include wireframes — not final UI, but enough to force real trade-offs. A four-slot mobile nav forces choices that prose cannot. Balsamiq for an afternoon beats a 10-page spec.

The 12-competency PM framework, built at Tripadvisor for a rotational hire program, clusters into four areas: product execution (specs, delivery, quality), customer insight (data fluency, voice of customer, UX design), product strategy (business outcomes, roadmapping, strategic impact), and leadership (stakeholder inclusion, team leadership, managing up). Exponential feedback — compounding returns — comes from identifying root-cause competency gaps, not surface symptoms.

On leadership: selective micromanagement is the correct response when a team heads in the wrong direction. The two actual failure modes are hands-off passivity that lets a team drift, and chronic micromanagement with no framework transfer. Effective senior leaders expand dynamic range rather than staying at altitude.

The woman behind Canva shares how she built a $42B company from nothing | Melanie Perkins

Canva reached $42B and $3.3B ARR by refusing to plan from present resources. Melanie Perkins calls this Column B thinking: start from the future you want, not the bricks you have. As a university student in Australia with no software experience, Column A would have produced nothing. Column B was the conviction that design would move online, collaborative, and simple — so she started with Fusion Books (school yearbooks), then built Canva.

Over 100 investors rejected the pitch, each rejection sharpening it: objections about market size got a market-size slide; objections about differentiation got a competitive-gap slide. The vision barely changed; only the articulation did. Problem-first framing was the decisive shift.

Canva operationalizes Column B through a mission decomposed into pillars — design anything, in every language, on every device — each carrying "Crazy Big Goals." Timing estimates routinely fail (a six-month front-end rewrite took two years with nothing shipped), but the goals land eventually. A 2021 deck for 2026 had largely materialized by recording date. Goals couple with ritualized celebrations — plate-smashing, dove releases, La Tomatina — so teams mark each rung climbed before resuming.

Product direction flows from two inputs: the mission pillars and over a million community feature requests per year, with 200+ loops closed in 2025. Perkins has personally run hundreds of remote user tests; participants are candid alone with a camera.

The philanthropy is structural. Perkins and co-founder Cliff Obrecht committed over 30% of their Canva equity to the Canva Foundation. $50M has gone to GiveDirectly for direct cash transfers in Malawi; another $100M is pledged over four years. The education product reaches 100M monthly users and is free to schools. Wealth, Perkins says, is not a goal — it is the mechanism for step two of the two-step plan: do the most good possible.

How to spot a top 1% startup early

People who joined Palantir, OpenAI, Stripe, Figma, Slack, Spotify, and similar companies early—as full-time employees, not investors—share three signals they acted on. First, ambition that sounds ludicrous: Spotify's Daniel Ek was literally laughed at by record-label executives; Dylan Field's browser-based Figma seemed "completely insane." If businesspeople mock it while users call it the obvious future, pay attention. Second, founders above everything—specifically learning speed ("clock speed"), ferocity, and founder-market fit; Patrick and John Collison's books-stacked-to-the-ceiling curiosity was more predictive than Stripe's early metrics. Third, ignore the current product: early Facebook disappointed, early Slack was "a giant piece of shit." Look instead for organic customer pull and a "Jurassic Park moment"—a demo so alive you sense an apex predator entering the world.

Sequoia CEO coach: Why it’s never been easier to start a company, and never been harder to scale one | Brian Halligan (co-founder, HubSpot)

Starting a company has never been easier — AI and cloud collapsed setup costs — but scaling one has never been harder because those forces flood every market, making distribution the decisive constraint. Halligan co-founded HubSpot in 2006, ran it as CEO for fifteen years, and now coaches Sequoia founders.

Adult-table CEOs (100+ employees) spend half their time recruiting. Best blind-reference question: "would you enthusiastically rehire this person?" Parker Conrad sends candidates the last board deck and treats pure flattery as a red flag. Hire spiky candidates over those with fewest weaknesses; avoid big-company transplants — near-100% attrition on Salesforce and Google hires at HubSpot. Build like the 2004 Red Sox: homegrown core plus a few free agents.

Halligan evaluates founders on LOCK-S: Lovable, Obsessed, Chip on shoulder, Knowledgeable, Student. A newer archetype is the five-tool CEO — can code, has taste and vision, can sell, can inspire; Brett Taylor is the prototype.

Key Halliganisms: don't nibble a shit sandwich — incremental layoffs compound. Never waste a good crisis — HubSpot's 2019 outage became a full overhaul of deployment. One DRI per initiative; shared ownership kills execution. EV > TV > MEV: tie exec comp to retention and NPS, not revenue.

HubSpot over-indexed on being a great place to work (Glassdoor's #1 company). Shifting toward customers required customer panels in every board meeting. Planning cycles have compressed from annual to quarterly; optionality now carries a real tax.

Go-to-market will be inverted: buyers research inside LLMs, making AEO more important than SEO. Enterprise sales — trust between humans — is the last white-collar function AI will displace.

A 2022 snowmobile accident left him with 20 broken bones and 33 screws. Lying in a Vermont ravine, he decided he didn't want to run an 8,000-person company and handed the role to Rangan.

How to know if you've got product-market fit

PMF has concrete signals at every stage. Pre-product: pupils dilate during demos; customers send unsolicited love letters; people pay before the product exists. Post-product: retention cohorts flatten rather than decay to zero. More than 40% of surveyed users say they'd be "very disappointed" if the product disappeared. Enterprise trials end in screaming when pulled. Growth is organic and explosive — over 50% of new accounts from direct traffic. Burn multiple stays low (market pulling product, not team pushing it). Sales yield exceeds 1.0. The clearest sign: people keep using a broken product anyway.

Joining as the first product manager

The first PM role is uniquely hard because founders are reluctant to hand over their product. Hire only after product-market fit with 7+ engineers (Jonathan Golden, first PM at Airbnb); signals include the team saying "you're slowing us down." Once in, spend the first month listening rather than fixing — moving fast to show value destroys trust. Within three months, know every product flow and competitor feature better than anyone. Build CEO trust incrementally over 90 days via frequent vision debates. Prove value to engineers by shielding them from ad-hoc roadmap asks. Align explicitly on what "product management" means — many founders haven't defined it. The ongoing tension is quick wins versus customer-empathy depth.

What is good retention?

Retention benchmarks vary dramatically by business type, and most founders are working off anecdotes. Surveying 20 growth practitioners (Casey Winters, Andrew Chen, Brian Balfour, Elena Verna, and others), the concrete thresholds at six months are: consumer social 25%/45%, consumer transactional 30%/50%, consumer SaaS 40%/70%, SMB SaaS 60%/80%, enterprise SaaS 75%/90%. Net revenue retention targets differ by customer segment — bottom-up SaaS (Slack, Zoom, Figma) should hit 100%/120%, enterprise 110%/130%. Public comps confirm the ceiling: Twilio and Zoom ran 140%+ NRR; Workday held 95% user retention. Low retention is acceptable only when CAC is near-zero or the business isn't venture-scale — otherwise it is the single most diagnostic metric for whether a business will survive.

How to increase your product's retention

Retention is the single most important early growth metric; Brian Balfour: "if you have poor retention, nothing else matters." Measure it via cohort curves, not aggregate churn — a flattening curve signals product-market fit.

Seven levers ranked by impact: (1) improve the product — solve problems 10x better, expand scope, or wait for network effects; (2) improve onboarding — Andrew Chen's finding is that connecting users to existing value moves retention more than adding features; Superhuman's 1:1 onboarding and Pinterest's "core product fast, but not faster" illustrate this; (3) make it stickier via habits, annual plans, or deep integration like Slack or AWS; (4) catch users at cancellation — Airbnb's "pause listing" reduced host churn; (5) remind users of value; (6) retarget lapsed users; (7) change the user mix — paid traffic has lowest intent, and studying who already retains reveals the real ICP.

The most important bottom-up SaaS metrics to track

Retention comes before revenue: for early-stage bottom-up SaaS, user retention (% still active 3–6 months later) and intra-org virality (invite rate, invite conversion, virality factor) are the priority metrics even before any paying customers exist. Post-revenue, the hierarchy is MRR growth → customer retention → paid conversion and payback period. Net Dollar Retention — cohort MRR at 12 months — is the key expansion signal. Google Sheets dominates actual founder dashboards, with Profitwell and Google Data Studio as runners-up.

The most important consumer subscription metrics to track

Consumer subscription success depends on six sequential stages: sustainable acquisition (payback period under 6 months; virality above 1.0), activation, ongoing engagement, free-to-paid conversion, retention, and profitable delivery. The decisive metrics differ by type: software businesses should prioritize activation rate and L7/L30 engagement intensity; content businesses should watch cohort engagement; physical goods businesses need second-order retention (users who don't cancel after order one) and contribution margin. Cohort retention above 70% at six months is the universal paid-subscriber benchmark.

The most important consumer metrics to track

Consumer metrics are determined by monetization model, not by product category. Trial-based subscriptions live on conversion rate and cohort retention (M1/M3/M6) plus payback period. Freemium adds free-user retention and WOM-driven user growth before conversion matters. Ad products need DAU/MAU intensity and cheap CAC. Marketplaces track both buyer and supply-side retention separately, plus GMV. DTC hinges on gross margin, AOV, and ROAS. Across all five, narrow to the two or three levers that most directly move the business.

What is good monthly churn

Good monthly churn varies sharply by segment: B2C SaaS aims for 3–5% (great: below 2%); B2B SMB/Mid-Market 2.5–5% (great: below 1.5%); B2B Enterprise 1–2% (great: below 0.5%). Price point drives this — higher-priced products sell to stickier customers and unit economics collapse at elevated churn. ProfitWell data across 13,000 companies shows B2C variance is narrow (under 2% at any price); B2B diverges significantly between top-10% and top-25% performers. Early churn (months 1–3) reflects activation failure or mis-targeted acquisition, not steady state. For B2B SMB, net revenue retention matters more than logo churn, since expansion can offset losses. Higher churn is defensible only with very low CAC or a non-venture-scale ambition.

How to win in consumer subscription

Consumer subscription is brutal: most apps get tried briefly and abandoned. The companies that survive — Noom, Grammarly, Duolingo, Spotify, Calm, Future, Flo — share three patterns.

Obsession with efficiency. Grammarly bootstrapped and ran 2–4x leaner than peers. Noom's founders shared a single apartment for two years; targeting one-month payback periods let them reinvest into paid marketing without outside capital, reaching $60M ARR before significant hiring. Future's CEO credits a small footprint with surviving years of being "misunderstood" while iterating.

Product strategy locked to acquisition strategy. Grammarly's choice to embed everywhere (browser extension, mobile keyboard) directly informed funnel experimentation. Duolingo kept all learning content free to drive word-of-mouth at scale, converting only on premium features. Spotify's freemium hit 20%+ conversion when the industry benchmark was 7%. Post-iOS-14, paid social is harder, but App Store and Play Store cuts dropping to 15% partially offsets rising CAC.

Sticky product through rapid iteration. Noom ran up to six experiments per PM per week, targeting only 20–30% improvements to move fast. Future deliberately fired happy paying members between cohorts to isolate what drove anomalously high 3-month retention before scaling. Centered doubled web conversions by stripping the above-the-fold to a single CTA.

Tactically: push annual or multi-month plans (Noom, Calm); consumers anchor pricing to Netflix ($10–20/month); B2B2C distribution through employer wellness programs is underused; and not all churn needs solving — focus retention energy on best-fit users.

Merci Grace (ex-Head of Growth at Slack) on PLG, interviewing, storytelling, building a diverse team, hiring salespeople, building a growth team, and much more

Slack's product-led growth started as "new user experience," run by a game designer. Grace was hired because Stewart Butterfield wanted a game-design sensibility applied to onboarding: frame the product before building it, since how you introduce something determines whether users see its value at all.

The activation metric: 3 real humans, 50 real messages. The counterintuitive finding was that Slack's needle-moving experiments clustered around synchrony — getting people into a team at the same time. Jules Walter's push notification work confirmed it: being greeted when you join is what makes adoption stick.

PLG fitness: can an individual at any seniority adopt this without organizational gatekeepers? DevTools pass; anything requiring HR sign-off on PII doesn't. A second test is day-zero value — transcript apps that only pay off after months of data are as easy to abandon as to adopt.

The most common PLG mistake is the informational carousel — no one finishes them. Run a prototype bake-off against embedded tooltips; let the user test decide. On invites: ignore research pushback ("I'd never invite anyone before seeing the product") — social connectors exist in every user base; everyone else skips the prompt.

Hire a first salesperson when the founder is maxed out, or when enterprise customers expect one — many cannot purchase without a human contact.

For PM hiring, Grace assigns a take-home project (candidate picks from three problems, ~3 hours). She filters for technically grounded solutions and narrative structure. At Slack, story won over data — a proposal without a compelling arc didn't move regardless of numbers.

On diversity: proactively source; inbound referrals skew homogeneous. Two people from an underrepresented group shifts team tone and triggers its own referral flywheel; one outlier doesn't. On pitches: open with the unique founder insight, not market size. Backfill after you have attention.

How to kickstart and scale a consumer business—Step 5: RETAIN: Iterate until enough people stick around

Product-market fit is a confidence spectrum measured by four signals in decreasing reliability: cohort retention curves flattening above zero (six-month benchmarks: ~25% for consumer social, ~30% for transactional, ~40% for subscription); explosive organic word-of-mouth that can't be bought; more than 40% of users saying they'd be "very disappointed" without the product (Sean Ellis's threshold); and visceral desperation — panicked calls when servers go down, refusal to leave a broken product.

Across dozens of B2C companies, over 80% found PMF within two years, but fewer than 20% felt it immediately. The journey follows five archetypes: Lightning (Tinder, Instagram, Dropbox — immediate); Delayed Lightning (Netflix took 18 months before no-late-fees subscription unlocked growth; YouTube found virality through MySpace embeds; Discord via a Reddit post in the Final Fantasy subreddit); Foothold (Instacart clear with early adopters, unclear until retailer partnerships; Lyft obvious in SF, opaque in LA for months); Milestone (Duolingo's year-long private beta then sustained DAU at public launch; Substack growing until there was no time left to doubt); and Grind (Thumbtack five years building supply then demand then revenue; Coinbase two years until adding a buy button). A sixth archetype — Clubhouse, Vine — had explosive growth that collapsed when retention never materialized.

True PMF requires three fits: enough people want the product, you can profit, and you can acquire sustainably. Justin.tv had 30 million monthly viewers and $8M revenue but no repeatable growth lever — not PMF. Shyp had demand but no viable margin.

When PMF is absent, four situations cover most cases: wrong product (Coinbase added a buy button after one user call; Duolingo added streaks); wrong distribution (a single channel spark can unlock a working product); wrong onboarding (early churn signals activation failure, not product); wrong audience (Pinterest stalled until Silbermann switched from tech friends to female bloggers).

How to measure cohort retention

Accurate cohort retention requires resolving five sequential decisions, each of which silently corrupts the number if gotten wrong. First, define "active" as the main user action rather than logins or app opens — those overcount by pulling in unauthenticated sessions and background refreshes. Second, separate free users from paid customers; blending them hides true free-user engagement and masks conversion opportunity. Third, choose between X-day retention (returns on exactly day N — suits daily-habit products and SaaS trials) and unbounded retention (returns on day N or later — the inverse of churn, suits irregular-engagement consumer products). The two methods produce substantially different numbers from identical data. Fourth, tools like Amplitude and Mixpanel default to N-day and lack payment data unless piped in via Segment, making SQL the reliable path for SaaS; build a clean activity foundation table first, then join sign-up to subsequent activity and bin elapsed time into periods. Fifth, visualize with heat-mapped cohort grids — color scaling makes behavioral patterns across signup cohorts legible where line charts hide them. Retention is an output metric; using it as an A/B test target is a common mistake because activity is only one of its components.

Growth tactics, retention strategies, and becoming a better writer | Julian Shapiro (Demand Curve, Hyper, Webflow, TechCrunch)

Viral growth is most durable when the product's core function forces acquisition — not when referral incentives are bolted on. Julian Shapiro's product-led acquisition (PLA) framework identifies four mechanisms: settling debts (PayPal, Venmo — recipients create accounts to collect), joining locked conversations (WhatsApp, Slack — outsiders sign up to participate), billboarding (Hotmail's "sent via" signature, Calendly links, Teslas as walking ads), and UGC spreading off-platform (TikTok watermarks, Reddit surfacing on Google). Referral programs are weaker — they self-select for reward-seekers who churn. PLA compounds because cost is zero and the mechanism is in the product's DNA, not subject to ad-platform volatility.

Retention follows the same logic: the best lock-in is state users build and cannot export. Non-transferable reputation (eBay seller ratings), non-transferable audiences (YouTube subscriber graphs), and embedded infrastructure (Twilio, AWS) make switching costly proportional to investment. "Moat" is really just accruing state faster than competitors.

On writing, novelty is the primary variable: new + significant + non-obvious. Five types: counterintuitive (world doesn't work how you'd think), counter-narrative (world doesn't work how you were told), shock-and-awe, elegant articulation (Naval-style compression of a rich idea into one sentence), and curiosity gap. Writing quality = novelty × resonance; draft one surfaces the novelty, draft two adds story and analogy to make it land.

Topic selection pairs an objective (prove the status quo wrong, say what nobody is saying) with a personal motivation (get something off your chest, solve a nagging problem) — both together sustain follow-through to completion.

