An internal Claude tool now lands 65% of the Claude Code team’s product PRs

At an AI Engineer World’s Fair session transcribed by Simon Willison, Anthropic engineers Kat Wu and Tariq Shihipar walked through how the team ships. Wu said an internal version of Claude Tag now merges 65% of their team’s product PRs, more than half of everything that goes in.

New Claude Code features roll out to Anthropic staff first. Only the ones that hold up on retention with that audience make it to production, and each feature has an internal bar for active users and retention.

Core changes still get a manual review from a code owner. For the outer layers, the team moved review to automation over six months: after every incident, the PR that caused the failure gets added to the code-review eval set. They also cut the Claude Code system prompt for frontier models by 80%.