Claude Codes New REVIEW.md System, What is it?

Summary

The video explains Anthropic’s approach to reviewing code as AI agents increase the volume of changes. The presenter says merged pull requests more than doubled from June to September, while reverts fell by over 40%. He argues that the main challenge is no longer generating code, but checking it quickly without increasing customer-facing failures.

Anthropic’s guiding question is not simply, “Was this reviewed by a person?” but, “What could go wrong, and which automated or human check is best suited to catch it?” The goal is to move routine review work to machines while keeping human oversight proportional to risk. Rollbacks and production safeguards are essential: faster merges are not a success if they merely shift defects downstream.

Anthropic’s Review and Safety Process

The process is described as six layers of checks, with rollback and staged-release practices providing additional protection:

1. Pre-PR verification: Agents run available tests, builds, browser or end-to-end checks, and may produce screenshots to demonstrate results. A separate agent can inspect the changes in a fresh context. Project guidance files such as CLAUDE.md or AGENTS.md, along with a REVIEW.md file, can specify conventions and review priorities.
2. AI review of each PR: Multiple agents examine the diff and search for defects in parallel. Other agents assess the findings, and comments are posted only when there is sufficient agreement, helping reduce false positives and low-value nitpicks.
3. Security review: Security findings must be resolved before a change can proceed in Anthropic’s internal process.
4. Standard automated checks: Linting, type checks, builds, and platform-specific tests remain merge blockers. PR templates also ask authors to state how they tested a change and how it can be rolled back.
5. Risk-based human approval: Approval requirements depend on the change’s risk, not the author. Low-risk changes may need no human approval; ordinary product changes generally need one person; and high-risk changes may need two. Agent approval never substitutes for required human approval. The person responsible for an agent-generated PR remains accountable, and a new commit requires approval again.
6. Merge queue: Checks are rerun against the current main branch so that concurrent PRs do not combine in a way that breaks it.

The presenter also describes staged releases using feature flags: changes are tested internally, then with a broader group, before being promoted to a stable release. Monitoring and quick rollback limit the impact of problems. Incident reviews are used to ask where in the process an issue should have been caught.

Lessons from an Excessive Review Loop

One Claude Code PR reportedly went through more than 50 rounds of AI review and fixes overnight. The presenter says around 59% of repeatedly flagged issues were in pre-existing code, not newly introduced changes. Because agents may keep trying to fix minor or unrelated findings, each fix can trigger another commit and review cycle.

Anthropic’s response was to limit the review loop: after three rounds, the reviewer should ask the author to reconsider the overall approach. Reviewers should also identify underlying classes of problems rather than repeatedly flagging individual lines. The team reportedly uses real past incidents as evaluation data to measure whether its reviewer is improving.

Claude Code Review Product

The video presents Claude Code Review as a productized version of the AI-review layer. It uses parallel agents to find issues, filters and ranks findings, and adds review comments. The presenter cites Anthropic’s claim that reviews can take roughly 6–10 minutes rather than about 20, with Quick, Standard, and Deep modes offering different levels of analysis and cost.

A key distinction: the product does not approve or block a PR. It supplies review findings, while merge decisions and required human approvals remain separate. The presenter suggests teams can use it as a model for building or adapting their own review agents and guidance files.

Other Technology and Practical Recommendations

Anthropic reportedly replaced Git in its large monorepository with Scion, alongside a viewing tool called Antique. The video describes this as a response to the scale of Anthropic’s repository and CI workload, rather than a change most teams need to make.

For teams adopting agent-assisted development, the presenter recommends:

• Give agents concrete checks they can run.
• Add a REVIEW.md describing what reviewers should focus on.
• Pilot an AI reviewer on one repository and tune it over time.
• Set approval levels according to change risk and potential impact.
• Keep human accountability and monitor reverts, incidents, and errors.
• Limit repeated automated review cycles to avoid agent-driven fix loops.

A sponsored segment briefly covers OpenRouter Armor, described as a layer that checks agent tool calls against policies before execution and can block disallowed actions.

Reviews, Guides, and Tutorials Mentioned

• Anthropic’s code-review process: The main source for the six-layer workflow, risk-based approvals, review loops, and reported outcomes.
• Claude Code Review documentation: Setup and usage, including REVIEW.md, @claude review, and the local /code-review command.
• Security-review guidance: The video mentions /security-review and a GitHub Action as ways to add security checks.
• Practical walkthrough: The presenter’s recommendations for adding review guidance, piloting an AI reviewer, and configuring risk-based approvals.

Main Speakers and Sources

• Rob Shocks (Switch Dimension): Narrator and analyst; presents the workflow and practical takeaways.
• Anthropic’s Claude Code team and published materials: The source of the review-process details and reported results discussed in the video.
• OpenRouter Armor: Sponsor featured in a separate segment.