The Creativity Faucet: John Mayer, Ed Sheeran, and Neil Gaiman independently describe the same process. The first output is always wastewater — bad ideas that must be emptied before good ones arrive. Resisting them stalls the process; emptying them trains pattern-matching. Sequence: weak imitation → identify what makes it weak → iterate until original.

Customer-led growth | Georgiana Laudi (Forget The Funnel)

Standard SaaS growth frameworks fail because they measure business metrics instead of customer value. Funnels, MQLs, and SQLs bucket every customer identically, ignore post-acquisition revenue, and omit the "struggle phase" — what life looked like before customers found you. That missing context is where marketing leverage lives.

Gia Laudi's alternative, built at Forget The Funnel with co-founder Claire Suellentrop, runs in three steps: identify your best customers, map their value journey, find where you're failing to deliver.

"Best customers" means a specific slice: happily paying, high value, signed up within three to six months — recent enough to remember life before. Surveying them surfaces jobs-to-be-done patterns that personas miss. King Charles and Ozzy Osbourne share age, country, dog, and cars — entirely different motivations. One customer job is then selected based on willingness to pay, urgency, and retention potential.

That job anchors a six-milestone journey: problem (experiencing pain), interest (evaluating solutions), first value (initial activation), value realization (the "hell yes" moment), continued value (habit formation), and value growth (expansion). Each milestone gets a KPI tied to actual product behavior.

The SparkToro case demonstrates the method. Free-to-paid conversion was underperforming despite strong traffic. Surveys revealed two underexposed features customers valued most: lists and export. An onboarding checklist pushing users toward search → lists → export doubled trial-to-paid. A separate social media tool updated messaging from the same research and cut trial length from 30 to 7 days; website conversion rose 89% and trial-to-paid 40% without any product changes — better-qualified entrants convert on their own.

The research output is a five-to-seven-page messaging guide covering value prop, competitive advantages, and emotional and functional benefits in customers' own language — baseline for all copy, email onboarding, and in-app prompts. Bob Moesta's *Demand-Side Sales* and April Dunford's *Obviously Awesome* are the named reading anchors.

How to build a high-performing growth team | Adam Fishman (Patreon, Lyft, Imperfect Foods)

Growth teams fail not because they lack talent but because founders pattern-match to a senior hire instead of diagnosing what skills they actually need. Adam Fishman's growth competency model maps four quadrants: growth execution (channel fluency, experimentation, productizing learnings), customer knowledge (data fluency, user psychology), growth strategy (loop modeling, capital allocation, roadmap sequencing), and communication and influence (strategic framing, team leadership, stakeholder management). Junior hires should be strong on the first two; strategy and influence develop through years of hard experience.

Fishman prefers internal transfers for early hires — they already carry customer knowledge and execution, and you know what you are getting. Promoting a marketer into a growth PM role at Patreon outperformed most external searches. External "surgeon" hires — SEO, paid growth — belong only once foundational loops are in place.

Onboarding is the one product surface 100% of users touch, and users are most motivated there. At Patreon, Fishman's team identified high-potential creators by pulling audience size and engagement data across platforms via OAuth, then routed them to a human advisor during onboarding. That raised first- and second-month creator revenue by 25%, a key LTV input. The human patterns were encoded as "opinionated defaults" — preset three-to-five pricing tiers with friction added when creators tried to deviate. Revisit onboarding only when a genuinely new insight about users emerges; tweaking without fresh learning rarely moves retention.

Company selection requires a PMF lens: People, Mission, Financials — all three, not two of three. Fishman's failures at Wyzant (founders expected a silver bullet, not a system) and Imperfect Foods (C-suite infighting blocked execution) trace to shortcuts on one dimension each. Diagnostic moves: attend an executive meeting before signing, ask how leadership navigated its last major disagreement, and back-channel references with people your future manager has managed, including those let go.

What is a good activation rate

Across 500+ products, the median activation rate is 25% and the average is 34%; SaaS alone averages 36% with a median of 30%. Marketplaces and e-commerce land lowest because they typically define activation as a first transaction; B2C freemium ranks highest because milestones like "logged first meal" or "listened to 40% of a track" are low-friction.

A valid activation milestone must be at least 2x predictive of long-term retention and actionable by a growth team. The most common errors are setting it too early (sign-up completion) or too late (multiple purchases). Only about 6% of companies time-bound their definition, with a median window of 10 days.

The most-cited levers for improving activation are simplifying onboarding UI, reducing friction via defaults and templates, email/push follow-up sequences, better top-of-funnel qualification, and white-glove sales outreach for B2B. Causality between milestone and retention is proven by running experiments that improve the metric and watching whether downstream retention moves with it.

How to determine your activation metric

A good activation metric is causal for retention, not merely correlative. Finding one follows three steps: brainstorm candidate "aha" moments, run regression analysis against long-term retention, then run experiments to confirm causality. Teams with little data lean on qualitative interviews first; one e-commerce team found two completed orders retained users at 2× the rate of one order.

For multi-player B2B SaaS, two parallel metrics work best: a user-level action experiments can move, plus a workspace-level outcome that confirms team value is forming. Snyk's "F30D" — one team fix within 30 days — predicts three-month retention and purchase propensity; a second fix adds no predictive signal. Airtable tracks W4MUA (two users active together in week four) alongside a build-rate sophistication score. Slack's early metric was a team of 3+ exchanging 50+ messages in seven days. Figma's is two users editing or commenting in the same file within 24 hours. Sprig found teams receiving their first in-product survey response are 21× more likely to stay active over two years.

Linear's contrarian view: real adoption takes months, so measuring artificial milestones risks optimizing for the wrong signal.

Five steps to starting your product-led growth motion

PLG replaces "show" (demos) and "tell" (marketing campaigns) with "do" — letting users experience the product directly — which makes it more efficient and scalable than sales-led growth for most B2B SaaS companies. Hila Qu, who launched GitLab's PLG motion on top of an existing sales team, lays out five steps.

Map your funnel. The key difference between PLG and SLG is that product usage becomes the primary nurturing channel, generating PQLs instead of MQLs. PLG fits prosumers and SMBs best; SLG fits complex enterprise products. Most companies should layer both eventually.

Pick a starting point. Identify the biggest constraint — acquisition, activation, or conversion — and focus there first. For activation, B2B has four sequential aha moments: individual user, team (e.g. GitLab's "2+ users, 2+ features in 14 days"), buyer, and paid customer. Activation is product-led but not product-only: onboarding, lifecycle emails, and human assistance all play a role. For conversion, build self-service checkout first, then a PQL system. GitLab's hand-raiser PQLs — a sales contact form embedded in-product — converted 3x higher than MQLs.

Anticipate pitfalls. The five most common: no strategic conviction (commit to 1–2 years); no free product vehicle (build one); no data infrastructure; no PLG expertise (hire an advisor first, identify internal talent second, hire externally only after early wins); and internal resistance from sales and marketing teams who fear quota cannibalization and credit disputes. Executive buy-in and cross-functional alignment resolve the last one.

Mastering onboarding | Lauryn Isford (Head of Growth at Airtable)

Onboarding is the highest-leverage, most undervalued growth lever in self-serve products. At Airtable, a six-to-eight-month rebuild produced a 20% lift in activation. The biggest driver was replacing tooltips with a guided wizard: users pick their project type on the left while their workflow assembles visually on the right, reducing cognitive load enough that one generic flow served more than 90% of users before personalization was added.

Personalization by learning style and building style outperforms role-based segmentation. A database-familiar builder needs different scaffolding than someone invited to fill in a sprint.

The right activation metric is counterintuitively low. Airtable's north star was "week-four, multi-user active" — more than one person contributing to a workflow in week four. A 5–15% rate signals strong future retention; 40% usually means the bar is too easy. Secondary metrics — individual week-two retention and a "Build" sophistication score — let the team tell whether a treatment deepened engagement or merely drove more invites.

Two common traps: naming features ("This is Automations") instead of enabling contextual one-touch setup; and letting pricing tiers drive onboarding content, pushing premium features at beginners for conversion reasons rather than because beginners need them.

The reverse trial is the recommended structure: start users on the full premium experience for 7–30 days, then drop to freemium. Users anchored to what they stand to lose convert at higher rates than those starting on a bare free plan.

The PLG funnel — Join → Evaluate → Upgrade → Expand — maps to team structure. Airtable ran acquisition, activation, and monetization teams from the start, adding an Expand team as upmarket traction grew. Its value is shared vocabulary across functions. B2B growth follows the same logic but demands more care per customer — conversations and beta testing substitute for the high-volume A/B tests consumer scale enables.

10 lessons on bootstrapping a $200m business | Patrick Campbell (ProfitWell)

Bootstrapping ProfitWell to a $200M exit was a mistake — the company could have reached a billion-dollar outcome had it raised earlier. Bootstrap through ideation and product-market fit, then raise when going for the fences. If the idea can't reach $1B in revenue, stay off the VC treadmill; a $10M cash-flowing SaaS stands on its own.

On team: manager tenure in tech averages 15.7 months — proof "team is everything" is lip service. Values need real trade-offs. ProfitWell's was "most charitable interpretation": assume good intent in conflict; those who couldn't were let go.

On shipping: tempo framework beats org design. Define what "good looks like" per function; leadership conversations become gap-closing exercises. Most underperformance is misaligned expectations, not bad hires.

On pricing: most companies touch pricing once every three years. One action per quarter compounds. The highest-leverage choice is the value metric (per seat, per visit): it segments acquisition, cuts churn 20-25%, and doubles expansion revenue.

On retention: 25-40% of churn is tactical — failed payments, weak cancellation flows — and product teams ignore it. Two questions dominate: "Why are you leaving?" (multiple choice) and "What did you like?" The second triggers nostalgia and stops the cancellation.

On customer research: 1 in 5 companies have buyer personas; 1 in 10 do research quarterly. Active customer development correlates with 15-20% higher growth and better NPS and LTV/CAC. Start with 10 non-sales conversations per month.

On competitive intelligence: 16x more SaaS competitors than a decade ago; B2B CAC up 110%. Campbell (ex-NSA analyst) ran white-labeled NPS surveys to competitors' customers and built positioning strategy from it.

On middle-of-funnel: 80% of budgets go to top and bottom. Freemium users who self-convert retain 10-20% better and show roughly double the NPS of trial-forced conversions — because they arrive on their own timeline.

How Duolingo reignited user growth

Duolingo grew DAU 4.5x over four years by targeting Current User Retention Rate (CURR), which modeling showed had five times the DAU impact of any other lever. Jorge Mazal, CPO during this period, recounts two failed attempts first: borrowing a finite-moves counter from Gardenscapes (neutral result — Duolingo answers require recall, not strategy) and an Uber-style referral program of free premium subscriptions (3% new-user lift; most active users already paid, excluding the best referrers). The lesson: ask why a feature works in its source context before adopting it.

The CURR finding came from a Zynga-derived model dividing all users into MECE buckets (New, Current, Reactivated, Resurrected, At-risk WAU, At-risk MAU, Dormant). Sensitivity analysis — moving each rate 2% per quarter over three years — showed CURR dwarfed every other lever through its self-compounding return loop.

Three vectors moved CURR. Leaderboards adapted from FarmVille 2 kept league-tier progression but stripped extra tasks; users competed through ordinary study, learning time rose 17%, and highly-engaged learners tripled. Push notifications were optimized on timing, copy, and localization but never increased in quantity — Groupon's aggressive testing had destroyed their channel, so volume became a CEO-level constraint. Streak optimization began when an APM found that a 10-day streak sharply cut churn; a streak-saver late-night alert, then calendar views and animations, compounded that effect.

Over four years CURR rose 21%, cutting daily churn of active users by over 40%. Users with 7-day-plus streaks grew to more than half of DAU, fueling Duolingo's 2021 IPO.

Career frameworks, A/B testing mistakes, counterintuitive onboarding tips, selling to developers | Laura Schaffer (VP of Growth at Amplitude)

Adding onboarding questions with no downstream personalization raised Twilio signup conversion by 5%. The result traces to user psyche: first-time signups arrive expecting difficulty, and screening questions ("what language do you code in?", "what product do you want to use?") gave reassurance rather than friction — confirmation they were in the right place. The same principle resurfaced when prescriptive onboarding put "get a phone number" as step one. Developers found telecom unfamiliar and threatening; kicking users out of the console into a docs page — code first, phone setup buried inside familiar context — converted better. Embed the scary pill in the hotdog.

On experimentation: Microsoft and Netflix data show 80–90% of hypotheses fail. A/B testing should be reserved for ideas already vetted through cheaper paths — mockups, painted doors, qualitative interviews. Teams judged on weekly metrics get pushed toward data-fitting; growth teams need at least a year's mandate for failures to compound into signal. On confidence intervals: 95% is unnecessarily conservative for consumer product work. Accepting a lower threshold, paired with qualitative corroboration, can double experiments per year and produce a net-positive outcome — but the threshold must be set before the experiment runs, not fitted to results after.

Developers skip marketing websites and enter signup context-free. They avoid sales so strongly that enterprise teams have used personal email addresses to dodge outreach. The reason is accountability: if the product fails, the developer who chose it owns that failure. Any product requiring developers to build must invest in self-serve as seriously as an enterprise sales org.

Career framework: build visibility by staying closer to customers than executives can, then surface insights proactively. At Twilio, a voice-of-customer digest started within months of joining grew into CEO-attended quarterly sessions — creating conditions for funding a growth team that hadn't previously existed.

Hot takes and techno-optimism from tech's top power couple | Sriram and Aarthi

Technology is the greatest equalizer in history — Sriram Krishnan (a16z partner) and Aarthi Ramamurthy (former Netflix/Clubhouse PM) ground this in biography: both grew up middle-class in India, parents saved to buy their first computers, and they met coding online. The richest person on earth uses the same iPhone you do; Google returns the same search results regardless of net worth.

On bootstrapping social networks: Eugene Wei's "Status as a Service" is the right frame — recruit high-status people underserved by existing platforms, then cultivate homegrown stars. TikTok rewarded video/dance skills Instagram didn't, producing Charli D'Amelio instead of Kylie Jenner.

On networking: two coffees a week with peers, no agenda, follow up once a year, let it compound. Publishing online works even when content feels basic — Sriram's most-shared post was "how to write a cold email." What's obvious to you isn't obvious to everyone, and the internet sends interesting people to whoever is visible.

Imposter syndrome is universal. The fix: identify the one narrow thing you genuinely know better than anyone in the room, lead from there, build outward — don't try to fake competence elsewhere.

Jobs-to-be-Done cannot model multi-agent trade-offs, which is where real product decisions live. Facebook's "People You May Know" made your feed worse to help new users reach 10 friends in 14 days. Amazon stopped emailing order details to prevent Google from mining inboxes. Twitter's algorithmic timeline infuriated power users but served newcomers. None of these fall out of asking "what job is the customer hiring this for?" Systems thinking — mapping all players' incentives and how they interact — handles these trade-offs. Duolingo's streak mechanic came from product intuition about psychology, not a JTBD brainstorm. First-principles thinking (would you build it the same way from scratch?) is the other tool worth reaching for first.

The ultimate guide to adding a PLG motion | Hila Qu (Reforge, GitLab)

PLG is fundamentally data-led growth — the free product is the entry mechanism; what you get in exchange is behavioral signal. Without usage data infrastructure, giving it away yields nothing.

The most common failure isn't a bad free product: it's treating PLG as a single launch event rather than a two-year motion. Red flags: assuming "add free trial = self-serve revenue," assigning one coordinator instead of a dedicated team, and skipping fit validation. Fit requires short time-to-value, low complexity, and a large user pool.

PLG diverges from sales-led at one point: product usage replaces marketing engagement as the leading indicator. Two conversion paths follow: self-checkout for low-price deals, and product-qualified leads (high usage + ideal company profile) where sales engages for a larger contract.

To find where to invest, walk the funnel as a naive user: landing page, sign-up, first use, aha moment, checkout. Activation is the right starting place for most B2B products. GitLab's activation metric — two users using two features within 14 days — was derived by correlating candidate actions against 90-day conversion and 30-day retention, then validated via experiment (correlation doesn't prove causation). Miro's bar: three questions, a matching template, value in under five minutes. Conversion is close to revenue too — one client immediately lifted India success rates by adding a locally-supported payment method.

Data infrastructure in priority order: product analytics (Amplitude, PostHog), event routing (Segment), experimentation tooling (Eppo), and a lifecycle marketing tool that triggers on behavior rather than lead score. Before any tooling: audit existing instrumentation against a data dictionary — garbage-in makes analytics misleading.

For the initial team, hire a data analyst before or alongside the growth PM. The growth PM owns funnel metrics — conversion rate, activated accounts — not feature roadmap. Match hire experience to the starting focus area.

How to do linear regression and correlation analysis

Correlation and linear regression answer different questions and should be run in sequence, not treated as alternatives. Correlation (output: −1.0 to 1.0) measures whether two variables move together — it's the right starting tool when scanning which features or behaviors associate with retention, churn, or activation. At MyFitnessPal, Olga Berezovsky used Amplitude's Compass feature to find a 0.564 correlation between food logging in the first 7 days and 14-day retention; Mixpanel's Signal returned 0.78 for the same pair and pinpointed that logging at least twice within 3 days drove the effect. Neither number proves causation or tells you how much to move the dial. That's where linear regression enters: it fits a trend line so you can estimate "if food logs increase by X, retention moves by Y." Regression requires addressing outliers — those far from the mean skew the trend line disproportionately. For execution: Amplitude and Mixpanel handle correlation natively; linear regression runs via Excel's Analysis ToolPak (`=LINEST`), Google Sheets, or calculators like DATAtab. Run correlation first; proceed to regression only if the score is meaningful.

