Skip to content

Review PR

Thoroughly review a pull request by examining the diff, checking CI status, running tests locally, and providing structured feedback. Can optionally post the review directly to GitHub.

How It Works

  1. Get PR details -- Fetch the title, description, changed files, additions, deletions, and commit list using gh pr view.
  2. Read the diff -- Examine all code changes with gh pr diff.
  3. Check CI -- Verify whether CI checks passed with gh pr checks.
  4. Review the code -- Evaluate changes against six criteria (see below).
  5. Run tests locally -- Check out the branch with gh pr checkout and run the test suite.
  6. Provide feedback -- Summarize findings with an overall assessment and specific line-level comments.
  7. Post the review -- Optionally submit the review to GitHub with gh pr review.

Review Criteria

Criterion What to Check
Correctness Does the logic do what the PR title and description claim?
Tests Are new features and bug fixes covered by tests?
Security Any hardcoded secrets, SQL injection, XSS, or command injection?
Style Does the code follow project conventions from CLAUDE.md?
Dependencies Are new dependencies added to requirements.txt or pyproject.toml?
Breaking changes Could this change break existing functionality or APIs?

Example Review

# Step 1: Fetch PR details
$ gh pr view 42 --json title,body,files,additions,deletions
{
  "title": "feat: add CSV export endpoint",
  "additions": 147,
  "deletions": 12,
  "files": [
    {"path": "src/app/export.py", "additions": 95},
    {"path": "src/app/routes.py", "additions": 22},
    {"path": "tests/test_export.py", "additions": 30}
  ]
}

# Step 2: Read the diff
$ gh pr diff 42

# Step 3: Check CI
$ gh pr checks 42
All checks passed

# Step 4-5: Checkout and test
$ gh pr checkout 42
$ pytest --tb=short -q
18 passed in 5.23s

Feedback Format

The review follows a consistent structure:

## PR Review: feat: add CSV export endpoint (#42)

### Overall Assessment
**Approve** -- Clean implementation with good test coverage.

### Feedback

**src/app/export.py:34** -- The CSV writer does not escape commas
in field values. Consider using `csv.writer` instead of manual
string concatenation.

**src/app/export.py:67** -- This query fetches all rows without
a LIMIT. For large tables, this could cause memory issues. Add
pagination or streaming.

**src/app/routes.py:45** -- Good: the endpoint requires
authentication. No security concerns.

### Suggestions
- Add a `Content-Disposition` header so browsers download the
  file instead of displaying it inline.
- Consider adding a row count limit parameter with a sensible
  default (e.g., 10,000 rows).

Good vs Bad Reviews

Bad: vague, unhelpful feedback

Looks good to me.
This code is bad. Please fix.

Good: specific, actionable, with line references

src/app/export.py:34 -- The CSV writer uses string concatenation
(`",".join(row)`) which breaks when field values contain commas.
Replace with `csv.writer` from the standard library:

    import csv
    writer = csv.writer(output)
    writer.writerow(header)
    writer.writerows(data)

Posting Reviews to GitHub

After completing the review, the skill can submit it directly:

# Approve
$ gh pr review 42 --approve --body "Clean implementation, good tests."

# Request changes
$ gh pr review 42 --request-changes --body "See inline comments for required fixes."

# Comment only (no approval or rejection)
$ gh pr review 42 --comment --body "A few suggestions, nothing blocking."

Notes

  • The skill always checks out the branch and runs tests locally before providing feedback. Remote CI alone is not sufficient.
  • When reviewing large PRs (>500 lines changed), the skill focuses on the most critical files first and flags the PR size as a concern.
  • Security issues (hardcoded secrets, injection vulnerabilities) are always flagged as "request changes" regardless of other code quality.
  • The review checks for consistency with CLAUDE.md project conventions if that file exists in the repository.