How the Claude Code team does code review — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Anthropic's Claude Code team more than doubled merged pull requests from June to September while cutting the share that had to be reverted by more than 40%, by moving most code reading to machines and investing in fast, easy rollbacks.
- Claude writes about 90% of Claude Code's own code, and across Anthropic's monorepo, changes landing per day went from roughly 2,500 in May to more than 30,000 in September.
- The Claude Code team runs automated code review on every pull request before any human looks, and a separate group of agents votes on each finding — only those with enough agreement are posted, while style-only and pre-existing issues are dropped.
- Code Review for Claude Code, the public version of the team's automated reviewer, is now generally available on Team and Enterprise plans, and a lighter version runs locally via the /code-review command.
- The research preview of the automated reviewer raised the share of internal PRs receiving substantive review comments from 16% to 54%, with most large PRs (over 1,000 lines) getting findings and most small ones returning clean.
- A risk-tier classifier reads each change and decides how much human review is required — sensitive paths always need a person, and an agent's approval never counts as a human reviewer.
- Boris Cherny, who created Claude Code, estimates that giving the model a way to verify its own work (running tests, builds, or comparing screenshots) makes the final result two to three times better.
Why it matters: Anthropic is publishing a concrete playbook for reviewing AI-written code at scale: verify agent output with executable checks before a human looks, use a classifier to set risk tiers per change, and reserve humans for cases only humans can solve. With reverts falling 40% even as merged PRs double, the team is arguing the bottleneck has shifted from reading code to directing where human attention goes — a model other engineering organizations will be pressure-testing.
Ask SkimNews