How to drive word of mouth | Nilan Peiris (CPO of Wise)

Wise built a business where 70% of new customers — 700,000 of the million joining each quarter — arrive through word of mouth. The constraint: Wise charges ~0.65% versus the industry's 6–7%, leaving almost no margin for paid acquisition.

The diagnostic is NPS, valued not for the headline number but for a doubling pattern: moving users from detractor (0–6) to passive (7–8) doubled referrals; passive to promoter doubled again; 9–10 doubled once more. NPS comments emailed weekly to the whole company consistently surfaced three levers: price, speed, ease of use.

Entering a market at 5.9% versus a competitor's 6% generated usage but zero advocacy. Recommendation appeared only when Wise was 8–10x cheaper. The bar is an experience people didn't know was possible — not incremental improvement.

The price mechanism: allocate every cost — partner fees, FX risk, support calls — atomically back to the transaction that generated it. The top 20% of cost-generating customers get price increases; everyone else gets cuts. Engineering eliminates each cost category. Securing central bank accounts at the Bank of England, Bank of Singapore, and Reserve Bank of Australia — years of lobbying each — cut partner fees dramatically.

Two additional drivers matter. Mission resonance: a rebrand email with no call-to-action went viral because authenticity was legible, generating more signups than any campaign. Perception gap: customers believed they saved money but didn't trust the stated number. A PM built a comparison graph showing what banks hide in the exchange rate versus Wise's fee; placing it on the success screen with a share button produced a 3x referral rate increase. Instant transfers needed an animation making speed legible — referral rate jumped again.

The governing principle: conviction over experimentation. Teams build qualitative conviction that a lever matters, then execute the hardest version rather than AB-testing toward it.

Lessons from going freemium: a decision that broke our business

Bobby Pinero (CEO of Equals, a next-gen BI spreadsheet) added a freemium tier after Series A pressure — and it nearly destroyed the company. Users 4x'd immediately, then stalled: engagement cratered, retention collapsed, ARR tanked. Removing the mandatory data-source connection made it worse — CSV workarounds and full bypass both shrank activated users, because the allure of seeing the product is the only motivation new users have to complete hard setup. Let them skip it and they never return.

The fix: kill free, require a credit card and a 14-day trial. ARR bounced back immediately; deeply engaged users rose with the same signup volume. Credit card commitment drove setup completion more than any in-product nudge.

Freemium requires: massive addressable user base, very short time-to-value, foundational early-adoption role (like AWS), low incremental serving cost, and viral loops from free users (Loom, Miro). Equals failed on three. The meta-lesson: users demanding less friction, advisors endorsing freemium, and SaaS darlings modeling it are noise — none carry your retention curve.

How to make an impact in your first 90 days

The first 90 days disproportionately shape your next two years, and the beginner's mind is the asset — you catch what insiders miss. Operators from Ramp, Canva, Dropbox, and Wiz offer quick wins across five areas.

Funnel friction: Linktree's Jiaona Zhang swapped a logo for a "Create a Linktree" CTA and nearly doubled signups; a signup form over a blurred UI lifts conversion via curiosity bias; fixing early-page latency signals quality before users evaluate core value.

Pricing: Madhavan Ramanujam (Simon-Kucher) argues $273 makes buyers overthink — round to $299 and lead with the monthly rate on annual plans. Elena Verna restarts paid trials for existing users to revive urgency. Discounts should extract concessions: references, price escalators, annual commitments.

Operations: cut meetings 50%; replace with focused rituals; send weekly written updates to build async visibility.

Data: only 20–30% of signups drive revenue — focus there even at the cost of volume. Measure ad ROI against retained customers, not conversions. The 40% of activated users who churn within seven days usually have two or three fixable drop-off reasons.

Customer conversation: ask "if you had a magic wand, what would you change?" to surface problems rather than feature requests. Canva's CMO counters: take time to just be new — those early relationships compound over the long haul.

Summary: The ultimate guide to adding a PLG motion | Hila Qu (Reforge, GitLab)

PLG and sales are not a choice but a sequence: PLG lowers the barrier for individuals to try the product and generates usage data; sales closes large targeted accounts. Hila Qu, who built GitLab's PLG motion, argues retrofitting PLG onto a pure sales-led company is far harder than starting with it.

Three things define a PLG product: low-barrier free entry (no "book a demo" gate), self-service checkout, and organic spread through use. The most common failure is treating "launch a free trial" as the whole motion — without a dedicated team, mapped funnel, and data foundation, you're giving away the product for nothing.

The funnel has three layers: marketing site optimized for free signups, free version that guides users to the three most valuable features, and a smooth checkout. GitLab defined their aha moment as "two users using two features within 14 days"; Facebook's was 10 friends in 7 days.

Prioritize activation first: Miro asks minimal questions and immediately supplies templates. Once activation works, make checkout as frictionless as Amazon. Only then does investing in product-qualified lead scoring pay off.

The minimum team is a growth PM (analytics-first), a data analyst, and engineers — start as a tiger team, formalize into growth product, growth marketing, and product-led sales functions after early wins.

The Subscription Value Loop: A framework for growing consumer subscription businesses

Consumer subscription apps are structurally hard to scale: RevenueCat data from 30,000+ apps shows top-quartile apps convert 1 in 20 installs to paid and lose more than half of annual subscribers in year one. Four forces drive this: app store lock-in (15–30% fees), paid acquisition weakened by Apple's ATT, high churn without network effects, and a single-SKU model that blocks ARPU expansion. Fewer than 50 have reached $1B.

Phil Carter's Subscription Value Loop frames growth as three reinforcing steps: Value Creation (fast connection to the core promise), Value Delivery (organic and paid acquisition), Value Capture (conversion and retention) — each stage funds the next, lifting LTV/CAC and shortening payback period.

Ten levers by rough impact: (1) PMF first — Superhuman raised its Sean Ellis 40% Test from 22% to 58% by targeting speed-obsessed founders. (2) Pricing is under-optimized at most apps; Van Westendorp analysis yields 5–15% revenue lifts; AllTrails raised its annual price $29.99→$35.99, confirming 8.3% ARPU gain. (3) Onboarding speed — 75%+ of trial starts happen on install day; front-load value screens, back-load registration (Headway overtook Blinkist doing this). (4) Web conversion flows bypass ATT and app-store fees; Palta reports 30–50% LTV lifts; Ladder fed predicted-LTV quiz scores back to TikTok and grew 9K→100K+ subscribers. (5) Paywalls need 80%+ Day-1 view rates; Blinkist's trial-timeline paywall lifted trial starts 23%. (6) Premium merchandising: Tinder's three-tier model plus IAPs drove hundreds of millions incremental. (7) Involuntary churn is 20–40% of total; Spotify recovers it with dunning and local payment methods. (8) Motivation mechanics: Duolingo's streaks and leaderboards grew DAUs 4.5× targeting Accomplishment, Ownership, and Avoidance. (9) Word of Mouth Coefficient (Zynga) measures actual referral conversion better than NPS. (10) Activity-based discounts — Flo triggers a lifetime offer after a user rejects the free trial, capturing low-WTP users without cannibalizing full-price revenue.

The original growth hacker reveals his secrets | Sean Ellis (author of “Hacking Growth”)

Product-market fit is measurable before retention data arrives. The "Sean Ellis test" — asking active users "how would you feel if you could no longer use this product?" — came from a workaround: senior managers gave lukewarm answers to satisfaction surveys, but loss framing drew honest responses. The 40% threshold was observed empirically across early YC companies; it's a coordination device to stop premature scaling debates, not a hard cutoff.

The real power is diagnostic. At Lookout Mobile Security, a 7% score climbed to 40% in two weeks: Ellis isolated the very-disappointed users (antivirus), repositioned around that benefit, and streamlined onboarding to deliver it first. Follow-on questions matter as much as the score: ask the very-disappointed segment their "primary benefit" (open-ended, then forced multiple-choice to a second cohort), then "why is that benefit important to you?" — at Xobni this surfaced "I'm drowning in email," which became the acquisition hook.

Ignore the "somewhat disappointed" segment — they signal nice-to-have, and optimizing for them dilutes the core. Superhuman's refinement: among somewhat-disappointed users who share the must-have benefit, ask what it would take to cross the line — targeting fence-sitters without polluting the signal.

Growth sequencing: activation before acquisition. At LogMeIn, 5% of signups ever completed a remote-control session; fixing onboarding over three months moved that to 50%, turned unprofitable channels into $1M/month at three-month payback, and drove 80% word-of-mouth growth. Priority order: activation → engagement → referral → monetization → then acquisition.

North Star Metrics should reflect customer value, not revenue — Amazon's monthly purchases, Uber's weekly rides, Airbnb's nights booked. Facebook's shift from monthly to daily active users shows how the metric reshapes behavior.

ICE (Impact, Confidence, Effort) lets cross-functional teams compare experiments without politics. Intercom's RICE adds "Reach," which Ellis considers redundant — reach is already inside Impact.

Behind the product: Duolingo streaks | Jackson Shuttleworth (Group PM, Retention Team)

Duolingo's streak is the biggest growth lever after the lessons — 9 million users hold a 365+ day streak at a $14B market cap. The consistent finding across 600+ experiments: streaks amplify an app users already want; without that foundation, a streak is a distraction.

The biggest mechanical shift: moving from an XP-based streak to a flat one-lesson requirement drove a large DAU jump. Testing the opposite extreme — count a single exercise — produced no gain and captured only the least engaged users. The right unit is the app's meaningful unit of use.

Loss aversion solidifies at day seven — the day-one-to-day-two retention jump is enormous; after that it flattens. Two standout zero-to-seven interventions: giving new users two free streak freezes immediately, and goal-setting. Telling users they are 7x more likely to finish a course with a 30-day streak won big. Adding an opt-out button won nearly as much — choosing "Commit to My Goal" over "Continue" mattered more than downstream mechanics. Pre-selecting a goal lost; the deliberate selection was where the effect lived.

Two streak freezes beat one; three only added days off. Earning back a lost streak by completing lessons outperformed charging in-app currency. A "Perfect Streak" visual — the number turns gold when no freeze is used — creates a counterweight toward perfection without cheapening the streak.

Notifications fire 23.5 hours after the prior session; a 10 PM streak-saver reads as welcome because users have positive emotion tied to the streak. Teams own metrics, not features — retention owns the streak because it drives CURR (current user retention rate) best.

Board member Bing Gordon's observation: users care about the streak because Duolingo visibly cares — the streak screen gets the most animation budget, appears after every session, and anchors most push notifications. Visibility signals priority.

How to find hidden growth opportunities in your product | Albert Cheng (Duolingo, Grammarly, Chess.com)

Growth's job is to connect users to product value, not to hack metrics. Albert Cheng, who led growth at Duolingo, Grammarly, and Chess.com, structures that work around an explore-then-exploit cycle: find a human-psychology insight that cuts across the product, then have adjacent teams apply the same principle to their surfaces. At Chess.com, discovering that 80% of game reviews happen after wins — not losses — led to flipping the post-loss screen to show brilliant moves with coach encouragement. That change grew game reviews 25%, subscriptions 20%, and retention materially.

The Grammarly monetization win followed the same logic. Free users experienced Grammarly as a spelling fixer because those were the only suggestions they saw. Interspersing paid suggestions — tone, clarity, full-sentence rewrites — into the free experience created a real-time reverse free trial. Upgrade rates nearly doubled. The lesson for freemium products: the free tier should show everything the product can do, not a stripped-down subset.

For consumer subscription businesses, existing-user retention is more leverageable than new-user acquisition once a product matures. At Chess.com roughly 80% of weekly actives are existing or resurrected users, so resurrection flows compound more than acquisition spend. Duolingo used social notifications showing friends joining to pull dormant users back; D1 retention of 30–40% is the floor.

On experimentation: Chess.com ran practically zero experiments before 2023 and is targeting 1,000 in 2025. Cheng treats that number as a forcing function — what would need to be true? No-code tooling for key screens, lifecycle copy tests, app-store asset experiments. When insignificant results cluster, the vein is mined out and the team returns to open exploration.

On teams: the highest performers prioritize clock speed and agency over domain experience. In AI-era environments, accumulated habits can be a liability; a beginner's mind compounds faster than seniority.

5 questions to ask when your product stops growing | Jason Cohen (2x unicorn founder)

Growth stalls at every SaaS company by mathematical necessity: cancellations compound with scale while marketing output doesn't. Signups divided by churn rate sets a hard ceiling — at 5% churn and 100 new customers per month, the company tops out at 2,000. Cohen's five questions run in fixed order; each only matters once the prior is satisfied.

Logo churn first. Customers who cancel survived a near-impossible gauntlet — ad, click, pricing page, purchase, onboarding — and still left. "Too expensive" is almost never the real cause; they accepted the price at purchase. Ask "what made you cancel?" not "why?" — Groove found that phrasing doubled usable responses. Cancellation clusters in the first 30–90 days; small gains in onboarding compound downstream.

Pricing and positioning second. Patrick Campbell's data from 4,200 startups: prices are almost always too low — founders guess and never revisit them. Raising prices often doesn't reduce signups — it selects a more credible market. The same product repositioned from "halve your AdWords cost" to "double your leads" commands 8x the price, selling against growth budget instead of savings.

NRR third. Logo churn sets the floor; NRR above 100% is the counter-force. Median NRR at SaaS IPO is 119%. A 20% loss requires a 25% gain to recover, so tracking logo count alongside NRR is necessary.

Channel saturation fourth. Channels follow an elephant curve: growth, plateau, sag as audiences exhaust. Constant Contact restarted growth with in-person city workshops; HubSpot built half its revenue through agency partnerships. Saturated channels call for a new channel type or a second product, not more spend.

Do you need to grow at all, fifth. 37Signals optimizes for profit rather than revenue. "If you're not growing, you're dying" applies more to the person than the company — stagnant companies still pay dividends.

The hidden pattern behind successful products | Mark Pincus (Founder of Zynga)

Most hit products follow a pattern Pincus calls Proven, Better, New: catalogue what works on the exact platform for the exact audience (proven), make one change every existing user calls unambiguously better (better), add one novel hook (new). Sid Meier's social Civilization on Facebook died in minutes because his first-time user flow violated proven onboarding norms. Words with Friends was Scrabble (proven), polished for mobile (better), Facebook friends already in the game (new). Slack was proven and better with no new.

Instincts are right 95% of the time; ideas are right only 25%. The fix is many fast tests, not one MVP sustained by hope. Asking "is this an A?" signals it is not; when you have a real hit you know before launch. Pincus spent four years and $25 million on a project before cancelling; two weeks later he had more ideas than in the prior four years.

Less ambition at the start produces bigger outcomes. Zynga launched as a Facebook poker game — embarrassingly small for a 41-year-old multi-time founder — and that humility was the key. After success founders over-raise before finding product-market fit; newcomers forced into humility have the advantage. Zynga's edge was Day 365 retention and a metric called ASN: one round-trip friend exchange predicted 80% chance of return the next month; four exchanges predicted 22 active days out of 30.

Consumer distribution is broken: AI is not yet a platform, discovery is dead (zero of 40,000 games last year cracked the top 10), social has lost its cocktail-party energy — people are proud to have quit Instagram. The next breakout restores productivity via agentic lead generation.

Management: give every direct report real CEO authority; best hires are frustrated expert witnesses who were right but never let run. The CEO's job is to be right.

The Power of Performance Reviews: Use This System to Become a Better Manager 🤝

Performance reviews done well accelerate careers; done poorly, they accelerate departures. A three-step system front-loads the work: Prepare by gathering peer feedback (start/continue/stop questions, anonymous, 5–8 colleagues) plus a self-review — and form your own view before reading responses to resist anchoring. Deliver using a template covering accomplishments, one named superpower (research favors strengths over weaknesses), a short narrative arc, and only one to two development areas; more than two prevents meaningful progress on any. Then follow up.

This Week #7: Effectively communicating about a failure to execs, managing founder expectations, and hiring a Director of PM

When an MVP fails, name it outright, reconstruct the journey (with early wins and user quotes), separate the idea from the execution, and arrive with a concrete next-step recommendation — that last part is where most time should go. Changing a founder's mind on a failing idea requires sharing data points as you discover them, aligning on testable hypotheses, and framing the pivot as an evolution of their own thinking rather than a refutation; accept that some founders won't be moved. Hiring a Director of PM differs from IC hiring on three axes: managing PMs, large scope, and senior-leader partnership — probe people leadership, stakeholder management under pressure, and hiring judgment.

This Week #11: What should new PMs over-index on, and empowering product in a sales-driven org

Jackie Bavaro's advice for new PMs targeting a senior role: four priorities in order. Ship wins — join high-cadence teams with strong teammates, stay long enough for results to register. Learn best practices at a company with real PM culture; rotational programs accelerate pattern-matching. Get baseline credentials (top undergrad or 2+ years at a known tech company) early, since credential-rich and learning-rich places overlap. Build your network now — junior peers become senior sponsors later. On sales-led orgs: product is structurally second-class in B2B companies; move to a product-led org instead.

Turning around an underperforming team - Issue 17

Joining a turnaround team is a career accelerant if you treat the risk as leverage. First, calibrate your manager's expectations — business impact will dip for 6–12 months and some people will resent you. Set a high bar from day one, calling out weak work privately and strong work publicly. Before acting, run through Observe → Identify → Share → Act: collect facts, diagnose root causes (process, people, charter), align peers and leadership, then move. Find allies early; the unsure majority is usually winnable.

