How Claude Code Reviews Its Own Code | made with ngram

About the video

This video was created by ngram to explain how Anthropic manages code review at scale now that Claude writes about 90% of Claude Code's own code, with daily changes across the monorepo jumping from 2,500 to over 30,000. Using HTML/CSS motion graphics, the video walks through seven scenes that lay out the core idea: review isn't one final gate but a stack of layered checks, most of them automated. It covers how Claude runs tests and a fresh agent reviews each diff before a PR even exists, how CI runs agents that hunt bugs in parallel with verifiers voting on findings, and how a security pass can block a merge. It explains that humans only decide after these automated layers, with a classifier scaling the need for zero, one, or two human reviewers based on risk, and notes that a bot's approval never substitutes for a person's. The video also covers what happens after merge, including a merge queue, feature flags, staged releases, and fast rollback, and closes with the measured outcome: merged PRs more than doubled while the share of reverted changes fell by more than 40%, meaning less human time per change with no increase in reverts. The result is a concise, data-driven explainer of an engineering review process at scale.