Cat Wu

AI Summary

→ WHAT IT COVERS Cat Wu, Head of Product for Claude Code at Anthropic, explains how her team ships features in days rather than months, why product taste has become the scarcest PM skill, how Claude Code and Cowork divide responsibilities, and what the PM role looks like when model capabilities change faster than any roadmap can accommodate. → KEY INSIGHTS - **Shipping velocity framework:** Anthropic reduced feature timelines from six months to one week or one day by creating a standing "evergreen launch room" where engineers post completed features, triggering same-day turnaround from docs, PMM, and DevRel. Labeling releases as "Research Preview" removes the commitment barrier, allowing the team to ship rough versions within days and iterate based on real user feedback rather than internal speculation. - **PM goal-setting specificity:** Vague goals create paralysis on AI-native teams. Effective PMs define the exact user segment, the precise problem, and the specific use case — for example, "professional developers at enterprises need zero permission prompts safely" — which automatically eliminates most solution candidates and lets engineers make independent decisions without waiting for PM sign-off on every micro-choice. - **Product taste over technical skills:** As code generation costs drop, the scarce resource becomes judgment about what to build. Cat Wu's team prioritizes hiring engineers with product taste over hiring more PMs, because engineers who can read user feedback on Twitter and ship a fix by end of week require almost no coordination overhead. Engineering background helps for roughly the next few months because it informs effort estimation during prioritization. - **Model-harness relationship:** Every new Claude model release triggers a full system prompt audit to remove prompting interventions that compensated for prior model weaknesses. The to-do list feature, originally added to force Claude to complete all 20 call sites in a refactor, became unnecessary with Opus 4. Teams should build features that don't fully work yet, then swap in newer models to test whether capability gaps have closed. - **Cowork as non-code output layer:** The practical split between Claude Code and Cowork is output type — code versus everything else. Cat Wu used Cowork to generate a 20-page conference slide deck overnight by connecting Slack, Google Drive, Gmail, and Google Calendar, feeding it a PMM draft and a narrative direction, then reviewing the output in the morning. The deck matched Anthropic's design system because she supplied the existing slide template as context. - **Automation completion standard:** A 95% reliable automation delivers almost no real leverage because it still requires human monitoring for the failing 5%. The correct target is 100% reliability, which requires iterating on Claude's preferences through explicit feedback loops — defining a skill, running it, correcting errors, and instructing the model to update the skill definition. Stopping at "good enough" means the automation cannot run unattended and the time investment yields minimal return. → NOTABLE MOMENT Cat Wu describes how Anthropic's source code leak happened despite passing two layers of human review — a developer used Claude to write a package release PR, and human error at both review stages allowed it through. Anthropic treated it as a process failure rather than an individual failure and hardened the release pipeline afterward. 💼 SPONSORS [{"name": "WorkOS", "url": "https://workos.com"}, {"name": "Vanta", "url": "https://vanta.com/lenny"}] 🏷️ AI Product Management, Claude Code, Anthropic, Shipping Velocity, Product Taste, Agentic Tools