Moving from IC product manager to manager of product managers

The PM manager role is leverage, not execution: your output equals your team's output plus neighboring teams you influence (Andy Grove's formulation). The five real jobs are stopping bad decisions from shipping, unblocking at every time horizon, holding the PM performance bar rather than letting mediocre work slide, pushing product quality without micromanaging, and building a cohesive peer-leadership group. To earn the promotion: mentor junior PMs, deliver high-stakes initiatives, develop a written strategy — then explicitly ask; managers don't assume you want it. Once there, teach rather than do.

A comprehensive survey of Product Management

PMs overestimate their own influence: when non-PM respondents (engineers, designers, CEOs) are isolated, only 70% say PMs have more influence than other functions — versus 80% when PMs self-report. Across ~1,000 responses from 600+ companies, the top hiring criteria are communication, execution, and product sense; design/UX and raw intelligence rank last. Promotion hinges overwhelmingly on showing business impact; keeping managers happy matters mainly at Apple, IBM, and Oracle. PM influence is highest at YouTube, LinkedIn, and Airbnb; lowest at Apple, Stripe, and Tesla.

Should I become a product manager

Whether to move from engineering to PM comes down to one question: are you more excited by business and customer problems than technical ones? PM work splits into three buckets — shaping the product (customer research, strategy, PRDs), shipping it (timelines, unblocking engineers), and aligning people (stakeholders, goals, communication). Don't switch expecting authority, higher pay, or an easier path — PM roles are scarcer than engineering roles and compensation is roughly equal. Shadow a PM first; don't rush with only three years in.

Becoming a senior Product Manager

The jump from PM to senior PM takes a median three years — longer than any other rung — so accelerating it has the biggest leverage on a PM career. Three differentiators separate the levels. Strategy: senior PMs spot market opportunities, build a vision-framework-roadmap, and convince executives to fund teams, rather than executing a given scope; block half a day and write one now. Autonomy: earn trust through proactive communication using a template (here's a challenge I faced / here's how I handled it / any thoughts?). Nuance: replace rigid frameworks with "if X then A, if Y then B" reasoning, and trace stakeholder concerns to their second-order consequences before overriding them.

14 habits of highly effective product managers

Clear writing is the most fundamental PM skill — people judge thinking by how it's communicated. Beyond that, great PMs build an "I've got this" reputation by saying no selectively and tracking every open thread. They hold the bar high on team output, hunt actively for misalignment ("What's your understanding of X?"), and maintain a source-of-truth document anchoring goals and key decisions. They hold strong opinions loosely, ruthlessly prioritize to give teams focus, unblock bottlenecks proactively, and form a tight triad with their engineering and design managers. They connect work to mission, talk to customers regularly (five user conversations before any major feature), amplify team members' contributions publicly, and set the emotional tone for the group.

My favorite PM interview questions

Each of ten PM jobs maps to one behavioral question. Impact: "what's the most important product you shipped?" — flag if they had little influence over it. Ownership: ask for a failure, then ask for a second; deflection is the tell. Execution: a 7–10 minute walkthrough of a 3–9 month project, probing delays and exec alignment. Strategy: flag if their articulation is hard to follow. Planning: co-created deadlines, not mandated ones. Communication is judged across all answers.

How to know if you're doing a good job as a product manager

PM performance reduces to three questions in priority order: is the team delivering business impact (growth, efficiency, sales, customer happiness, launches)? Are you personally contributing — shaping strategy, surfacing insights, proposing what to build, not just coordinating? Do stakeholders regard you highly? Aim for roughly 80% loving your work, a few feeling uneasy — 100% approval signals insufficient pushback. Shipping impactful features while key leaders think the product was wrong still hurts your review.

The 10 commandments of salary negotiation

Never volunteer a salary number — ask for the band instead. At companies over 5,000 employees, a compensation committee sets offers formulaically; recruiters fish for your number to save a round-trip to that committee. Read where your initial offer lands: low means fix interview misconceptions before countering; the middle-of-band template (over 80% of candidates get it) signals no strong internal advocate — build one first; top-of-band is leverage for an out-of-band bump. At startups, interrogate equity like an investor: get the strike price, outstanding shares, and runway. Before countering, hold follow-up meetings with decision-makers to build pull. Counter by phone, not email; no competing offer is required, and citing Glassdoor annoys recruiters. The strongest ask names obstacles the recruiter must overcome to close you. Sequence: total comp first, then equity, then signing bonus.

What is product management

A PM's output equals the output of their team (Andy Grove's formulation). The job breaks into three parts: deliver business impact, marshal cross-functional resources to do it, and identify the most impactful customer problems to solve. Day-to-day that means shape the product, ship it, and synchronize stakeholders. Skills weighted most in hiring: communication, execution, strategy, collaboration, product sense. The role varies sharply by company — influence ranges from mini-CEO to pure project manager, and every dimension differs between startup and large-company contexts.

The most common pitfalls of new product managers

New PMs fail as one of five types: the Coordinator executes others' ideas without forming a POV; the Dictator treats teammates as executors rather than peers; the Dreamer fixates on strategy before earning it through execution; the Feature Factory ships relentlessly without a strategy defining what *not* to build; the Busted Umbrella lets exec whims reach the team instead of absorbing them. Avoid all five while simultaneously building good habits from the other direction.

The top 5 things PMs should know about engineering

PMs don't need to code deeply, but five areas improve shipping speed and credibility. Know your tech stack — language and framework choices determine whether a feature takes hours or weeks; a NoSQL database skips migrations but risks data quality. Know the fragile codebase sections — poorly-organized areas cause delays that look mysterious until you ask. Know your build and deploy process — multi-day deploys change batching decisions. Learn to submit small code changes yourself (copy fixes, frontend tweaks via GitHub's editor) so you don't block engineers on trivial polish. Build conceptual grounding to parse "frontend-only change" statements: read engineering blogs from Slack and Netflix, look up unfamiliar terms, and keep a developer friend to consult.

How to get promoted

Consistent delivery of impact is the core driver of promotion, but seven tactics close the gap when you're stuck. Take on next-level scope before being asked — mentoring junior PMs or leading cross-functional projects signals readiness better than doing your current job well. Name 2–3 specific gaps blocking you with your manager, build an action-plan tracking them monthly. An influential champion fighting for you in calibration is the most underappreciated variable — document wins to give them ammunition. Watch who actually gets promoted to read your company's implicit priorities. If nothing moves after two-plus years, leave.

How to know when to stop

Sustained overachievement produces physical collapse, not just burnout feelings, and most high performers don't recognize the crossing until damage is done. Andy Johns — growth at Facebook, Twitter, Quora, Wealthfront — landed in the ER at 35 with an enlarged, thickened heart shortly after becoming president. His LinkedIn surveys found ~50% of tech workers had considered quitting at least three times, and ~50% had cried from work stress.

His three-part framework:

Range of tolerance. Every organism has an optimal zone; pushed outside it, humans show hormonal dysregulation (irritability, anxiety, shutdown), physiological breakdown, and destructive coping (nightly alcohol, etc.). A 29-item inventory surfaces the crossing — Johns had checked 17 boxes before his hospital visit. The output is three lists — intolerables (hard stops), boundaries (guardrails), flourish behaviors (resilience-builders) — mapped to a red/yellow/green visual target. Johns uses daily sauna as his anchor; a 20-year Japanese study correlated it with a 28% drop in cardiovascular disease.

Career S-curve. Careers run foundation, acceleration, and peak phases. What overachievers undercount is "mind hours" — 10–30 extra hours per week of nervous-system load (hiring, firing, organizational fires) on top of 50–60 office hours. Johns ran three acceleration cycles; by the third costs outweighed gains. Year 25 would have yielded ~$10M cumulative salary (70% in the final eight years); he left it; cortisol dropped 30% within 30 days.

Life S-curve collision. From the late 20s through the 50s, career peak overlaps with adult-life load — mortgages, dependents, health crises. Johns avoided the Bay Area, relocated to a low-cost town, and set his sufficiency floor at $35,000 post-tax — removing the lock-in that keeps most high earners in roles that breach their tolerance.

Short windows of intolerance can be worth it — hormesis, the mechanism behind exercise adaptation — but only when entered consciously.

Jackie Bavaro on getting better at product strategy, what exactly is strategy, PM pitfalls to avoid, advancing your career, getting into management, and much more

Good product strategy has three components most PMs conflate or omit: vision (an inspiring picture of the future), a strategic framework (the market you're targeting and your bets on what it takes to win), and a roadmap used not as a commitment but as a feasibility check — work backward; if the roadmap shows thirty years, take bigger swings or hire faster. A revenue target is an input, not a strategy; the PM's job is to connect that number to specific choices through a chain of explicit assumptions. Missing links surface as repeated confusion — people don't volunteer when they're lost, so listen for the nod that hides disagreement.

Strategy improves through experience and cross-domain borrowing. Shishir Mehrotra cracked a YouTube linking debate by importing a "consistency vs. comprehensiveness" framing from a Google Shopping review. After six months on a product, block half a day, draft whichever pillar is fuzziest, and test it with your engineering and design leads. Persistent feature disagreements are almost always strategy disagreements — surface the underlying principle and resolve it there.

Spend the first six months executing the existing strategy. Common early-career errors: treating the role as a "no machine" burns collaboration (Bavaro nearly got fired for this at Asana); deferring to a strong designer makes you invisible; crowding engineers out of credit alienates the peers whose reviews drive promotion. At any level, the most effective move is telling your manager "I want to reach X someday — what should I work on now?" It enlists them as an ally, filters feedback to what matters for the next level, and positions them to advocate in committee reviews. Senior IC pay at top companies matches doctor or lawyer salaries; management is not the only path upward, and its loneliness and information asymmetry are costs to weigh.

Brandon Chu on building product at Shopify, how writing changed the trajectory of his career, the habits that make you a great PM, pros and cons of being a platform PM, how Shopify got through Covid

Writing publicly accelerated Brandon Chu's career by roughly a decade — not because posts went viral, but because writing crystallized his thinking at the moment he was working things out. He stopped in 2018; posts still function as onboarding: incoming PMs know how he thinks before day one.

Shopify's product culture rests on three pillars: technical fluency expected of everyone (early on, even marketers had to commit to deploy a blog post), no pedestal for the PM function (engineers and support own product thinking too), and a 30–40% ex-founder share of the PM org. Job description: "Help teams ship the right thing at the right time in the right way" — servant leadership, not CEO of the product.

Annual planning issues "investment plans" — VP-level bets covering ~20% of the company — then directed chaos governs execution, with discipline to discard plans mid-stream. During Covid, merchants went to zero revenue overnight; Shopify collapsed its quarterly scope to gift cards, buy-online/pick-up-in-store, and a few survival features — shipped in two weeks. Covid made remote-by-default permanent. "Bursts" — quarterly sprints to self-selected locations booked through an in-house app — now substitute for office proximity.

Platform PM work demands a different psychology: validation cycles run five to ten times longer because the developer builds the app and the end user appears only two years later. The decisive early move is locking in which constituent wins ties. Amazon's "consumer over seller" and Shopify's "merchant over developer" produce opposite tradeoffs; without that principle, the CEO blocks the launch the day before it ships.

What differentiates at senior levels is storytelling, mobilizing teams through ambiguity, and high-conviction bets — not analytical hygiene, which is table stakes. Build a side project: one weekend with a Rails tutorial produces a working Twitter clone and permanently demystifies engineering.

Manik Gupta (ex-CPO Uber, Google Maps) on how to build consumer apps, why it’s useful to be optimistic about technology, creating inflections in your PM career, the changing CPO role, and more

Consumer products are harder and slower than founders expect — you cannot force adoption, so virality, distribution, and product must all earn their way. Two career-long patterns that compound: surrounding yourself with the best people, and maintaining genuine technology optimism, which steers you toward ambitious problems worth solving at scale.

Before chasing product-market fit inside a large company, ask a prior question: company-product fit. Does this product slot into the company's existing strength portfolio? Entering an adjacent segment because a competitor is there wastes teams and kills morale. Google succeeded on long-horizon technology bets distributed through Search; Uber succeeded through operations discipline and real-time P&L. The culture shaped what the product could be.

The five-capability "consumer stack" is a self-audit scorecard: (1) design-led thinking — pixel-level craftsmanship is table stakes; (2) radical focus — identify the critical user journeys, build only for those; (3) precise metrics, fully instrumented — teams debate "active user" because nobody locked the definition; (4) high ship and experiment velocity — not learning means not progressing; (5) strong cross-functional talent across PM, design, data, and engineering.

PM career inflection points correlate with product inflection points — causality between a PM's work and a shift in product trajectory is the signal to bet on them. The second filter is the IC-to-manager-of-managers transition: it demands building structure and coaching capacity, not direct execution. Early-career traps: letting process override progress, believing the "PM as CEO" myth (PMs are enablers), and refusing to admit mistakes. For promotion, weight demonstrated end-to-end impact, the ability to create clarity and energy, and followership — do smart people actively seek to work with this person?

The CPO role is morphing toward a GM model — owning PM, engineering, design, and data under single-threaded accountability — with the pure-functional C-suite title becoming less common.

How to unlock your product leadership skills | Ken Norton, Ex-Google

Most senior product leaders plateau not from skill gaps but from operating reactively — responding from fear, seeking approval, needing to be right, or asserting control — rather than from a creative mode grounded in purpose and vision. Bob Anderson and Bill Adams's research shows creative leadership correlates positively with every measurable success dimension; reactive negatively. Roughly 75% of leaders are primarily reactive.

Norton identifies three reactive postures: wanting to be liked (complying), needing to be right (distancing), needing to control (asserting). Each starts as a genuine strength that grinds to a halt as seniority increases. His own was people-pleasing: effective when influence without authority required making everyone feel heard, but eventually fatal to decisiveness. The unlock wasn't eliminating the underlying care but redirecting it — from "I need people to like me now" to "I want them to say, a decade later, they'd work with me again."

The shift isn't achieved by learning tactics. It requires interrogating belief systems that lock you into reactive patterns — especially unexamined archetypes of what "real" leadership looks like, which model only other reactive styles. Coaching surfaces buried assumptions through questions; advice delivers a sugar high but rarely changes behavior because it bypasses the actual constraint.

The most common PM blind spot: underinvesting in the people dimension — persuasion, difficult conversations, collaboration — while over-indexing on frameworks. The art grows in importance with seniority but is rarely treated as something to train.

Imposter phenomenon is near-universal in PM roles — the job is always adjacent to specialists who outrank you in their domain. Naming the inner critic and treating it as a protective part rather than your identity creates useful distance. Leaders also carry an obligation to address systemic forces that produce imposter feelings disproportionately for women and people of color rather than just coaching individuals through them.

How to build trust and grow as a product leader | Fareed Mosavat (Reforge, Slack, Instacart, Zynga, Pixar)

PM skill is "specific knowledge" in Naval Ravikant's sense — earned only by doing real work on real products. No pre-training qualifies you; courses don't substitute for reps. The learning loop runs four steps: execute, generalize into transferable rules (high-friction onboarding helps high-activation-bar products; low-friction helps high-intent ones), communicate those generalizations, then earn bigger scope — moving from known-problem/known-solution toward unknown-problem/unknown-solution work.

Career inflections come from sponsorship, not mentorship. Sponsors — Nabeel Hyatt at Conduit Labs, Merci Grace and April Underwood at Slack — handed Mosavat larger scopes because he taught the organization how the growth model worked. The route to sponsorship is "two stack levels up and down" curiosity: know your boss's priorities and their boss's, understand what sales, marketing, data, and finance are doing, and be the person others consult to connect the dots.

The transition from senior IC to manager has a named failure mode: the manager death spiral. New leaders keep the best work, hand off scraps, and burn out while the team stagnates. Fix: shift from doer to editor — what is the least input that makes this output as good as possible? A second failure: treating resources as fixed. Making the case for what the problem actually requires is IC extra-credit; at the leadership level it is the baseline.

Mosavat and Casey Winters identify four types of product work: feature (deepen engagement), growth (connect more customers to existing value), PMF expansion (new audience or new product for same audience), and scaling (trust, safety, technical debt from size). Leaders over-apply whichever type made their name. Building fluency across all four is what separates senior ICs from genuine product leaders.

For a fractional or advisory path: build a top-five-in-the-world intersection of depth and breadth, and work at known companies — their brand becomes your sales infrastructure.

Lessons from one of the world's top executive recruiters | Lauren Ipsen (Daversa Partners, General Catalyst)

The biggest mistake founders make hiring their first senior product leader is chasing big names — ex-Google CPOs, YouTube veterans — who have spent years managing managers, far from the work. The right question isn't "who is the best talent?" but "who is best for this role at this stage?" Someone closer to the work, with something to prove, who can be a player-coach, will outperform the whale hire at a pre-IPO company.

Before opening a search, define the mandate precisely. Platform infrastructure, core consumer product, and growth/monetization are genuinely different roles — one person covering all three is a unicorn brief that dooms the search. Map success at 90 days, 12 months, and through IPO. Default to "Head of Product" early-stage; over-titling creates demotion problems later.

Whether a hire is imminent, maintain a warm network of benchmark candidates. Approach great people with no agenda — they're typically receptive. A seven-month courtship of advisory conversations is how you convert "why would I ever leave?" into a signed offer. The best source is a trusted shortlist of elite connectors who vouch for referrals; LinkedIn cold outreach reaches the wrong pool.

