How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
TL;DR
Anthropic compressed product timelines from 6 months to as little as 1 day — Cat Wu says the biggest PM shift in AI is not roadmap coordination but removing every barrier between an idea and users, with Claude Code shipping most things in research preview to learn fast.
The new moat for PMs is product taste, not document-writing — as code gets cheaper, the scarce skill becomes deciding what to build, which GitHub issues matter, and how to design the “golden path” for today’s models rather than some abstract AGI future.
Anthropic’s speed comes more from process and culture than from secret internal models — Wu credits low process, weekly metrics readouts, tight launch handoffs across engineering/docs/PMM, and engineers with strong product sense more than any single model like the previewed “Mythos.”
Claude’s personality is treated like product infrastructure, not cosmetic polish — Wu says users love Claude because it feels competent, low-ego, positive, and collaborative, and Anthropic has people like Amanda explicitly shaping that character because it affects how well the product works.
Anthropic willingly trades product consistency for velocity — shipping overlapping features, rough edges, and constant updates creates confusion for users, but Wu says that’s the cost of testing multiple form factors while models improve fast enough to obsolete product scaffolding every few months.
Her practical advice for the AI era is brutally simple: automate the repetitive 100%, then use the freed time on real leverage — Wu argues a 95% automation is not really an automation, and urges people to stop one-shotting toy apps and instead build workflows they actually use every day.
The Breakdown
Cat Wu’s actual job: translating Boris’s vision into shippable reality
Wu describes Boris Cherny as the product visionary for Claude Code — the person seeing the “AGI-pilled” version of the product 3 to 6 months out. Her role is turning that future into something users can use now, while handling the messy cross-functional work across marketing, sales, finance, and capacity so features don’t get stuck right before launch.
Why most PMs are playing the AI game wrong
She says many PMs still operate like it’s the pre-AI era: long planning cycles, multi-quarter alignment, careful dependency management. At Anthropic, feature timelines have collapsed from 6–12 months to one month, one week, or even one day, so the real job is shortening the path from idea to user feedback and making weekly shipping feel normal.
The mechanics behind Anthropic’s shipping speed
Wu’s answer is surprisingly operational: set extremely clear goals, ship in “research preview” to lower commitment, and build a tight launch process so docs, PMM, and DevRel can jump on a feature immediately. She says engineers post ready features in an “evergreen launch room,” and people like Sarah, Alex, Tar, and Lydia can help turn them into public launches by the next day.
PRDs didn’t disappear — they just got demoted
Anthropic still writes PRDs, but mostly when a project is especially ambiguous or infrastructure-heavy. Day to day, Wu relies more on weekly metrics readouts and a written set of team principles — who the core users are, what tradeoffs matter — so people can make decisions without waiting on PM approval.
Leaks, Open Claude, and the reality of operating at scale
On the Claude Code source leak, Wu says Anthropic traced it to human error during a package release process, even after two layers of human review, and kept the person involved because it was a process failure, not a blame story. On restricting third-party products like Open Claude from riding first-party subscriptions, she’s blunt: Anthropic had to prioritize its own products and API as demand surged, even though it tried to soften the transition with credits.
The PM role is blurring into engineering — but taste still wins
Wu says Anthropic has around 30 to 40 PMs across research, developer platform, Claude Code, enterprise, and growth, but the bigger story is role convergence. Many PMs used to be engineers, designers can ship frontend code, and Anthropic prefers engineers with strong product taste who can go from seeing user feedback on Twitter to shipping by the end of the week with almost no PM help.
What humans still do better than models
Her answer isn’t “strategy” in the abstract — it’s common sense, prioritization, and social navigation. Models still struggle with stakeholder nuance, communication venue, EQ, and the thousand tiny moving pieces in a launch, so humans remain most useful at choosing what matters, spotting what’s broken, and making judgment calls under chaos.
Shipping fast means accepting mess, and Anthropic openly does
Wu says the cost of this speed is product inconsistency: overlapping features, confusing paths for new users, and users feeling like they have to check Twitter every day to keep up. That’s partly why Anthropic added onboarding like /powerup even though the team originally wanted the product to be intuitive enough to need no tutorial at all.
Mission, focus, and why Anthropic seems unusually aligned
She argues Anthropic’s edge comes from two things: a unifying mission around bringing safe AGI to humanity, and a willingness to put company goals above any one team’s goals. Her line is memorable: if Claude Code failed but Anthropic succeeded, she’d be “extremely happy,” and that kind of attitude makes hard prioritization much easier.
How she actually uses Claude Code, Desktop, and Cowork
Wu’s split is clean: Claude Code in the terminal is the most powerful and gets the newest features first; Desktop is best for visual work and as a control plane for sessions; mobile/web are for kicking off tasks while literally “touching grass.” Cowork is for non-code output — Slack zero, inbox zero, docs, and decks — and she gives a vivid example of asking it to pull from Google Drive, Slack, Twitter, and internal launch rooms to draft a 20-page conference deck overnight.
The hardest PM skill now: being the right amount of “AGI-pilled”
Her sharpest point is that it’s easy to design for a superhuman future where the model just needs a textbox. The hard part is designing for current models: understanding their strengths and weaknesses, asking them to introspect on strange behavior, finding the five users whose feedback you really trust, and writing enough evals — maybe just 10 good ones — to define what success actually looks like.