Claude Code Team Playbook for High-Volume Automated Pull Request Reviews
The Claude Code team detailed their internal review flow handling over 30,000 daily changes with 90 percent machine-written code. By shifting verification upstream and requiring consensus across agent reviewers, PR reverts dropped over 40 percent while merge volume doubled.
Why it matters
You can structure your repository using CLAUDE.md and REVIEW.md to automate multi-pass verification before PR creation.
TL;DR
- 01Give coding agents direct access to execute tests, builds, and design diffs before they finalize a code change.
- 02Add short REVIEW.md files to code subtrees defining concrete failure modes and capping nitpick comments.
- 03Use multi-agent voting to filter hallucinated bugs and purely cosmetic feedback from PR comments.
Key facts
- Monorepo Daily Changes
- 2,500 in May to >30,000 in September
- Internal Code Written by Claude
- ~90% (self-reported)
- Substantive Review Comments Share
- 16% to 54% after preview launch
- PR Revert Rate Reduction
- >40% reduction while PR volume doubled
Shifting Review Upstream
Anthropic reports that Claude writes roughly 90% of Claude Code's codebase. As changes in Anthropic's monorepo surged from 2,500 daily commits in May to over 30,000 in September, manual review became a major bottleneck. The team solved this by moving the burden of reading diffs to machines. Between June and September, merged pull requests doubled while revert rates fell by more than 40%.
Boris Cherny emphasizes that giving Claude executable verification—such as running a test suite, executing a bash build command, or running an app inside a simulator—improves output quality by two to three times. Authors run internal skills to spot bugs and simplify diffs locally before opening a pull request.
The Role of CLAUDE.md and REVIEW.md
Automated reviews enforce constraints through structured markdown documentation:
CLAUDE.md: Outlines the project stack, architectural goals, non-goals, and core coding conventions. Reviews fail if a proposed diff breaks documented conventions or leaves references outdated.REVIEW.md: Maintained by module owners to specify failure definitions (e.g., routing errors or documentation drift), ignore rules, and enforce that reviews cite concrete file paths and line numbers rather than guessing.
Multi-Agent Consensus and Gated Merges
On pull requests, multiple agent instances evaluate diffs concurrently and cast votes on each bug candidate. Findings only post as inline comments if multiple agents reach consensus, filtering out noise, diagnostics, and style squabbles. The public Code Review feature supports Quick, Standard, and Deep inspection levels, while developers can trigger lightweight checks locally via /code-review and /security-review.
Try it in 2 minutes
# REVIEW.md
## Scope & Rules
- Failures: breaking API changes, state desynchronization, unhandled edge cases.
- Skip: autogenerated client bindings under /src/generated/.
- Formatting: Ignore stylistic nits; require exact filename and line number for defects.
- Verification: Require bash test output confirmation before approving.markdown
✓ When to use
- High-velocity teams integrating agentic tools like Cursor and Claude Code across shared monorepos.
- Codebases where human code review has become a severe delivery bottleneck.
✕ When NOT to use
- Small personal repositories where automated multi-agent passes add latency without catching real regressions.
- Prototyping codebases without continuous integration test suites or reproducible verification scripts.
What to do today
- Create a scoped REVIEW.md in your repository root defining fatal error conditions and ignored paths.
- Run /code-review or /security-review locally in Claude Code before opening large diffs.
- Add executable verification steps to your CLAUDE.md instructions so agents test their own output.
What the community says
“No one cares about good code. It’s a means to an end... The only thing people care about is performant and error free code.”
“If agents made me 2x faster (they have, review is still slow; I currently have 40 PRs open) that’d be the same rate.”
“Poorly I'm afraid, given how many new bugs make their way to the claude code CLI every update.”