This exists so that a thorough review does not have to come out of your personal Claude subscription. Each repository has its own limited budget, so use it with care: pick the depth the change actually needs rather than reaching for the deepest one.
Comment on a pull request:
@claude review
The phrase has to start the comment. A quote reply or a mention mid-sentence does nothing.
To rerun, post a new comment. Editing one that already starts with the phrase does nothing.
Add a depth after the phrase. The default is medium. Cost and time scale with both the depth and the size of the pull request, so max on a large pull request is expensive.
@claude review high
| Depth | What you get |
|---|---|
low |
Clear bugs only, in the changed lines. No tests, cleanup or docs. |
medium |
Bugs, tests, cleanup and docs. Only findings it is confident about. |
high |
The same four areas, each separately, plus less certain findings. |
xhigh |
high, then another look for what it missed. |
max |
xhigh, then double-checks every finding. |
Depth is the only argument. Everything else in the comment is ignored.
You and the pull request’s author both need write access to the repository.
The review comes from a skill in the repository being reviewed, at .claude/skills/<name>/. The workflow runs /<name> <depth> and posts whatever the skill returns as its final message.
Each repository’s .github/workflows/claude-review.yml picks the skill through the review-skill input, which defaults to pr-review. Edit the skill to change what reviews look for. The workflow itself is shared and does not need touching.