Claude writes about 90% of Claude Code's code

Claude writes about 90% of Claude Code’s code, so review had to scale from a human gate to a continuous, machine-first process to handle a surge in changes. The goal is to minimize human time without increasing reverts or incidents; metrics tracked include median and 95th-percentile human approvals per PR, human comments, wait time, and revert/incident rates as guardrails. Between June and September merged pull requests more than doubled while reverts fell over 40%. Most routine mistakes are caught and often fixed by Claude before a person sees a change, preserving human attention for novel or high-risk decisions.

The review flow is layered: pre-PR verification runs tests, builds and screenshots via Claude skills; every PR gets an AI review where multiple agents hunt for bugs and verifiers vote on findings, which are posted as categorized inline comments and must be answered. Automated security reviews and blocking CI (lint, types, tests, builds) run on every PR. A risk-classifier assigns approval tiers (none, one, or two humans) and trusted authors can take lighter paths. Merge queue rechecks against main, deployments roll out behind flags, and fast rollback and on-call authorship handle failures. Conventions live in CLAUDE.md and REVIEW.md, and recurring failures are fixed at the source via new skills, lints, or docs.