For product leaders: breadth beats depth if the goal is leadership. Touch platform, core product, and growth. Short stints are a red flag when repeated — but don't hide them; omitting a company looks worse. The test for when to move on: can cross-functional peers name something you built? Impact only you know about won't survive back-channel references. Key reference questions: "Would you hire them again? Report to them? Stake your name on it?"

Recruiters' most consistent failure is transactional behavior — ignoring stated constraints (vesting cliffs, no-crypto preferences), never building rapport. Relationship-first recruiting compounds into warm pipelines that transactional competitors can't replicate. Trust lost with a candidate doesn't come back.

Leaving big tech to build the #1 technology newsletter | Gergely Orosz (The Pragmatic Engineer)

Gergely Orosz left Uber — where he earned $320–330K/year as a manager of managers — to write a newsletter, and now earns more than he did there with no theoretical ceiling. The decision grew from a promise when he joined Uber: if the IPO paid out and left him with savings, he'd take a real risk. COVID layoffs broke his trust in the corporate system. His original plan was VC-backed platform engineering, but six months of writing produced two self-published books making $100K in year one, and he recognized the Mexican fisherman parable: he'd been planning to do this work anyway — after a decade of startup grind first. He skipped straight to it.

The Pragmatic Engineer reached 189,000 subscribers by late 2022, growing ~1,000/day after Substack introduced recommendations. Paid subscribers are a small single-digit percentage but yield revenue exceeding his Uber comp. The apparent ease of "one email a week" conceals a near-full work week per post: research, draft, feedback, editor pass. He added a second weekly post for deadline pressure — self-discipline without external structure tends to collapse. A hosts-file Python script that kills Twitter and LinkedIn on demand is his most reliable focus tool.

That fast launch rested on six years of prior work: a blog from 2015, a Hacker News hit on code comments, posts on engineering salary tiers that circulated widely. When he announced the newsletter, thousands already knew the byline.

Core advice: build domain depth first — credibility is the product. Set process goals you control (publish monthly) not outcome goals (subscriber counts). Follow pull signals hard when something lands. The structural trap: a personal newsletter can't be sold at fair value and can't pause without revenue loss, so the exit is either building a media company around other writers or running indefinitely.

What it takes to become a top 1% PM | Ian McAllister (Uber, Amazon, Airbnb)

Communication is the foundational PM skill at every level — not as polish but as a test of thinking. McAllister (Amazon, Airbnb, Uber) traces his development to a Microsoft rebuke: asked when something would ship, he gave context rather than a date. He and his first Amazon boss compiled the "Book of Kim" (after SVP Kim Rackmiller): avoid weasel words, answer first then explain, own your mistakes.

For new PMs, the three core skills are communicate, prioritize, and execute. Prioritization generates 5x the impact for a PM with equal resources — covering theme selection, project sequencing, scope of build, and time allocation. McAllister credits his early Amazon success entirely to this, not seniority or technical depth. Execution is largely a team function; the PM's job is to keep the team resourced and moving.

For senior PMs the emphasis shifts. Think big: always ask whether the idea could be larger before locking scope. Earn trust: the currency of product leadership, built by calling shots and hitting them; evasion or missed commitments destroys it. Drive for impact rather than promotion: McAllister spent his first ten Amazon years not raising promotion with his manager, just growing his book of business — promotions followed as a byproduct.

Amazon's working backwards process is a mechanism to enforce problem-first thinking, not a ritual. The principle is to start from a genuine customer problem, not capabilities you happen to have. The most common failure is retrofitting the problem after the solution already exists — McAllister's team once built and shut down "ASIN to ASIN linking," a Bezos-adjacent idea with no customer problem behind it. Bezos's three investment gates: is it a big idea? Should Amazon be doing it? Is there a legitimate plan to succeed? The FAQ stress-tests that third gate before resources are committed.

Leveraging mentors to uplevel your career | Jules Walter (YouTube, Slack)

PM skill development splits into IQ (execution, product sense, strategy, interview performance) and EQ (communication, leadership, self-awareness), and the two require different learning approaches. Jules Walter—first growth PM at Slack, later product lead at YouTube's Primetime Channels—mastered IQ quickly and spent years catching up on EQ.

For IQ: set a concrete outcome first (double-digit activation gains within six months at Slack), work backwards to the knowledge gap, read minimally to sharpen the right questions, then go to the best practitioners. Bangaly Kaba at Facebook gave Walter the understand-identify-execute framework; Walter applied it, reported results, and repeated. Reverse-engineering internal artifacts—strategy memos, exec updates—works the same way. Sitting in on a peer's iteration process reveals the backstage thinking that polished outputs hide.

For EQ: the feedback loop is years, and what to fix is person-specific. Lawrence Ripsher (head of product at Pinterest) helped Walter see a recurring pattern: going quiet under pressure read as disengagement. Ripsher's strength diagnostic—"What do others praise that you dismiss as obvious?"—surfaced asking questions as Walter's core strength, and its shadow side as asking without framing, which read as junior.

Getting honest feedback requires engineering: ask specifically ("did I show executive presence?"), offer self-critical observations for others to react to, respond with visible enthusiasm. EQ feedback—"you come across as angry"—stays in calibration rooms unless you build the trust for it to reach you directly.

Finding mentors: make the smallest possible ask (a two-minute email question, not a coffee request), close the loop by reporting what you did with the advice, then escalate to calls. Walter built his whole roster this way—Bangaly at a Facebook event, Ripsher at a dinner, Bradley Horowitz at a fundraiser, Nikhyl Singhal via email.

Build one skill at a time over six-month cycles; expect EQ to keep demanding work long after IQ feels settled.

Understanding the role of product ops | Christine Itwaru (Pendo)

Product ops exists in two forms: a system any strong PM builds to help their team thrive, and a dedicated role that acts as strategic partner to PMs and advisor to CPOs. The role crystallized around 2019 — not because the underlying problems were new, but because rapid growth, the shift toward product-led tactics, and an expanding CPO mandate (from "ship features" to "drive business outcomes") made it impractical to keep loading these responsibilities onto PMs.

The core case is that PMs' most irreplaceable work is time with customers and engineers. Everything else can be handed off: aggregating voice-of-customer data from CS, sales, and NPS into usable PM readouts; managing the tool stack (Pendo, Salesforce, Looker); treating in-app education and documentation as part of the definition of done; and creating internal readiness so revenue teams know not just what is shipping but what to do with it.

Pendo's origin story: a major launch flopped in Christine's fifth week — not because teams didn't know it was coming, but because they didn't know how to prepare. The fix was a product digest giving revenue teams actionable readiness context rather than just release dates. That transparency-and-readiness framing is not B2B-specific; large B2C companies face the same cross-functional alignment problem.

The line between product ops and product marketing: PMM positions and sells; product ops educates internal teams on value and usage.

Product ops leaders should come from PM backgrounds to quickly locate where effort belongs. Day-to-day practitioners typically arrive from management consulting, customer success, and technical success. The essential temperament is comfort building systems and then letting them go — automating or handing off what's stood up so energy can advance to the next strategic layer.

One high-impact, low-cost change at Pendo: bringing engineers into customer calls. Initially reluctant, engineers quickly wanted more exposure, and their voice carried more weight in planning as a result.

Lessons from scaling Stripe | Claire Hughes Johnson (former COO of Stripe)

Building a company is a separate job from building a product, and most founders start too late. Claire Hughes Johnson scaled Stripe from 160 to 7,000+ people as COO; her argument is that operational structure must go in before you think you need it — companies that wait until 800 people to introduce job levels pay for it in painful retroactive categorization.

Her four personal operating principles scaffold the approach: build self-awareness to build mutual awareness (values exercises, DISC/Myers-Briggs, expose your tendencies to your team); say the thing you think you cannot say (detoxify internal commentary by asking questions and owning observations, not passing judgments); distinguish management from leadership; and use your operating system as a stabilizer when everything else is haywire.

The company-building architecture has three layers. Founding documents — mission, long-term goals (Stripe's: increase the GDP of the internet; advance developer tooling; accelerate globalization), and operating principles — give everyone decision-making context without a standing meeting. The operating system adds OKRs, quarterly business reviews, dashboards, and planning; the main failure mode is cycling through new systems every few weeks, creating a grab bag of half-committed processes. Cadence should be calibrated, not borrowed: Stripe ran six-month planning cycles and shortened QBR frequency to six-week cycles when quarterly felt stale.

On the COO role: fewer than 20–30% of companies have one, and it shouldn't be a default hire. The right signal is a founder who needs leverage on company-building while still driving product. The right dynamic has just enough friction — mutual trust plus willingness to say "I'm not prioritizing that."

On decisions: use an explicit framework (Bezos's type-1/type-2, SPADE, RACI — pick one and commit). If you don't know who the decision-maker is, assume it's you. The career-defining trait is being a force for positive momentum.

How to make better decisions and build a joyful career | Ada Chen Rekhi (Notejoy, LinkedIn, SurveyMonkey)

Unexamined advice is usually bad — not because people are wrong, but because it's rarely contextual. Ada Chen Rekhi's primary decision tool is the Curiosity Loop: send a scoped question to 5–10 people (subject-matter experts plus people who know you well), lightweight enough to answer in minutes, then report what you did. A question naming the specific decision and asking for rationale produces real signal; "what should I do with my career?" produces garbage.

The complementary tool is a values exercise: work through a word list, stack-rank into three to five sentences about what matters now. Ada used this to turn down a high-profile executive role that would have failed her top three values via constant travel and status-chasing — Buffett's inner scorecard versus outer. She traces this to Asian tiger parenting she had to consciously override.

Early-career framework: explore with a thesis until you have real signal, then exploit. Ada's path: Microsoft, Mochi Media, founding Connected (acquired by LinkedIn), LinkedIn's growth team (100M–200M members), SVP Marketing at SurveyMonkey at 27–28. At LinkedIn she told her manager she was there to learn subscriptions and growth, not get promoted — that clarity earned her exactly the roles she needed.

On when to leave: track which direction the temperature is moving, not the current level. On coaching: most people don't need one — a Curiosity Loop beats one opinion, a Reforge course covers technical topics more broadly, a peer tribe outlasts any engagement. Worth it in hypergrowth or for sensitive long-running interpersonal work. Talk to two or three first; half hire the first they meet.

"Eating your vegetables": repeat an uncomfortable skill 10–12 times before concluding you dislike it. Ada forced herself to weekly networking events, 10 business cards per visit, and touching the back wall before leaving. Those relationships became foundational.

How Ramp builds product

Ramp hit $100M ARR in two years with fewer than five PMs and 50 engineers — the fastest SaaS company to that milestone — because velocity is the business strategy, not a cultural aspiration.

Planning evolved across three stages: pre-PMF was two-week sprints with no backlog; post-PMF scaled to one-to-two quarter horizons to coordinate sales and marketing; at 500+ people, bi-annual company OKRs replaced quarterly cycles, cutting planning overhead from 33% of time to a fraction. OKRs cover cross-functional projects only, never performance management.

Strategy is written as seven-field docs — Goal, Hypothesis, Right to Win, Metric, Initiatives, Risks, Long-Term Outcomes — built bottom-up from pod level, then anchored to the financial model. Trade-offs are made explicit on the roadmap to force sharper debates about X vs. Y rather than whether to do everything.

PM, Design, Engineering, and Data all report to the CTO. Splitting product and engineering up to CEO and CTO creates stakeholder dynamics instead of co-founder ones. Teams are organized around business outcomes ("drive 50% of SQLs through outbound automation") rather than product surface, with published contracts covering goal, strategy, roadmap, metrics, UX flows, tech, and stakeholders. Platform teams stay embedded in product pods until they have cross-team wins, then spin out. Teams cap at 5–10 engineers; staffing flexes across reporting lines without destroying manager relationships.

Hiring prizes slope over intercept (learning rate over domain expertise), high agency, and humility. CTO Karim Atiyeh spent year one solely on engineering hiring — sourcing through programming competitions, a campus internship pipeline, and Fortnite. Managers must earn their role as ICs first. Headcount is treated as a last resort: automation first, internal flex second, new hire third.

Building a long and meaningful career | Nikhyl Singhal (Meta, Google)

Career success and career fulfillment are not the same problem. Nikhyl Singhal — mentor to hundreds of PMs at Meta, Google, and Credit Karma — calls the core mistake chasing the next job rather than the skip: the role after next. Plan current work in service of that skip; tell the story first-person, with a named problem, a named skill, and a named headwind overcome.

Early career: pick one of five PM ambiguities to master — craft, market uncertainty, organizational complexity, domain expertise, or growth. Without a specialty, the resume reads as presence rather than contribution. Avoid ex-growth companies: still searching for product-market fit while carrying hundred-million-plus valuations from the zero-interest-rate era. If user pull isn't organic, the equity is likely worthless.

Four causes block promotion. Missing advocacy means changing teams or companies. No open role is structural, not a performance verdict — common after layoff-driven delayering. Impatience hits high performers at leadership levels where impact lags effort by months. The hardest is a gap the individual won't accept, often the shadow of a superpower: the collaborator who can't form opinions with peers, the decisive entrepreneur who won't update on evidence, the storyteller who avoids detail. These shadows are invisible because the strength is the person's self-concept. Contradictory feedback managers dismiss as anomalies is exactly what to reexamine.

Management: the manager is a sidecar, not a driver. Entry requires invitation. Treat the meeting operating system as a versioned product, revised quarterly.

Act three — the back thirty years of a sixty-year career — demands a new North Star before reaching it. Catching the rabbit leaves the greyhound without reason to run. Two arcs sustain people: scaling financial ambition, or shifting toward giving — coaching, community, mission-driven work. The latter is what Singhal plans to devote thirty years to.

The unconventional Palantir principles that catalyzed a generation of startups

Palantir's seven unconventional practices are now traceable through nine unicorns — Anduril, Handshake, Amplitude, Peregrine — and 100+ venture-backed companies.

Forward Deployed Engineering meant living inside customer operations for months. In one engagement that became a $100M+ contract, Palantir ran the analysis themselves and briefed the client's CEO directly, with zero active customer users for the first 6–12 months. The goal: do the customer's actual job with your product until you've accomplished something they couldn't.

Hiring targeted the literal world's #1 for each critical discipline, not the top 1%. The leverage isn't just skill — the #1 person has disproportionately more access to other top people, capital, and media, which cascades across every subsequent hiring generation.

Sales was handled by operators and engineers instead of a sales team, forcing every customer interaction to create genuine mission value rather than close on commission.

Iteration used working code rather than wireframes, in daily cycles — but only where engineers could ship meaningful changes in under 24 hours. The speed made building cheaper than running most experiments.

Features compounded over time. Gotham's "tagging" let analysts write structured entities back from unstructured documents, permanently enriching the data graph for every subsequent user.

No formal PM organization was created. Palantir's first effective CPO was an engineer with no product title. Formal product hiring was deferred until after product-market fit.

The hardest available problem was always chosen first. Solving what competitors treat as impossible becomes proprietary. Impact earns access; access earns contracts — reversing that order is what makes sales feel forced.

How to pass any first-round interview (even in a terrible talent market)

Most first-round interviews fail not from lack of experience but from misaligned preparation. Coach Erika's Minimum Viable Interview Prep (MVIP) — built from coaching 200+ candidates with a 93% placement rate at top tech firms — reframes the first round as a single filter: does the employer want to keep investing time in you?

Four preparation pillars carry the system. First, audit and clean your digital footprint (LinkedIn, GitHub, public writing), then build one decisive reason for wanting the role that conveys what you bring and how it accelerates your career — never compensation or a critique of your current employer. Mine the job description for keywords, put them in a table, map your experience against each one, and use that language in answers; mirroring the interviewer's vocabulary subconsciously reinforces fit.

Second, replace scripted answers to hundreds of questions with deep recall of three to five large, complex, recent projects (past two to four years), documenting every detail — context, decisions, quantified outcomes. This gives memory reliable anchors to draw from under pressure.

Third, the behavioral question section evolves STAR into three formats. Pure questions (70%) call for STAR++ — add what you learned and how you later changed your approach. Theoretical questions (20%) want a structured framework with three to five elements, each detailed, ending with a STAR example. Situational questions (10%) need clarifying questions first, explicit assumptions, a step-by-step plan, a check-in with the interviewer, and a summary.

Fourth, the questions you ask the interviewer often decide close calls. Research the company's competitive landscape, then formulate questions that signal you've done the homework and whose answers genuinely matter to your decision.

Total prep: roughly ten to fourteen hours, most of it reusable across the full job search.

Inside Linear: Building with taste, craft, and focus | Karri Saarinen (co-founder, designer, CEO)

Linear (project/issue tracking for software teams) reached profitability and a net-negative lifetime burn rate — less cash spent than raised — while staying at around 50 people, running no A/B tests, no per-feature metrics, and just one product manager. Karri Saarinen argues this is a coherent system, not a collection of quirks.

On craft: Linear ships features internally within the first week, then to a small cohort (one to ten customers), polishes only before general release. Vercel co-created the roadmap feature this way. Readiness is judged by Karri clicking through every state before launch, not by metrics. An engineer named Andreas independently built macOS-style diagonal safe zones for sub-menus because the culture rewards unsolicited care.

On the absence of PMs: each project has a lead — engineer or designer — who owns scope and communication. The one head of product (Nan Yu, from Mode) sets company-level direction without attending every meeting. Engineers are hired partly on product judgment: can they articulate *why* a choice is better, not just that they dislike it?

On opinionated software: flexible tools force teams to spend time configuring instead of working. Linear provides strong defaults — cycles (automated-schedule time-boxes), a specific workflow — betting that alignment on an opinion beats infinite configurability.

