How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)

Summary of How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)

by Lenny Rachitsky

1h 25m•April 23, 2026
Listen to Episode

Overview of How Anthropic’s product team moves faster than anyone else | Cat (Kat) Wu (Head of Product, Claude Code)

This episode (hosted by Lenny Rachitsky) features Kat Wu, Head of Product for Claude Code and Co‑Work at Anthropic. The conversation explains how Anthropic ships at an unprecedented pace, how the PM role is changing in an AI‑first world, and practical ways teams can design, measure, and ship agentic / AI-native products. Kat shares concrete processes, tooling, team structure, hiring philosophy, and tactical advice for PMs, engineers, and designers adapting to rapidly improving models.

Key takeaways

  • Anthropic ships extremely quickly by removing barriers to release: many features move from planned 6+ months to 1 month, 1 week, or even 1 day.
  • Core levers for speed: strong product vision, low friction launch processes (Research Preview), tight cross‑functional launch workflows, and engineers with product taste who can ship end‑to‑end.
  • The PM role is shifting: less gatekeeping and roadmap coordination, more enabling rapid experiments, setting clear goals, and removing blockers.
  • Build for current model capabilities (the “right amount of AGI‑pilled”) — not only for imagined future super‑AGI.
  • Evals, small rigorous metrics, and trusted human evaluators are crucial to measure improvements and prioritize work.
  • Mission & focus (safety-first) helps Anthropic make fast cross‑org tradeoffs and stay unified.

Topics discussed

  • Kat’s role vs. Boris (tech/product split of responsibilities)
  • How Anthropic reduced release timelines and the internal cultural/process reasons
  • How PMs should change approach in AI era: goals, principles, fast experiments
  • Research Preview launch strategy and “evergreen launch room” workflow
  • PRDs: when they still matter (long/ambiguous projects)
  • Design and product taste as the most valuable emerging skill
  • Role blending: engineers doing PM work, PMs doing engineering, designers shipping code
  • Source‑code leak incident (human error, process fixes)
  • OpenClaw/cloud subscription policy: prioritizing first‑party product & capacity constraints
  • Tools: Cloud Code (CLI), Cloud Desktop, Cloud Web/Mobile, Co‑Work — when to use each
  • Co‑Work demos: slide deck generation, meeting dossiers, personalized sales decks
  • Tokens, internal usage, and cost tradeoffs
  • Evals and human evaluators (who’s good at model feedback)
  • How product changes with new model releases (remove harness cruft; unlock new features like reliable code review)
  • Advice for individuals: automate repetitive work to free creative time; get automations to near 100% reliability

How Anthropic actually moves fast — concrete processes

  • Research Preview: ship early and clearly label features as early/experimental to lower shipping commitment and gather feedback quickly.
  • Evergreen launch room: engineers post dogfooded features; docs/PMM/DevRel can turn around marketing and docs the next day.
  • Weekly metrics readouts: team‑wide metric reviews keep everyone aligned on goals and trends.
  • Team principles & user definitions: explicit principles that allow teammates to make decisions without constant PM sign‑off.
  • Empower engineers with product taste: many engineers can ship full features end‑to‑end, reducing handoffs.
  • Low process, high permission: remove approval friction; create clear rules for when to pull cross‑functional partners.
  • Regularly re‑read system prompts and harnesses after model upgrades: remove obsolete prompts/interventions.

Product breakdown — when to use what

  • Cloud Code (CLI): best for one‑off coding tasks, the most powerful surface and where features land first.
  • Desktop: better for front‑end work, previewing UIs side‑by‑side, and a graphical control plane to view sessions across devices.
  • Web/Mobile: great for kicking off tasks on the go (no laptop required).
  • Co‑Work: non‑code outputs — slide decks, meeting summaries, inbox/Slack triage, planning documents; integrates with Google Drive, Gmail, Calendar, Slack to synthesize context.

Use cases & examples (Co‑Work + Cloud Code)

  • Kat asked Co‑Work to create a 20‑page polished slide deck for a conference by connecting Drive/Slack and providing outline constraints — Co‑Work generated a strong draft that she iterated on.
  • Sales team built a web app that auto‑customizes decks per customer by pulling data from Salesforce/Gong/notes.
  • Applied AI team uses Co‑Work to prepare meeting dossiers summarizing customer history, action items, and the latest ETAs from Slack.

