Cat Wu, Head of Product for Claude Code at Anthropic, says feature timelines have collapsed from six months to one month, sometimes to one day.
Cat Wu says Anthropic ships almost everything as a research preview and holds weekly metrics readouts with the entire product team.
Cat Wu argues that product taste, deciding what to build, is the scarce skill as code gets cheaper, and that PM, engineering and design are merging.
Anthropic ships in days by cutting process and trusting taste
Anthropic's product team moves fast by cutting process and shortening the path from idea to users. Features that once took six months now take a month or a day, so long planning gives way to shipping and learning from feedback. With code cheap, taste about what to build is the scarce skill, and a shared mission keeps teams aligned.
Takeaways
Feature timelines shrank from months to days. Cat Wu says features that took six months now take one month, sometimes one day. PMs should skip multi-quarter roadmaps with partner teams and find the fastest route to users. She credits low process more than the Mythos model.
Product taste becomes the scarce skill. PM, engineering and design roles are merging, and Anthropic hires engineers with strong product taste. As code gets cheaper, deciding what to build matters most. Humans still add common sense and stakeholder judgement that models lack.
Mission outranks any single product line. Anthropic's goal of bringing safe AGI to humanity lets teams deprioritise their own goals. Cat says she would be happy if Claude Code failed but Anthropic succeeded. Overlapping features ship so users can choose between form factors.
Each new model reshapes the product. New models let the team remove scaffolding such as the to-do list and unlock features such as code review running several agents before merge. Users go from one task to six at once. Cat expects dozens or hundreds of remote instances.
Cowork brings agents to non-code work. Connected to Calendar, Slack, Gmail and Drive, Cowork turned internal sources into a 20-page deck in about an hour. A sales colleague built an app tailoring customer decks from Salesforce and Gong data. Applied AI is the second heaviest token user.
Anthropic addressed the leak and OpenClaw. Cat calls the Claude Code source leak human error that passed two layers of review, and says safeguards were added. On OpenClaw, Anthropic prioritised first-party products and the API, since third-party usage differs. Subscribers got some credits.
Cat Wu's product playbook
Set clear goals for a key user.
Ship almost everything as a research preview.
Hold weekly metrics readouts with the whole product team.
Keep a list of team principles.
Write PRDs only for ambiguous features or long infrastructure projects.
Building skill with AI
Ask models to introspect on their mistakes.
Trust a small group of about five feedback givers.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Write around 10 good evals, which matter most for features like memory.
Automate repetitive tasks until they work every time, and avoid over-customising tools.