Claude Code review skill

A Claude code review skill has one job: check your changes for bugs before a human reviewer sees them. Claude Code already ships one as a bundled skill, /code-review. There is nothing to download. This page covers what it checks, the options that change its behavior, and the related review tools Anthropic offers.

What the skill does

By default /code-review reviews your branch's commits ahead of its upstream plus any uncommitted changes. It reports correctness bugs. Depending on your model and effort level, it also suggests reuse, simplification and efficiency cleanups. The review runs as a background subagent with its own context window, so it does not fill your conversation, and the findings arrive when it finishes.

You can point it somewhere else by adding a target: a file path, a PR number, a branch name or a range such as main...my-feature. Its full form, from the Claude Code command reference:

/code-review [low|medium|high|xhigh|max|ultra] [--fix] [--comment] [--max-findings n|all|default] [pr#|branch|path]

• Effort. At low and medium it reports only the findings it is most confident in. high through max cover more ground and may include findings it is less sure about. If you type no level, it reuses the last one you typed.
• --fix applies the findings to your working tree. A background review's edits sit outside your checkpoints, so use git, not /rewind, to undo them.
• --comment posts the findings on a GitHub pull request as inline comments, or on a GitLab merge request as one note.
• --max-findings raises or lowers the usual cap on how many findings it reports.

/review is an alias. Claude can also start the review itself when you ask in plain words, and a scheduled task can run it.

Who publishes it

Anthropic. It is listed as a bundled skill in the Claude Code documentation. Anthropic's claude-code repository also contains a separate code-review plugin that runs four agents on a GitHub pull request, scores each issue for confidence from 0 to 100 and drops anything under 80. That plugin needs the GitHub CLI (gh) installed and signed in.

Where it works

• Claude Code in the terminal and in the desktop app, where the findings appear as a list with file locations.
• Non-interactive runs with claude -p or the Agent SDK, where Claude waits for the review and returns the findings.
• Not in Claude.ai chat. Paste a diff there and ask for a review, or use our Claude Code prompts.

How to install it

It is bundled, so open Claude Code in your repository and type /code-review. To keep it for manual use only, so Claude and scheduled tasks never start it, add this to a settings file such as ~/.claude/settings.json:

{
  "skillOverrides": {
    "code-review": "user-invocable-only"
  }
}

Review my changes on [branch name] for correctness bugs.
Pay extra attention to [risky area, e.g. auth, payments, migrations].
List each finding with file and line, and do not change any code yet.

Deeper reviews: ultra and security

• /code-review ultra runs a multi-agent review in a cloud sandbox. The command reference says it includes 3 free runs on Pro and Max, after which it needs usage credits. It needs a claude.ai login and is not available on Amazon Bedrock, Google Cloud's Agent Platform or Microsoft Foundry; there, it falls back to a local review.
• /security-review checks the diff between your branch and the default branch on origin for risks such as injection, auth issues and data exposure.
• Code Review for GitHub is a separate managed service for Team and Enterprise plans, in research preview. It installs a GitHub app, reviews pull requests automatically and posts inline comments tagged Important, Nit or Pre-existing. Anthropic says each review averages $15 to $25, billed as usage credits. You can steer it with a REVIEW.md file, which the local /code-review does not read.

When it helps

• Before you push, to catch logic errors while the change is fresh.
• Before you ask a teammate to review, so their time goes to design questions.
• On a pull request from someone else, by passing its number.

Limits

• It looks for correctness bugs. Formatting, style and test coverage stay the job of your linter and test suite.
• Higher effort levels may include findings that turn out to be wrong. Read each one before you accept a fix.
• It follows your CLAUDE.md, so project rules written there shape what it flags. Rules kept only in REVIEW.md are ignored locally.

For the rest of the commands, see our Claude Code cheat sheet. Other skills are in the Claude Skills catalog, and our skills manual explains how bundled skills relate to the ones you write. More prompts for developers are in coding prompts.

FAQ

Do I need to install a code review skill for Claude Code?

No. /code-review is bundled with Claude Code. Type it in a session inside your repository.

Can it fix the bugs it finds?

Yes. Add --fix to apply the findings to your working tree, or ask Claude to fix them after you read the report.

Can it post comments on a GitHub pull request?

Yes. --comment posts the findings as inline comments on a GitHub PR, or as a single note on a GitLab merge request.