On growth: the first year was private beta, inviting cohorts of about ten users, fixing each cohort's problems before opening to the next. Founders emailed invites personally and followed up a week later. At public launch, nearly every beta company converted to paid. Product-market fit was treated as segment-by-segment — early startups first, then larger companies — following Zoom founder Eric Yuan's advice: double down on the category already working before expanding.

The personal operating principle: go slow to go fast. Rushing produces rework; thinking before building is often faster end-to-end.

The engineering mindset | Will Larson (Carta, Stripe, Uber, Calm, Digg)

Engineers have been coddled — sheltered from hard problems because retention dominated management metrics. Post-ZIRP accountability makes real senior IC paths credible, the core argument of *Staff Engineer*.

Systems thinking (stocks, flows, feedback loops — Meadows's *Thinking in Systems*) is a learning tool, not an operating mode. Stripe's incident team trapped itself: measurement crowded out improvement. Reality always beats the model; that gap is where you learn, but you have to act on it.

Engineering strategy is boring by design. Every company already has one — usually unwritten. Larson uses Rumelt's frame (*Good Strategy, Bad Strategy*): diagnosis, guiding policies, actions. Uber's no-cloud policy enabled launching in China in three months; Stripe's Ruby monolith kept engineers building features instead of tooling. Both hated, both focusing. Bad strategy traces to willful misdiagnosis of constraints.

EM/PM friction usually reflects misunderstood constraints, not bad actors. Larson's fix at Carta: paired ratings calibrated jointly by CTO and CPO. Shared incentives remove the most common failure mode.

Engineering productivity: DORA metrics (*Accelerate*, Forsgren et al.) diagnose where to improve; they don't grade teams for boards. Sprint points are noise. A roadmap of shipped impactful work is the right board artifact.

Values must be honest (you do what you claim), applicable to real decisions, and reversible (a credible company could hold the opposite). Identity values — integrity, high standards — fail reversibility.

On writing: roughly a thousand posts over sixteen years by writing only what's energizing. Two or three polished pieces beats a dead newsletter for career positioning. The real risk is quitting, not starting late.

The Digg V4 rewrite: complete rewrites don't work. Servers crashed for a month; a Python default-parameter bug took weeks to find. The company sold for parts; Larson became engineering manager two and a half years in by being willing to stay.

Taking control of your career | Ethan Evans (Amazon)

Most managers are too busy to actively develop your career — waiting to be noticed is a losing strategy. Ethan Evans, former Amazon VP (Prime Video, Appstore, Prime Gaming, Twitch Commerce, 800-person org), argues growth is in your own hands via a five-step framework: the Magic Loop.

The loop: (1) Do your current job well enough that your manager isn't wishing you were different. (2) Ask how you can help them — almost no one does this, and simply asking builds an alliance. (3) Do what they ask, even if unglamorous. (4) Return and say: "Is there something you need that would also help me reach my goal?" Managers help those who help them; this exchange is the mechanism. (5) Repeat. It works even with mediocre managers.

Senior managers stalling before director face a structural bottleneck — a director may have only six to eight slots. The behavioral gap is equally real: functional excellence gets you to senior manager; director and above requires influence, peer coordination, and letting go of detail. Marshall Goldsmith's title: what got you here won't get you there.

On invention: Evans holds 70+ patents from a simple method — two focused hours monthly, combining things from different domains. His drone-via-roving-truck patent came from crossing delivery logistics with aircraft-carrier thinking. One idea takes years to fully execute; Prime is 20+ years old and still expanding.

On failure: when Evans's Appstore launched broken and Bezos was publicly furious, recovery followed three moves — own it immediately, update leadership hourly with a concrete plan, and meet in person (anger dissipates face-to-face). Two years later Evans was promoted to VP. Jeff Wilke told him he had been close to being fired for prioritizing launch date over reliability — the lesson Evans drew was "fear the New York Times headline."

How to discover your superpowers, own your story, and unlock personal growth | Donna Lichaw (author of The Leader's Journey)

Effective leadership starts not with influencing others but with understanding yourself first — what Donna Lichaw calls leading from the inside out. Ken Blanchard's sequence: lead yourself, then one-on-one relationships, then groups, then the business. Skipping the inner circle produces leaders whose drive evaporates.

The mechanism is the stories we tell ourselves — the brain processes self-narrative as real whether or not it's true. A CEO who believed "nobody takes me seriously" was mis-diagnosing: his senior hires didn't want orders, they wanted a clear vision and room to execute. The fix was co-creating a new working contract, not storytelling training.

Superpowers are best identified not through StrengthsFinder quizzes but by examining three peak experiences — childhood, past decade, how you entered your current work — and finding the energizing theme. One executive labeled "detail-oriented" in 360 reviews traced her peak moments and found the real pattern: connecting ideas and people — which became her identity anchor.

Apparent weaknesses often serve a function. Imposter syndrome in one founder reliably triggered learning bursts — functional at moderate doses; over-application caused burnout and absorbing others' emotional labor. Dyslexia in many founder-CEOs comes bundled with spatial and big-picture thinking that outweighs the reading friction.

Change is introduced through small in-session experiments. An executive ashamed of her silence tried sitting with it for 30 seconds in a coaching call; the experiment surfaced deep listening, not disengagement. Change followed from communicating her processing style to the team.

To surface subconscious goals, Lichaw uses backward planning: imagine the future state in vivid sensory detail, then sketch the journey that led there. Running behaviors through "head, heart, hands" — thoughts, emotions, physical sensation — reveals what rationalization conceals: a client who said "that was fine" while her face turned red learned more from her body than her words.

How to learn the most about a candidate from a single interview question

Questions that bypass rehearsed success stories reveal the most in a 45-minute window. Compiled from ~150 podcast guests, 25 high-signal questions fall into four groups.

Failure and adversity: "Your biggest product flop" (Annie Pearl, ex-CPO Calendly) surfaces honesty about why things went wrong; "worst product you've shipped" (Maggie Crowley, Toast) tests humility; disagreeing-with-management stories (Ethan Evans, Amazon) reveal backbone paired with willingness to commit.

Thinking style: "What does everyone take for granted that you think is hogwash?" (Nikhyl Singhal, Meta) forces genuine opinion and breaks interview-mode performance. Shishir Mehrotra (Coda) uses a teleportation-device scenario as a covert Eigenquestion test: sharp candidates immediately find the two questions that crack the core problem.

Ownership: asking about a non-cherry-picked release (Laura Schaffer, Amplitude) exposes frameworks rather than curated wins.

Character: Andrew Bosworth (CTO Meta) asks what references will say — candidates are accurate about strengths but not critiques. Nilan Peiris (CPO Wise) asks what frustrates them now: what someone couldn't fix predicts future struggles. "What question should I have asked you?" closes by forcing genuine self-disclosure.

The happiness and pain of product management | Noam Lovinsky (Grammarly, Facebook, YouTube, Thumbtack)

Product management rewards advocating for the business over your own position — even when that means killing your project or asking to be demoted. Noam Lovinsky did both at YouTube: in his first months he recommended winding down the product he'd been acquired to build, then asked CEO Salar Kamangar to move him under Hunter Walk. Both moves paid off. The condition is knowing the decision is right for the business and trusting the team's culture to reward honesty over self-interest.

At Thumbtack, triple-digit growth masked a structural problem: pros paid to quote while customers waited 24 hours. When Google algorithm changes caused the first year-over-year decline in company history, recovery required switching to instant quotes, rebuilding the monetization model, and diversifying away from SEO. Sequoia's Brian Schreier called the resulting smile graph the prettiest he'd ever seen. Single-channel growth is structurally fragile; high growth hides supply-demand imbalances that only surface when the channel dries up.

At Facebook's New Product Experimentation team, judging success by "did we find the next Instagram" sets a lottery-win bar. The real value was teaching a large org to work small: experiments with 100 users, direct customer contact without legal intermediaries, unconstrained infrastructure. The incentive system must be redesigned first — semi-annual performance cycles in a 0-to-1 incubator create adverse selection and wrong time horizons. Nike's internal lab, which plugs in Nike distribution only after product-market fit, was the best external model encountered.

Grammarly's durability comes from frictionless ambient value: install once, and it improves every text box across all applications without configuration. Fifteen years of bootstrapped profitability from Ukraine built a culture where engineers ask how their work translates to revenue.

Career principle: seek roles that stretch you, but keep one or two anchor skills solid so discomfort stays productive rather than destabilizing.

Bending the universe in your favor | Claire Vo (LaunchDarkly, Color, Optimizely, ChatPRD)

Career progress comes from knowing exactly what you want and framing it as solving a company problem. When a marketing head departed, Claire Vo drew an org chart with her name at the top and walked in with a full transition plan — and got the job. She told each subsequent CEO directly what role she wanted next. Rule: 0.005% of boss conversations should be about your career; the rest should be producing results.

Organizations are more fluid than they appear — "the universe is bendable to your will." She expanded into running engineering at Color by spotting a scaling problem she was confident she could fix. The CPTO role (product, engineering, design under one leader) eliminates function-versus-function debates and creates a single R&D owner; genuine technical depth is required. Culture principle: "no lanes" — engineers can write specs, PMs can sketch designs.

On pace: don't let meeting cadence set execution cadence. Her instruction is "one click faster" — if it needs to happen this half, make it this quarter; this week, today. She keeps her own SLA short because a slow approver stops everyone below.

On talent bar: define expectations specifically and measurably, normalize candid feedback ("clear is kind"), and act quickly when someone isn't a fit. She told two feuding leaders their behavior was not meeting expectations and they would not survive unchanged — both turned it around within months.

ChatPRD (chatprd.ai) began as a tuned prompt used to spec a platform feature in real time. It became a standalone app with per-user customization; 60% of users go idea to PRD, 30% improve existing specs. She sees AI replacing lowercase-c communication — mechanical coordination and synthesis — while capital-C influence and forward-looking judgment stay human. Contrarian view: sales-led product organizations are unfairly stigmatized. Life motto: fast beats right.

A framework for PM skill development | Vikrama Dhiman (Gojek)

PM career growth depends less on which product area you work on and more on three dimensions — what you produce, what you bring to the table, and your operating model. Strong PMs excel at two of three; those rising fast perform well on all three.

What you produce follows a progression: outputs first (shipping features, drafting briefs, ranking options), then outcomes (ownable product goals), then directional leadership. The critical mistake is skipping ahead or dropping output discipline as you rise. Dhiman watched a peer win the bigger role despite a less prominent area because that person never abandoned IC craft: precise launches, tighter design collaboration, cleaner experiments.

What you bring to the table is your impact on impact. PMs are evaluated on four axes — data, design/research, technology, strategy — but must demonstrate them daily through artifacts: PRDs, product notes, experiment designs, design briefs, pre-sprint planning docs. When a PM's product does well but career doesn't move, the diagnostic is: show me your last PRD. Something is usually missing or thin.

Operating model — how you work with others — matters most at the mid-to-senior transition. Three mantras: raise difficult issues without being difficult to work with; surface important topics without drawing importance to yourself; get decisions made without making all the decisions. The key skill is pulling an emotionally charged conversation back to a logical register.

Three mindsets stall growth: drifting focus outside your control; slowing your rate of skill change as seniority rises; and the identity stories you tell yourself. "High-agency" can rationalize brashness; "collaborative" can rationalize indecision. Identify the story and correct it.

Skill sequencing: pair data/technology with design/research early, then add strategy. Two contrarian positions: good intent alone is insufficient — behavior must match; and hours invested correlate strongly with skill growth.

The paths to power: How to grow your influence and advance your career | Jeffrey Pfeffer (author of 7 Rules of Power, professor at Stanford GSB)

Political skill — networking, brand-building, projecting power — predicts salary, promotion speed, and job satisfaction more reliably than almost any other factor. Gerald Ferris established this empirically. Jeffrey Pfeffer has taught it at Stanford GSB for 50 years as Paths to Power, organized as seven rules.

Get out of your own way. Believing power is dirty causes people to skip the behaviors they need. Preemptory apology ("I don't know if this is useful") signals self-doubt that others immediately adopt.

Break the rules. The unexpected creates memorability. Jason Calacanis defies VC conventions — many small bets, no partners, a journalist's questioning habit — and became extremely wealthy.

Appear powerful. Eye contact, no notes, louder voice, open posture, and humor are learnable. Regis McKenna's team turned Steve Jobs from someone who "couldn't convince you to buy water if you were dying of thirst" into a person who could sell anything.

Build a personal brand. No one promotes someone they don't know. Laura Chau made VC partner at Canaan in four years through writing, podcasting, and dinners. Tristan Walker got hired at Foursquare by closing Starbucks as a partnership before anyone offered him a job.

Network through generosity. Omid Kordestani spent a year at Netscape ignoring his job to network across Silicon Valley; when Google polled everyone for the best business person, his name appeared on every list — employee 11, $2.5 billion. Mark Granovetter's research shows the best jobs come through weak ties, which reach non-overlapping parts of the market.

Use your power. Results attract resources; deploying authority is self-reinforcing.

Success excuses almost everything. Martha Stewart's post-prison brand, Gates's appropriated code, Epstein's continued royal dinners — proximity to power makes people overlook past behavior.

Pfeffer's own tradeoff: he chose autonomy over institutional power. "You can have power or autonomy, but not both."

Making an impact through authenticity and curiosity | Ami Vora (CPO at Faire, ex-WhatsApp, FB, IG)

Genuine curiosity about disagreement is the most durable leadership tool Ami Vora developed at Meta. CTO Boz describes her responding "fascinating, tell me more" to ideas she found wrong; her explanation is unsentimental: she loved being right, and a manager showed her she was sacrificing better outcomes for ego. The reframe: curiosity produces better decisions, and being collectively right outweighs being individually right.

At product reviews, she coaches against "dinosaur brain" thinking: leaders hold three facts and pattern-match; they cannot replicate deep analysis. The compact: "my manager owns context, I own the recommendation." Reviews should extract principles, not just decisions, so teams stop escalating and decide independently.

The "hill climb" metaphor names a structural problem: crossing from a local optimum to a better one requires going through a valley where progress looks like regression. Knowing you're in the valley is what makes it survivable.

On execution vs. strategy: perfect strategy with poor execution teaches nothing — you can't tell what failed. Good-enough strategy with strong execution gives a learning machine. Spend 20% on strategy, 80% confirming it with customer feedback. Even senior leaders mostly execute, just at system level.

Org design failure she calls "toddler soccer" — everyone chasing the same metric. The fix: assign each team a distinct input metric that ladders into the shared output.

Her gender observation: feedback to shrink and become "less objectionable" produces short-term wins while hollowing out leadership. Expand your toolkit, not compress your identity.

At Faire she joined as CPO after 15 years at Meta partly to test herself outside that scaffolding. She wrote quarterly "hot takes" for the CEO — provocations about what could change — building trust faster than conventional onboarding. Closing point: stay close to the customer; virtually everything else in a PM's job is distraction from that signal.

On asking for help (even when you really don’t want to)

Refusing to ask for help is not self-sufficiency — it is a hidden tax paid in burnout, isolation, and slower work. Executive coach Natalie Rothfels argues that high-performing product leaders, particularly PMs who internalize "radical ownership," are especially prone to this trap, and that asking well is the single highest-leverage interpersonal skill for career trajectory.

The framework has three steps. First: navigate the fear before making any ask. The six fears she names — burden, incompetence, rejection, loss of control, uncertainty, and wasted vulnerability — are real but distorted. Her process moves through awareness (an emotion log tracking when fear appears), allowance (naming it without debating it), a narrative reframe ("knowing what I don't know is a leadership competence"), and small experiments that generate disconfirming data.

Second: make the ask with a six-part template — signpost, clear request, rationale, why this person, timeline, opt-in. Make yes easy: establish relational connection first, start small, frame the request around the recipient's goals. She also names five types of help to clarify what you need: perspective, information, task progression, empathetic support, or advocacy.

Third: receive the help properly. Follow through with impact and reciprocity — it reinforces the relationship. When help misses, redirect specifically. When they decline, keep the door open and ask for a referral.

Client Meena, a GPM near burnout who had never considered asking for help, went through all three steps and was subsequently asked to expand her scope to two additional teams.

How embracing emotions will accelerate your career | Joe Hudson (executive coach, Art of Accomplishment)

Suppressing emotions doesn't protect you from them — it guarantees you'll keep experiencing them. Joe Hudson, an executive coach to leaders from OpenAI, SpaceX, and Apple, identifies two blocks that hold ambitious people back: a destructive inner critic and an inability to feel emotions.

The critical voice repeats shame-based pressure rather than helping. Silencing it doesn't work. Treat it as a frightened child — "I see you're scared, I've got you" — or experiment daily with different responses. The brain's habenula punishes repeated failure, so framing each response as an experiment removes the failure condition.

All decisions are made emotionally; logic just predicts how we'll feel (Damasio's *Descartes' Error*: emotional-center brain damage destroys decision-making even with intact IQ). Emotions you avoid cost you solution sets. A conflict-avoidant CEO ends up presiding over exactly the unresolved tension they feared. Whatever you resist, you invite in. The goal is active welcome, not non-judgmental awareness: "Joy is the matriarch of a family of emotions and she won't come into a house where her children aren't welcome." Release also requires physical expression — all mammals shake to discharge fear.

Self-improvement framed as "I should" produces shame loops and stagnation. "I want to" carries aliveness. Fully understanding a problem dissolves it.

