RC RANDOM CHAOS

Claude Writes 90% of Its Own Code — Here's How They Review It

· via Hacker News

Original source

How the Claude Code team does code review

Hacker News →

Anthropic’s Claude Code team faces a volume problem: the monorepo went from roughly 2,500 daily changes in May to over 30,000 by September, and merged PRs on the Claude Code repo more than doubled in that same window while reverts dropped by more than 40%. Their answer is to shift most code reading to machines and reserve human attention for judgment calls. Review now runs continuously during development rather than as a final gate. Engineers give Claude executable verification targets—test suites, builds, design screenshots—and a subagent re-reads the diff with fresh context before a PR even exists. Internal skills then hunt bugs, simplify diffs, and check end-to-end behavior. The team’s creator, Boris Cherny, estimates that giving Claude a way to verify its own work makes final results two to three times better.

Before a human sees anything, every PR passes through an automated Claude review that reads the diff against the full repo and the stated intent. Findings from multiple agents are cross-validated by a voting group, and only substantive bugs get posted—style and cleanup notes are filtered out. Each finding must be answered: fixed or resolved with a reason. Security review runs in parallel, and a PR cannot merge while any security finding remains open. A classifier assigns a risk tier to each change based on its content rather than its author, and only trusted authors on the repo’s list get the lighter review paths. The team measures success by human approvals per PR, wait times, and—critically—revert and incident rates, treating the latter as the guardrail that proves human time was actually saved rather than just relocated downstream.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.