ClaudeCode reviewLevel 2: Context

How Do You Give Claude a Code Review Checklist?

Write down the checks your team actually cares about, and put them where the review will read them. For one review, paste them into the prompt. For every local /code-review, put them in CLAUDE.md. For Anthropic's Code Review service on GitHub, use a REVIEW.md at the repo root.

Prompt to copy
Review this branch against the checklist below. For each item, report either the findings or "no issues", so I can see that it was covered.

Always check:
- [New API routes have an integration test]
- [Log lines don't include email addresses, user IDs or request bodies]
- [Database queries are scoped to the caller's tenant]
- [a rule of yours]

Don't report:
- Anything CI already enforces: [lint, formatting, type errors]
- Generated files under [src/gen/] and lockfiles

For each finding give the file and line, why it matters, and the smallest fix. Behavior claims need a file:line citation, not an inference from a name. Skip style.
  1. 1In the promptthis one review
  2. 2CLAUDE.md/code-review and the service
  3. 3REVIEW.mdthe Code Review service only
Where a checklist goes, and which review reads it, according to Anthropic's docs.

Try it now

  1. 1Add two or three always-check rules to CLAUDE.md, then run /code-review on your branch. They apply to every local review.
  2. 2On a Team or Enterprise plan with the service on, commit REVIEW.md at the repo root and merge it. It takes effect only after it merges.
  3. 3Protect your default branch. Lowering severity, excluding paths and capping nits are honored when the review reads the file from a protected default branch.
  4. 4Comment @claude review on a test pull request with a deliberate violation, and watch for the **Claude Code Review** check run. It worked if your rule comes back as a finding.

Use it when

A team with rules the code can't show, such as tenant scoping or keeping personal data out of logs. Also a payments or auth change that deserves a heavier pass than a copy edit.

Skip it when

Anything your CI already enforces. Anthropic's sample REVIEW.md tells the reviewer not to report lint, formatting and type errors. A checklist also can't tell you whether a change should exist at all. The service works from reading the diff and surrounding code, without building or running it, so design calls stay with you.

Why it works

An open-ended request gets whatever stands out. The service's default focus is correctness: bugs that would break production, not formatting or missing test coverage. So a team rule such as new API routes have an integration test won't be checked unless you write it down. Each tool reads different files. /code-review follows your CLAUDE.md, including in -p runs, and doesn't read REVIEW.md. The service reads both: its finding and verifying agents receive REVIEW.md as review instructions, and a change that violates CLAUDE.md becomes a finding. It reads REVIEW.md from the base or default branch, never from the pull request's own changes, so a pull request can't rewrite its own review rules.

Example

The sample REVIEW.md in Anthropic's docs has four parts for a backend service. It says what Important means here, with style and naming as Nit at most. It caps volume with Report at most five Nits per review. It lists what not to report, including src/gen/ and any *.lock file. And it lists what to always check, such as database queries being scoped to the caller's tenant. Say a pull request adds a route with no integration test. With that file merged, the rule gives the review something specific to flag.

Going further

REVIEW.md has limits. It can include other files with an @ line, up to eight files and 150,000 bytes combined, and a file that would cross the byte limit is skipped whole. Includes work only when the file is read from a protected branch. On an unprotected default branch, the @ line stays as plain text and subdirectory REVIEW.md files aren't read. For a path that needs some review but not full scrutiny, set a higher bar: in scripts/, only report if near-certain and severe. To cut false positives, the docs suggest requiring evidence: behavior claims need a file:line citation in the source.

Common mistakes

  • Expecting /code-review to follow REVIEW.md. Put anything you want in both places in CLAUDE.md.
  • Editing REVIEW.md in a pull request and expecting that same review to use it. Merge the change first.
  • Writing be thorough into the prompt. On /code-review, the real control is the effort level, as in /code-review high.
  • Leaving a repo on After every push, where repos added from the admin page start. The service bills by token through usage credits, with no flat price, so each push adds cost. Pilot with Once after PR creation or **Manual (@claude review)**.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course