Team & org structure (PM org)

  • ~30–40 PMs across functionally scoped PM teams:
    • Research PM (model feedback & shepherding launches)
    • Cloud Developer Platform (APIs, managed agents)
    • Cloud Code & Co‑Work core product
    • Enterprise (cost controls, RBAC, security)
    • Growth (product growth across suites)
  • Hiring emphasis: engineers and designers with strong product taste; PMs often have engineering backgrounds or ship code on Cloud Code.

Evals, metrics, and the PM’s toolkit

  • Build a small set of high‑quality evals (10 great evals > 100 mediocre ones).
  • Weekly metrics readouts to align team on measurable goals.
  • Find a handful of trusted human evaluators (people who can give high‑signal feedback about model behavior).
  • Use model introspection: ask the model to explain its behavior to identify harness weaknesses.

Decisions, tradeoffs & risk management

  • Source code leak: human error in PR automation; result was two layers of human review failing — Anthropic hardened processes and added safeguards.
  • OpenClaw subscription limits: due to heavy compute costs and demand, Anthropic prioritized first‑party products and offered credits for transition; business sustainability influenced the decision.
  • Token/compute costs: model improvements increase token usage per user; internal teams have significant capacity but are expected to use tokens responsibly.

What to expect as models improve

  • Remove prior “crutches”: features added as workarounds for older models (e.g., manual to‑do lists) can be simplified or removed as models naturally handle tasks better.
  • New model generations unlock fundamentally new product capabilities (reliable automated code review, large multi‑task orchestrations).
  • Infrastructure shifts: as users run many tasks/agents concurrently, you’ll move from local to remote execution and need management surfaces for large numbers of tasks.

Practical advice / Action items for PMs, designers, engineers

  • Focus on product taste: decide what to build — as code becomes cheaper, choice becomes the key skill.
  • Shorten idea → user feedback loop: make it possible to ship concepts in days or a week via Research Preview.
  • Automate repetitive tasks for yourself: aim to reach near 100% reliability before calling something automated.
  • Invest time in model usage: spend hours interacting with models, ask them to introspect, and learn failure modes.
  • Build a small set of meaningful evals to quantify progress and anchor decisions.
  • Define clear user & success criteria early (helps rule out irrelevant approaches).
  • Be low ego and willing to wear multiple hats — roles will continue to blur.

Notable quotes & soundbites

  • “It is very hard to be the right amount of AGI‑pilled. The hard thing is figuring out, for the current model, how do you elicit the maximum capability?”
  • “We want to remove every single barrier to shipping things.”
  • “The PM role is changing a lot… the thing that is extremely important for building AI‑native products is iterating so quickly.”
  • “Product taste is still a very rare skill… deciding what to write is becoming more valuable.”
  • Kat’s life motto: “Just do things.”

Challenges and tradeoffs Anthropic accepts

  • Short‑term product overlap and inconsistency: experimenting rapidly creates overlapping features and may confuse users — Anthropic invests in onboarding/education (e.g., PowerUp).
  • Prioritization tradeoffs: mission‑first thinking (safety & Anthropic success) means teams will sacrifice some product or KR for org goals.
  • The pace of change: features may need rework when models jump — sometimes you remove a harness you previously built.

Lightning round — select personal highlights

  • Recommended reads: How Asia Works; The Technology Trap; Paper Menagerie (short stories).
  • Favorite shows/movies: Drive to Survive; Free Solo.
  • Favorite external product: Waymo — saves Kat ~30 minutes/day and allows productive calls en route.
  • Future non‑work plan: more rock climbing and reading (goal: 1–2 books/week).

Closing / Where to follow Kat

  • Twitter: @kat_wu (reads DMs; Kat asks for specific reproducible feedback about Cloud Code and Co‑Work failures — edge cases help them improve the product)

If you want a single‑page checklist of actions for PMs from this episode (e.g., quick eval checklist, launch room template, or Research Preview playbook), tell me which one and I’ll condense it into a ready‑to‑use template.