Claude writes about 90% of Claude Code's code

Claude's team at Anthropic have optimized the code review process to handle the increased volume of code being produced. They aim to minimize human review time without compromising code quality. Their system reviews about 90% of the code automatically. The review process evolved over a few months, with a significant increase in merged pull requests and a decrease in the need for code reverts.

Rather than reducing overall review time, they moved most of the reading to machines and focused on making human review faster and easier to correct. Claude Code, a separate repository from the monorepo, receives heavy scrutiny given its rapid development pace. Instead of a traditional gate-like review system, review now happens continuously, similar to spellcheck, during the writing process.

Most changes have been reviewed and fixed by Claude multiple times before a human looks at them, catching routine mistakes and saving human attention for more complex issues. The goal is to keep human time and attention for issues only humans can solve, while letting automated checks handle the rest. Human time is measured by tracking approvals, comments, and PR wait times.

A typical change goes through multiple layers of review before a PR is created. Claude verifies its work, checking for bugs, simplifying the diff, and testing the change end-to-end. If an automated review flags an issue repeatedly, it gets addressed by fixing the skill, lint rule, or instructional files. Each PR undergoes a Claude review first, which examines the diff against the entire repo and the PR's intent.

The review comments get verified by other agents before being seen by humans. After verification, the findings get posted as inline comments, sorted by severity. Important issues must be fixed, while Nit and Pre-existing items simply get resolved with a reason. Security reviews also occur before a PR can merge, with any findings needing to be fixed or resolved before the PR can be completed.