Enjoyment is a productivity mechanism. Ten percent more enjoyment yields 10% more efficiency and better quality. The experiment: ask "how can I enjoy this 10% more right now?" — releasing internally, not changing externally.

For teams, the diagnostic unit is meetings. Five-star meetings surface every organizational problem and typically halve meeting count within two months.

The single recommended experiment: seven minutes of felt gratitude daily with another person — not a list, but speaking from the physical sensation. Applied to areas of lack, Hudson credits this with clearing $40,000 in personal debt within months.

Time management techniques that actually work

Ten tactics that outlasted years of productivity-book trial. Put to-dos on your calendar as events — things not on the roadmap don't happen. Do anything under two minutes immediately (David Allen's GTD rule: tracking costs more than doing). Keep a "waiting for" list so delegated threads stay visible. Each morning write 1–3 must-do items before opening email; do the hardest one first. Block 2–3 weekly deep-work slots and guard them — Cal Newport: real output needs 30–60 uninterrupted minutes. Push meetings as late as possible. Keep Do Not Disturb on permanently. Convert meeting requests to async email by default. Delegate low-value tasks to a virtual EA. Say no by filtering asks through 2–3 core priorities, naming the tradeoff explicitly, and asking "would I want this tomorrow?"

Unorthodox PM wisdom: Automating user insights, unselling job candidates, logging every decision, more | Kevin Yien (Stripe, Square, Mutiny)

Don't go straight into product management — start as an engineer, designer, or salesperson first. The PM role converts a team's potential energy into realized customer value; that job is clearest after front-line work. Not every team needs one: companies building for themselves (Stripe, Figma) famously delayed.

The PM's job is to draw the perimeter — apply constraints on who the customer is and which tradeoffs apply — then let engineers and designers fill the box. At Square, Yien and a designer spent a week tuning a restaurant POS menu animation to millisecond precision to match bartender muscle memory. PM territory: stay obsessed with the final deliverable.

Writing is clarity at scale. A PM who can't write compelling messaging in the customer's voice can't be trusted to build the product. Read writing that compels action, not PM artifacts.

Product sense is the ability to make good decisions with insufficient data. Build it through a decision log: tag judgments (#decision), write the rationale, set a reminder to check. Log what you'd predict other companies will ship, then compare. When Shopify launched the Shop app, Yien mapped the flywheel publicly; Shopify employees confirmed the call that day.

The unsell email fixes candidates who leave within six months because reality didn't match the pitch. After the full interview — when unstated fears have surfaced — Yien sends up to eight bullet points naming the hard truths. He lost 30% at offer stage the first time. Those were the ones who would have left anyway.

Automate user research by routing Gong keyword alerts through Zapier into a customer.io sequence with a Calendly link — qualifying interviews book themselves. Sales calls are a research stream PMs underuse. Never accept "bent glass" — pre-processed reports or secondhand notes. Direct exposure to raw customer material is non-negotiable.

Land your dream job in today's market: negotiation tactics, job search councils, and more | Phyl Terry (Author, "Never Search Alone")

Job searching is a marketplace problem, not a personal one. When tech contracts, a director competes against laid-off VPs — the mismatch is supply and demand, not personal. Phyl Terry's framework applies product-market-fit logic to the search: a candidate who'll take anything is forgotten the moment the conversation ends.

The method has two phases. First, write a "Mnookin two-pager" — what you want and don't want — then run a listening tour of 10–15 conversations with former colleagues and recruiters. Ask each: "If you were in my shoes, how would you approach this?" People become invested and turn into active listening posts. From this, distill a three-to-four-attribute candidate market fit statement: role, level, industry, stage. One chief data officer spent a year going nowhere, narrowed to mid-size regional bank CTO, and had three offers in three weeks. The narrowness is counterintuitive but necessary — people expand from a specific statement; they never reduce from a vague one.

Second phase is interview and negotiation. Before salary, write a private job-mission document with OKRs for the role, share it with the hiring manager on a live call, and invite corrections. Every hiring manager struggles to distinguish good talkers from people who execute; presenting OKRs resolves it. One Amazon executive who has hired 2,000+ people said no one in his career had done this. Then negotiate success conditions — budget for tech debt, headcount, training — before base salary. Simply asking "Are you open to X?" gets a yes the large majority of the time.

Both phases hold up better inside a Job Search Council — a free, volunteer-run group of six to eight peers meeting twice weekly, providing accountability and countering the emotional erosion that stalls searches. Average council search time is three months against a national average of three to six.

A PM's guide to influence

PMs who can't order anyone still move controversial decisions by treating alignment as an engineered outcome. Jules Walter (YouTube, Slack, Gemini) runs five tactics. First, map each stakeholder before engaging: their OKRs, who they trust, what they fear. At YouTube, he won a reluctant partner team by tying his forecast to their metric and suggesting which of their projects to deprioritize. Second, frame from the other side's POV: when Slack CEO Stewart Butterfield resisted a monetization experiment, Walter led with user-benefit data ("we're hiding value"), not revenue targets — it lifted paying teams by 20%. Third, hold a "meeting before the meeting": debrief detractors to steelman their objections, then brief champions to rebut them; his YouTube pitch ran 14 pre-meetings and placed top-3 at the exec summit. Fourth, paraphrase objections back so people feel heard before they'll listen. Fifth, state the desired decision at the start and redirect tangents to protect decision time.

Why cash is king

In a 5,000-person compensation survey of tech workers, 75% ranked salary above equity — a striking reversal of the "take more equity" orthodoxy. Post-Covid inflation, layoffs, and repeated equity disappointments have turned RSUs and startup options into "Monopoly money" for most respondents. The core driver is stability: salary pays rent; equity may never materialize. Equity preference clusters predictably — men are 2.3x more likely than women to prefer it; each seniority step doubles the odds; US workers are 1.75x more likely than non-US workers; early-stage employees lean equity while public-company employees lean salary. Founders seeking talent can't rely on equity alone — pairing a solid base with a credible liquidity story is what moves candidates in 2024.

Why most public speaking advice is wrong—and how to finally overcome your speaking anxiety | Tristan de Montebello (CEO & co-founder of Ultraspeaking)

Speaking anxiety stems from treating speech as a conscious, controlled process rather than the subconscious flow it naturally is. Tristan de Montebello, who reached the World Championship of Public Speaking finals in 2017 with no prior experience, built Ultraspeaking on that diagnosis: trying to control words before they leave your mouth blocks the natural hardware you already possess.

Three tactics address root causes rather than symptoms. "Think up" — looking upward while gathering thoughts — makes pauses read as confidence. "End strong" counters the fMRI-documented phenomenon where the brain exits flow just before the finish line; summary prompts ("so to wrap up…") close that gap. "Stay in character" exploits the fact that audiences cannot detect internal anxiety unless speakers announce it — voicing insecurity forces a critical lens onto everything before it.

Practice works through games that create turbulence. Conductor flashes intensity numbers (1–10) while the speaker responds to a random prompt; energy leads, emotion follows, words fill in. Triple Step injects random words mid-speech to be integrated seamlessly, building resilience to distraction. Conviction Prompts inserts openers like "I genuinely believe that" or "it astonishes me when," pulling speakers into executive-presence states rather than hedged speech.

For prepared talks, the Accordion Method replaces the write-memorize loop: speak at full length, compress progressively (three minutes → two → one → thirty seconds), then expand back up. By the bottom, only essential structure survives; by the return to full length, the talk is internalized rather than memorized, flexible across time constraints, and immune to a single dropped line.

The Bow and Arrow frames the design question: audiences retain at most one thing, so identify that single sentence (the arrow), then build supporting anecdotes and data (the bow) to drive it home. Applied per slide, it eliminates slides crowded with competing ideas.

How much do U.S. product managers really make?

Senior ICs at U.S. tech companies now out-earn people managers: median base salary for senior ICs is $187,500 versus $162,500 for managers — a $26,920 gap that inverts the traditional assumption. The IC track compounds significantly: entry-level starts at $112,500, mid-level reaches $137,500, and senior ICs land at $187,500. Moving into management produces an actual salary dip; senior managers only catch up at the director level and above.

The strongest predictor of salary is job level, not years of experience — a PM with ten years could sit at IC 3 or IC 5 depending on the company, making tenure a misleading benchmark. Satisfaction correlates only moderately with total compensation (r = 0.31); negotiating higher comp showed no significant effect on satisfaction at all.

How to break out of autopilot and create the life you want | Graham Weaver (Stanford GSB professor, founder of Alpine Investors)

Most people never consciously choose their lives — 95–98% of thoughts are subconscious, shaped by parents, media, and social pressure. The diagnostic: if a genie guaranteed that whatever you threw yourself into would work out, what would you do? That answer surfaces the goal people carry but are talking themselves out of. A Brazilian MBA student set to return to consulting answered: start an education nonprofit in his home country. Writing down limiting beliefs breaks the block — "how would I fund this?" is paralyzing as subconscious fear; on paper, a task.

Two principles reframe what change feels like. Life is suffering — choose something worth suffering for: Weaver ran Alpine for 14 years before being confident it would survive, doing the same late nights he'd have done in a job he didn't care about. Everything you want is on the other side of worse first: every change has a negative first move, and optimizing for tomorrow locks you at a plateau.

The nine lives exercise: name nine lives you'd find genuinely exciting, all starting today. You can't live them at once, but pulling the highest-energy one in as a side practice raises energy everywhere.

Accountability is the mechanism: executive coach or committed peer, weekly goal-writing, three daily actions. Talking activates broader neural regions than thinking alone, making the social format superior.

After Alpine's first liquidity event in 2015, Weaver felt euphoric for two days, then noticed nothing internally had changed. Joy is constructed, not delivered by achievement. The quit signal is not hardship but absence of green shoots — evidence the vision is still becoming real. Alpine's PE thesis emerged by scaling bright spots: three unlikely top performers shared one trait — Alpine had installed its own management. Good-enough industry plus world-class leadership became the formula across 600 deals.

The definitive guide to mastering product sense interviews

Product sense interviews fail candidates not from lack of ideas but from lack of structure. Interviewers are busy PMs filling a standardized rubric who collect specific signals across five dimensions in roughly 35 minutes. A weak score on any single dimension fails you regardless of the others.

The five dimensions run in order: clear communication, product motivation, segmentation, problem identification, solution development.

Clear communication comes first. Open with 2–4 scoping assumptions (role, geography, platform) and state your game plan before touching content. Waypoint at each transition — pause, walk through your section, check in. Never ask the interviewer what to do next.

Product motivation (3–5 min) connects the product to a deeper human need, situates it competitively, and closes with a mission statement specific enough to guide later decisions but broad enough to preserve solution space. Describing features without purpose fails this dimension.

Segmentation (8–10 min) maps the full ecosystem — supply, demand, partners — picks one group with explicit rationale, then defines three segments by behavior and motivation, not demographics. Test mutual exclusivity: if a user fits two segments, refine until they can't. Prioritize one on reach vs. underserved degree and build a named persona with concrete constraints.

Problem identification (8–10 min) maps a day-in-the-life journey — not a generic pre/during/post arc — to surface pain points. Needs are desires; problems are obstacles. Prioritize one by frequency × severity and connect it back to the mission.

Solution development (8–10 min) generates three meaningfully different approaches, places them on an impact-vs-effort grid, selects one, scopes a v1 with go-to-market logic, and names risks with mitigations. Closing back to the mission completes the narrative arc interviewers look for.

Deliberate practice across 20–30 questions with the template internalized — not just read — is what separates candidates who pass from those who don't.

Become a better communicator: Specific frameworks to improve your clarity, influence, and impact | Wes Kao (coach, entrepreneur, advisor)

Most writers waste space on the what and why when readers already agree with the premise and want the how. The "super specific how": skip justifying why communication matters to PMs; explain how to get buy-in without positional authority, or how to present a hypothesis that might be wrong. Cut backstory the same way — start "right before you get eaten by the bear," dropping everything before the moment that matters.

Her content hierarchy of BS ranks formats by how much undefended assertion they allow. Twitter sits at the bottom (mic-drop, no rebuttal possible); cohort-based courses sit at the top because students push back live in Zoom chat. Long-form articles and books fall in between.

The state change method addresses monologue fatigue: every three to five minutes, shift what the audience is doing — Zoom chat, a poll, screen-share toggle, breakout rooms, asking them to guess before you reveal. Nathan Barry treats it as a rule: one state change every three to five slides.

Managing up is not junior behavior — senior people do it best. The practice: proactive context-sharing calibrated to decision reversibility, structured BLUF-style (bottom line up front, context below), biased toward over-communication in remote work. A weekly "state of me" email covering priorities and blockers is one implementation.

Saying no: reframe as trade-offs — "yes I can do this; that means deprioritizing X, which do you prefer?" — converting a team-player judgment into a prioritization conversation.

Writing: unintentional leading — a paragraph that steers toward a conclusion you didn't mean to endorse — is a technical error. Pros-and-cons lists imply objectivity; state your recommendation first and let the cons be honest risks. Books on craft over mimicry: *It Was the Best of Sentences, It Was the Worst of Sentences* (June Casagrande) and *Better Business Writing* (HBR Press).

The ultimate guide to negotiating your comp

Two equally qualified candidates walk into the same negotiation; one leaves with double the comp. The gap isn't credentials — it's a shift from "what can you offer me?" to "here's how I'll solve your problems." Jacob Warwick, who guided 3,500+ senior leaders, structures this as GAINS.

Gather intelligence others miss: map who influences decisions (one SVP lost an $800K role by dismissing a Chief of Staff the CEO deferred to on every hire) and surface pain points the company won't state. A VP of Product Design doubled a $600K cap to $1.2M by uncovering that leadership wanted customer enthusiasm rather than roadmap execution, then painting the post-hire future.

Align with key people rather than impressing them. Diagnose problems as a peer, share the messy middle of past failures, ask bold questions others avoid, and "sell the vacation": a 30/90/180-day picture after joining. One leader shifted a range topping at $285K to $1.1M by reframing around their aspiration to be a product visionary.

Influence broadly by delivering value before being hired. One candidate sent an unsolicited UX prototype in follow-up and received an offer 30% above initial. Converting one skeptic into an advocate can unlock a 40% comp increase when negotiations stall.

Navigate with patience — haste equals risk. Use questions instead of demands, signal enthusiasm while keeping space open, leverage competing offers transparently.

Secure rigorously: get an employment attorney, nail every vague term in writing. One client lost $20M because verbal equity promises went undocumented; another delayed their start to define metrics and captured $1.2M in equity acceleration on acquisition.

Carol's case: a 16-month campaign during a CEO transition — building cross-functional champions and creating options rather than reacting — produced an $800K signing bonus plus $500K equity at a new company, more than twice her previous total.

How Revolut trains world-class product managers: The “local CEO” model, raw intellect over experience, and a cultural obsession with building wow products | Dmitry Zlokazov (Head of Product)

Revolut produces disproportionately many world-class product leaders — alongside Palantir and Intercom in Lenny's triangulated alumni data — because of three mutually reinforcing practices: radical ownership, a cultural obsession with "wow" products, and hiring for raw intellect over résumé experience.

Product managers are called "product owners" and treated as local CEOs, line-managing engineers, designers, data analysts, and ops people on fully cross-functional teams, with end-to-end accountability for customer outcomes and business metrics. Depth is non-negotiable: product owners sit with engineers reading code, reverse-engineer regulatory compliance across 50 jurisdictions, and build scalable frameworks from those specifics rather than handing off complexity to specialists.

The "wow" obsession is operationalized by having founders Nick and Vlad review 100% of shipped screens in weekly product reviews with every team. The logic: a scrappy MVP makes it impossible to distinguish a bad idea from a bad product. By mandating high-quality UX even on narrow first versions, Revolut removes that ambiguity. Feature scope can shrink; aesthetics and polish cannot.

On hiring, Dmitry Zlokazov argues experienced professionals from established companies often rest on their laurels and resist the toil required to change the status quo. Revolut targets hunger and raw intellect instead — early-career candidates, startup founders, or internal transfers from engineering and operations who already proved culture fit. After 400–500 interviews, this profile consistently outperforms pedigreed hires. Sourcing targets teams behind products Revolut itself admires.

A counter-intuitive management practice: out of roughly 100 simultaneous projects, Zlokazov picks 7–10 to scrutinize at code depth. The signal disciplines the rest — ambitious people in a high-talent-concentration environment self-manage to that bar knowing any project could be next. The execution maxim: if something is 99% done, it's closer to 0% than 100% — because shipping without enabling customer care and marketing to use it makes it useless.

Unconventional product lessons from Binance, N26, Google, more | Mayur Kamat (CPO at N26, ex-Binance Head of Product)

Binance went from zero to a peak $400B valuation with 2,000 employees in five years by eliminating management layers. CZ had 55 direct reports; the entire leadership team met at 11 PM Singapore time every night including weekends, so no decision was blocked more than 24 hours. Urgent problems got a single owner at that call, who ran a daily all-hands until resolved. For KYC onboarding, this meant tracking conversion across 500 country-document combinations with a team scaled from 20 to 500 in three months — CZ's rule: no user left behind, including one in Congo.

Product strategy is overrated. Jonathan Rosenberg's rule at Google: "Come to me with data. If you come with ideas, we'll go with mine." For anything testable, the only real strategy question is how fast you can go from hypothesis to data. Experimentation platforms like Statsig replace slide decks and solve the PM credibility problem: six rounds of data showing something failed beats any executive's intuition.

Four career levers: join fast-growing companies where learning compounds like daily versus annual interest; work in roles that use your strengths rather than ones you're trying to fix; don't optimize for early comp (90% of lifetime earnings come later); decide early whether you genuinely want the C-suite — if the word "balance" comes up, you probably don't.

