At the AI Engineer World’s Fair, Simon Willison moderated a fireside chat with Cat Wu and Thariq Shihipar from Anthropic’s Claude Code team. The full video is now on YouTube — and it’s worth it, because the team talks openly about how it builds Claude Code with Claude Code. A few points stuck with me.
’@Claude’ writes the majority of PRs
Maybe the most surprising number: Claude Tag — the @Claude in Slack — now lands 65 percent of the product engineering PRs for the Claude Code team. In other words, the team delegates a large share of its own feature work to the very tool it’s building.
Its own people first, then the world
New features go to Anthropic employees first. Only what shows real user retention in that cohort gets rolled out — what actually helps, not just what sounds good. Critical changes are still reviewed manually by humans; for the “outer layers” of the product, the team increasingly relies on automated code review.
Prompting: more context, fewer prohibitions
The most practical part for anyone working with Claude directly. The team prefers to give more context and fewer “don’t do this” instructions. The reason: negation is a very strong impulse for Claude — too many hard prohibitions, and too many instructions overall, tend to cause confusion. And rather than prescribing exact rules, it’s often better to let Claude use its own judgement — for instance, letting Fable decide when tests make sense instead of spelling it out.
That prompting works isn’t a gut feeling, either: the team calls it “definitely a science” and built out automated evals for it.
My take
Insights like these are rare, and for Claude Code users they’re pure gold. The prompting lesson matches my own experience: long lists of prohibitions make Claude more hesitant, not more precise. Provide context and leave the model room to judge, and you get better results. And the 65 percent figure is more than a fun fact — it’s evidence that in a top team’s daily work, agentic coding is no longer an experiment but a standard tool.
Sources: