How Anthropic’s product team ships faster and thinks about AI PM work

Apr 23, 2026 · ai · Source ↗

Published 2026-04-23 - Runtime about 86 min - Watch on YouTube

TLDR

Cat Wu says Anthropic’s product speed comes from shrinking feature cycles from 6 months to 1 day and stripping barriers to shipping.
AI-native PM work is shifting from roadmap coordination to product taste, first-principles prioritization, and tight launch execution.

Key Takeaways

Anthropic ships many features in research preview, using it to lower commitment and get feedback in 1-2 weeks.
Weekly metrics readouts and team principles replace heavy PRD dependence for many decisions, though ambiguous or infrastructure-heavy work still gets PRDs.
Cat Wu says the Claude Code leak was human error during a package-release change, caught after two layers of human review.
Anthropic prioritized first-party products and the API over third-party Claude subscription usage, while offering credits during the transition.
Newer models like Opus 4.5, Opus 4.6, and Sonnet 4.6 unlocked reliable code review and reduced the need for prompting interventions.

Notes

Wu leads product for Claude Code and Cowork; Boris Cherny is the tech lead and product visionary setting the 3- to 6-month direction.
Her role focuses on the path from current product to future vision and on cross-functional alignment across marketing, sales, finance, and capacity.
Anthropic’s PM team is roughly 30 to 40 people, with separate teams for research, product, and other functions.
Anthropic’s product timelines have fallen from 6 months to 1 month, and in some cases to 1 week or 1 day.
AI-native PMs need to shorten idea-to-user time, not just coordinate multi-quarter roadmaps.
Clear goals matter because LLMs are general: Anthropic narrows scope by defining the key user, target problem, and intended use case.
A concrete example Wu gave was professional developers getting safely to zero permission prompts.
Most Claude Code features ship in research preview, explicitly labeled as early, so the team can iterate quickly without long-term commitment.
Engineering, docs, PMM, and launch coordination are tightly coupled; once a feature is dogfooded internally, launch work can turn around the next day.
Anthropic uses weekly metrics readouts and written team principles so people can make decisions without waiting on PMs or other stakeholders.
PRDs still exist for ambiguous features and projects requiring heavy infrastructure that take many months.
Wu says the model improvements explain only part of the speedup; process, expectations, and low-friction shipping culture matter more.
The source-code leak was investigated immediately, attributed to human error, and followed by hardened safeguards.