The "moving desk" principle from White Pages CEO Alex Allgood: physically relocate to whichever team has the highest-leverage problem until it's solved, then move again. A full calendar is a badge of shame, not honor.

Biggest failure: the first Google Hangouts PM role. Thousands of engineers, full resources, Larry Page involved — still couldn't ship a great messaging product. Diagnosis: Google's DNA resisted social the way Microsoft's resisted mobile. The team did invent WebRTC, which now underlies every video call on earth.

How tech workers really feel about work right now

Nearly half of tech workers (44.7%) are significantly burned out in 2025, yet 54–58% remain optimistic about their careers and roles. Sentiment has shifted measurably negative over the past year; only founders and new grads have grown more optimistic.

Manager quality is the single strongest lever: only 26.6% rate their manager highly effective while 42.3% call them ineffective. Workers with great managers report 62% more job enjoyment, 63% more belonging, 48% more engagement, 31% less burnout, and are 4.3x less likely to be at attrition risk.

Founders stand apart on every metric — highest optimism, lowest burnout, most engagement — pointing to autonomy and ownership as core drivers. Small-company employees outperform large-company peers on 9 of 11 sentiment measures, with belonging showing the widest gap. Mid-career workers (7–14 years) hit a clear slump: highest burnout, lowest commitment, least career clarity. Hybrid arrangements produce the best day-to-day satisfaction; in-office workers feel marginally more optimistic about long-term prospects.

AI anxiety pervades open responses — workers struggle to separate hype from reality and fear role obsolescence, while a notable subset is pivoting toward founding their own products as a hedge.

35 years of product design wisdom from Apple, Disney, Pinterest, and beyond | Bob Baxley

Software is a medium — like film, music, or books — and treating it that way changes everything. Every confused user at an airport kiosk is a failure someone caused; those who caused it never see who's on the other side of the glass. Building well is a moral obligation.

Design isn't a function or aesthetic output — it's imagining the future you want and taking steps to make it real. "Design-led" doesn't mean "designer-led." Apple, Airbnb, and Patagonia are design-led because that thinking is in their founding DNA; Baxley has never seen it grafted on afterward. Apple ran a global store across 30-plus countries with six designers because clear vision eliminates ambiguity. The Mac had 20 people on it; the iPhone patent names 24. Small is the model, not the committee.

The mechanism for that clarity is tenets, not principles. Principles ("be clear," "be simple") are non-arguments nobody disputes. Tenets resolve recurring forks: at ThoughtSpot his three were "documentation is a failure state," "start simple, let users opt into complexity," and "the whole product should feel like it came from a single mind." Each ended a repeating debate and let dispersed teams make aligned decisions without supervision.

On process: wait as long as possible before drawing. The first mark collapses the possibility space — high-fidelity prototypes redirect feedback toward color and typography rather than the concept. ThoughtSpot used "block brain diagrams," chunky low-fidelity layouts, for weeks, then converted to polished comps in a day once conceptual work was settled.

Apollo illustrates the point: John Houbolt bypassed NASA hierarchy to champion Lunar Orbit Rendezvous — the weight-optimized approach that got humans to the moon. His memo opened "a voice in the wilderness." Great ideas need champions willing to risk their career; advocacy must be for the idea, not the self.

How tech’s most resilient workers handle burnout

Interviews with ~175 low-burnout tech workers show that avoiding burnout means structurally designing your career so it becomes unlikely—not managing stress after it arrives. This yields the ARMOR framework.

Autonomy: audit your calendar, earn trust by delivering, and overcommunicate your working style so others leave you alone. Rock-solid boundaries: set hard time- or task-based stops, automate after-hours responses, reframe every "no" as a tradeoff, and take breaks that are rhythmic, reflective, or restorative. Maintenance rituals: treat sleep, exercise, and nutrition as non-negotiable infrastructure; track via simple heuristics (three disrupted nights = crisis); use therapy as preventive care. Original thinking: burnout is organizational dysfunction, not personal failure—reject the "effort = outcomes" equation and let yourself feel bad rather than optimize feelings away. Role architecture: evaluate jobs by four tests—relationships, Sunday gut-check, net energy, and values alignment.

The most-cited single choice: prioritize people, at home and at work.

The definitive guide to mastering analytical thinking interviews

Analytical thinking (AT) interviews don't test for correct answers — they evaluate a structured thought process, and candidates without a framework fail to generate the signals interviewers need.

The five-step structure: First, narrow scope by stating assumptions (product, geography, platform, user type) and sharing a game plan for the 35 usable minutes — this signals preparation and prevents wasted detours. Second, ground everything in product rationale: the problem solved, business model, maturity stage, competitive positioning, and a mission statement. For Spotify, that means the piracy-access problem, freemium revenue, late-growth positioning against Apple Music and YouTube Music, and a mission that anchors all subsequent metrics decisions.

Third, build the metric framework by mapping ecosystem players (Spotify: listeners, creators, advertisers, platform), the value each receives, the actions they must take, then specific implementable metrics with time frames. The North Star metric must grow indefinitely — avoid ratios and averages, which can look healthy while the ecosystem shrinks. "Total streaming hours per week" works for Spotify; pair it with a guardrail preventing passive-platform drift. Fourth, make an altitude shift to a team-level goal for the next 3–6 months: pick the ecosystem player with highest NSM leverage, map their journey, score opportunities on impact and ability to influence, then commit to one recommendation. Fifth, evaluate tradeoffs decisively — name the common benefit, identify the crux, state your choice with rationale tied to mission and metrics, and specify what would change your mind.

Debugging questions require multiple hypotheses across segments before narrowing. Estimation questions require a decomposition formula with round-number assumptions sanity-checked at the end.

Brian Chesky's secret mentor who died 9 times, started the Burning Man board, and built the world's first midlife wisdom school | Chip Conley (founder of MEA)

Older and younger brains are different cognitive tools, and mixing them is underused leverage. Chip Conley joined Airbnb at 52 — average employee age 26, reporting to Brian Chesky 21 years his junior. Younger brains run on fluid intelligence: fast, focused, linear. Older brains develop crystallized intelligence: holistic, dot-connecting, peripherally aware. Conley blocked a mobile-only pivot by asking whether older hosts could actually manage their listings from a phone. He warned Chesky early that avoiding occupancy taxes was unsustainable; the decade of regulatory battles in New York and other cities confirmed it. His label for the role: "modern elder" — as curious as they are wise, mentor and intern at once.

Working for Chesky had three friction points: the assumption everyone could keep founder-mode hours; a Steve Jobs–inflected belief the CEO's product intuition superseded others'; and stretch goals calibrated to be unreachable, generating stress when missed. Conley's fix: open every meeting by establishing shared agreement on what success looks like before the agenda derails.

His Peak framework — a Maslow-derived employee pyramid — puts compensation at the base, recognition in the middle, meaning at the top. At Airbnb this yielded the insight that the company wasn't in home-sharing but in "belonging anywhere," which reoriented host training and all marketing.

An antibiotic allergy caused nine flatlines over 90 minutes — visions of birds telling him to slow down. That near-death experience preceded selling Joie de Vivre and joining Airbnb. The arc led to the Modern Elder Academy: five-day workshops in Baja and Santa Fe, 7,000 graduates from 60 countries, built on three claims. A pro-aging mindset adds 7.5 years of life on average (Becca Levy's Yale longitudinal study). Life satisfaction bottoms between 45–50 then rises steadily into one's eighties — the U-curve of happiness. Midlife is not crisis but chrysalis.

The one question that saves product careers | Matt LeMay

Most product teams get laid off not for breaking best practices but for doing "work around the work" — Daniel Ek's phrase from Spotify's 2024 layoff announcement. LeMay's diagnostic: would you fund your own team if you were CEO? Most PMs he asks can't answer confidently, and that hesitation is the risk signal.

The failure mode he calls the Low Impact PM Death Spiral starts when teams choose safe work — UI polish, incremental features, cosmetic changes. Each team adds rhinestones to the car until the hood is too heavy to reach the engine. Coordination layers multiply, high-impact work gets harder, and teams sink further into low-impact work until the next layoffs.

The antidote has three steps. First, set team goals no more than one step from company goals — not cascaded through OKR levels until the arithmetic can't be traced back. At a FinTech company, LeMay's team found that converting single-product users to multi-product users carried a known revenue multiplier; they set that as their goal and leadership got it immediately, no deck required.

Second, keep impact visible at every step. One growth team spent a full OKR season scoring objectives and key results — then couldn't say how any of it contributed to the company's million-user goal. Their leader put "1 million" on a whiteboard: if a conversation doesn't start here, don't have it.

Third, connect work to impact in the same unit as the goal. Don't use abstract ICE scores — translate impact into users converted or revenue generated. The uncomfortable implication: the highest-impact option is usually the highest-effort, most dependency-laden, and you have to pursue it anyway.

On executive pushback: don't say yes or no — present options with trade-offs and a recommendation, converting a confrontation into an information exchange the decision-maker can use.

The art of influence: The single most important skill that AI can’t replace | Jessica Fain (Webflow, ex-Slack)

Executives operate like a strobe light — finance, legal, people problems, and product reviews back to back, with no time to revisit your proposal since the last meeting. The core mistake PMs make is forgetting their user-empathy skills the moment they sit across from a leader. Treat the executive as your key user and their buy-in as product-market fit for your idea.

Several tactics follow. Spend 30–60 seconds at the opening to reground context, state goals, and ask "anything else you hoped to cover?" Match the format to how that exec absorbs information — design mock, data dashboard, or silent doc-reading time. Present at least three options, not one; showing the alternatives you considered closes the gap when an exec says "I don't see how you got here."

Before pitching, surface the executive's real incentive structure. "What's the board pushing you on?" unlocks more signal than "what's top of mind?" — it reveals pressures beneath polished weekly updates. Connect your proposal to the OKR they own, and build cross-functional groundswell — if customer success, marketing, and engineering peers all advocate independently, the exec is far less likely to refuse.

Pick up subtle threads fast. When an exec says "I wonder if we should look at design reviews," the best responders have a framework back within an hour. Waiting a week lets the moment die.

Trust compounds from killing things. Willingly deprioritizing your own work signals company-first thinking — which earns the budget for bigger swings later. "Shrink the change": frame bold bets as two-week experiments with pre-specified success criteria, show results, then ask for the full investment.

On AI: as execution collapses in cost and everyone can build, the scarce resource becomes deciding what survives and rallying people behind it. Influence is the 10x skill AI cannot replicate.

Why half of product managers are in trouble | Nikhyl Singhal (Meta, Google)

About half of product managers built their careers on moving information — synthesizing status, translating between management layers, brokering alignment. That role is becoming obsolete. Nikhyl Singhal (ex-Meta, Google, CPO at Credit Karma, founder of the Skip community for heads of product) argues the split is binary: builders versus information movers. Builders are thriving — compensation at all-time highs, PM openings at a three-year peak, the idea-to-shipped cycle collapsed. Information movers are in serious trouble.

The structural cause: companies that doubled headcount over five years are calculating they didn't get twice the output. The next 12–24 months bring mass layoffs followed by smaller AI-first rehires — shed 30,000, hire 8,000. Legacy brand names matter less; interviewers now ask what tools you use and what you've shipped recently.

What survives is judgment: deciding what not to build, evaluating whether a change is good, maintaining product coherence as the cost of generating changes approaches zero. At Singhal's CPO roundtable, heads of product are building internal agents to automate standups, reviews, and status reports — replacing the mechanical infrastructure of product management itself.

The psychological barrier is real. People who mastered the old game resist reinvention most — their system still works for them. Mid-career PMs face compounding personal obligations on top of the pressure to stay current. The fix: find one moment of genuine joy with new tools — a personal app, an agent replacing a hated task. That crossing generates energy and time rather than consuming them.

Builders will spread into every function as change agents. Designers bifurcate between pixel production (commoditized) and taste-making (scarce). Engineers converge toward PM-like judgment. The disruption is transitional — comparable to hardware-era product management collapsing when internet companies arrived. A stable model emerges within a couple of years; the window to position correctly is now.

A Three-Step Framework For Solving Problems 👌

Misdefining the problem is the surest way to fail a project. At Airbnb in 2012, a team built a polished social-travel feature only to discover the real need was "find high-quality non-touristy activities" — meetups were a solution, not the problem; Airbnb Experiences fixed the right thing years later.

Three steps prevent this. First, crystallize: write a single-sentence, solution-agnostic problem statement referencing an unmet user need ("Lyft drivers cancel because passengers are too far away," not "build a loyalty program"), backed by three to five strong evidence points, with a concrete success metric. Second, align before design begins — each teammate imagines a different problem, so circulate the draft and get stakeholder sign-off explicitly. Third, keep returning: open every design review with the statement and invoke it to cut scope creep.

This Week #13: Balancing outcome-thinking with design and technical requirements ⚖️

Airbnb organized teams around outcomes—"increase host supply," "improve trip quality"—not around product surfaces, freeing them to touch any part of the product to hit their goal. This correlated strongly with impact across seven years. For technical work that doesn't directly move an outcome—e.g., migrating from a monolith to SOA—extend outcome-thinking further out: map timelines, resource cost, and short- and long-term goal impact, then decide within the team or escalate based on ROI.

The PM 🤝 Design Partnership

PMs and designers share the same end goal but diverge in method, and that gap drives most collaboration friction. Katie Dill (Airbnb, Lyft, Stripe Head of Design) offers five fixes. Trust the designer's expertise: creative work is non-linear, needs uninterrupted flow time per Paul Graham's maker schedule, and some choices are pure taste — defer on UX. The PM is the conductor: clarify problems, goals, and constraints so each discipline can lead in its lane. Include designers in planning from the start — waterfall handoffs produce slower, thinner solutions. Invest in north stars even at small scope: Lyft's transit-modes vision took six months; a sign-up-flow sketch took a week; without one, products become disjointed Frankensteins. Stop framing "design goals" as separate from "business goals" — consistency and usability are sub-goals of conversion. Say "design partner," not "my designer."

Prioritizing

Most prioritization frameworks add complexity without reducing subjectivity — skip them. List every idea, T-shirt-size each on impact and cost (XS–XL), and sort by impact-to-cost ratio; expect 10–20% adjustment for dependencies. T-shirt sizes beat numeric ratings because they force manual sorting rather than trusting a formula. Set strategy before building any roadmap: mission → vision → strategy → goals → roadmap. Within each strategic pillar, allocate 80% to incremental work that moves KPIs while 20% of big bets mature behind that cover.

How to develop product sense

Product sense — consistently making changes that have intended impact on users — rests on empathy to discover real needs and creativity to solve them. Jules Walter (Slack growth, YouTube) treats it as a learnable skill built through four practices.

Observe users directly rather than reading research. At Slack in 2016, watching mobile newcomers revealed: homepage too abstract, competing calls to action, users rejecting address-book permission. Fixing these lifted retention. People drop out the moment they feel confused.

Deconstruct everyday products — an hour monthly, using Julie Zhuo's critique questions. Comparing Cash App vs. Venmo: Venmo leans on social graph, Cash App on breadth and ease.

Learn from great product thinkers. Stewart Butterfield's principle — "The details are not the details, they make the design" — explains why Slack users loved the product without knowing why. Key corollary: frame problems deeply before seeking solutions; strong constraints eliminate most options.

Track macro shifts (web3, AR/VR, remote work) and micro platform changes. Figma's browser app became viable once WebGL was fast enough. Cash App's Bitcoin bet drove 76% of 2020 revenue.

Progress shows up as anticipating non-obvious problems before reviews, sharper hypotheses under ambiguity, and being right more often about what a change does to metrics.

Gibson Biddle on his DHM product strategy framework, GEM roadmap prioritization framework, 5 Netflix strategy mini case studies, building a personal board of directors, and much more

Good product strategy means delighting customers in hard-to-copy, margin-enhancing ways — the DHM model Biddle absorbed from Reed Hastings at Netflix. Delight is 10x improvement, not incremental satisfaction. Hard to copy means advantages competitors can't quickly replicate: original content, personalization across a billion taste profiles, brand trust. Margin-enhancement means it must make business sense.

DHM runs best through A/B testing. Netflix's "Perfect New Release" test (2005) showed that next-day DVD delivery — what everyone said they wanted — reduced churn by only 0.05 points, saving ~5,000 customers worth $1M against $5M in inventory cost. The word-of-mouth multiplier (Netflix used 2x; Amazon used 10x) was the live variable that flipped the investment decision. Netflix's social features failed twice — a friends recommendation network and Xbox Watch Party (5% adoption, killed) — showing network effects don't transfer automatically across categories.

The 2020 auto-cancel decision shows the tension in reverse: canceling half a percent of inactive members cost $100M but built brand equity. Biddle frames such calls with Amazon's two-door test: is it reversible, and how large is it relative to total revenue?

For roadmap prioritization, Biddle uses GEM: force-rank Growth, Engagement, and Monetization. At Chegg his CEO wanted growth first, his CFO wanted monetization first — the conflict was destroying the team. A committed stack rank (Growth → Engagement → Monetization) gave product a single decision function.

On career: the path from IC to manager is primarily a hiring skill — one to two days a week recruiting. CPO-level work adds vision communication, culture as a management lever, and picking the right companies. Biddle's edge was a personal board of directors — peers and mentors consulted informally for calibration. Mentors are cultivated by being useful first, not by asking. Career hacking mirrors product work: form a hypothesis, experiment, read the results, iterate.