Cat Wu Interviewed Hundreds of PM Candidates — Almost None Got the AI PM Question RightCat Wu 面试了几百个 PM 候选人，几乎没人答对一个问题：AI 产品经理到底应该干什么？

Cat Wu argues most PM candidates still think in 6-12 month roadmap cycles, while Anthropic ships features in a day — the job has shifted from cross-team alignment to compressing idea-to-user latency.

The mechanism behind the speed is threefold: sharply defined targets, a 'research preview' release framing that lowers commitment, and a cross-functional rapid-response launch pipeline (the 'evergreen launch room').

The interview opens with Cat's working relationship with Boris Cherny, framed as '80% mind-meld, 20% each running their own lane.' Boris sets the multi-month directional vision; Cat translates it into execution paths and handles cross-functional buy-in across sales, marketing, and capacity teams. The deliberately fuzzy division of labor, she argues, is precisely what enables speed — formal RACI-style ownership would slow things down.

The core diagnosis: most PM candidates fail interviews because they're optimizing for a world where code was expensive and roadmaps needed 6-12 months of cross-team coordination. AI has inverted the cost structure — shipping now takes a week or a day — so the real skill is compressing idea-to-user latency, not roadmap alignment. This reframes the PM's job from planner to accelerant.

Three concrete mechanisms explain the one-day cadence: (1) sharply pinned target users and failure modes that eliminate irrelevant solution space; (2) 'research preview' as a release framing that lowers the bar for shipping half-finished ideas; (3) the 'evergreen launch room,' a standing cross-functional channel where docs, marketing, and dev-relations react within a day once an engineer flags a feature as ready. PRDs aren't dead but shrink to weekly metrics readouts and a living team-principles doc that lets individuals decide without waiting on a PM.

Cat's most substantive framework is 'the right amount of AGI-pilled': designing for the ultimate AGI endpoint (a single text box, infinite capability) is trivially easy but useless now, since it ignores where the current model actually breaks. The hard skill is reading signals of where users hit today's capability ceiling and building precisely at that edge — a discipline she practices by spending 30% of her time deliberately stress-testing Cowork.

On role fusion, Cat's answer contains a latent contradiction: she describes PM/engineer/designer boundaries dissolving, with Anthropic choosing to hire engineers with product taste over scaling dedicated PM headcount. Yet the most efficient teams she describes need 'almost no PM involvement' end to end. Her fallback justification — product taste as the scarce, irreplaceable skill — stays undefined throughout, making it functionally unfalsifiable as a hiring or value criterion.

The editorial annotations add the piece's sharpest value: two leaks within a week get chalked up to isolated human error with no reflection on whether ship-fast culture raises systemic risk; the OpenClaw subscription ban has sound unit economics (a $200/month plan can burn $1,000-5,000 in API cost via always-on agents) but its timing right after Cowork shipped overlapping features undercuts Anthropic's stated open-source enthusiasm; and Cat's own AGI-pilled framework implies her product function is ultimately transitional.