GitHubSource
anthropics/claude-code-action
by anthropics · 2025-05-19 (repo created)
✓ Repo metadata and README fetched from GitHub on 2026-10-15
Read the original ↗Part 1 · What the source did
Anthropic's official action runs Claude Code on your own GitHub runner to review PRs, answer @claude questions and implement small changes.
- Detects whether it was triggered by an @claude mention, an assignment or an explicit prompt.
- Reviews PR changes, answers code questions and can implement fixes.
- Shows progress as checkboxes and can return structured JSON outputs.
Summary written by Agent Cookbook. See the original for the full context.
Part 2 · Our copy-paste version
Adapted by us from what the source described. Fill in the [BRACKETS].
The prompt
Review this PR as a senior engineer.
1. Summarize what changed in 3 bullets.
2. List bugs, security issues and missing tests as review comments on specific lines, highest risk first.
3. Do NOT push commits. If a fix is trivial, suggest it as a code suggestion block.
4. End with: APPROVE-WORTHY / NEEDS CHANGES and why.
🛡
Approval ruleRestrict tools so it comments rather than pushes; humans merge.
Set it up
- Add the API key as a repository secret.
- Install the workflow from the repo's quick start.
- Put the prompt above in the workflow's prompt input.
- Open a test PR and check the comments.
Common failures
⚠
What usually goes wrong
- Prompt injection from PR text or issues. Limit tools and don't give write tokens.
- Costs grow on busy repos; trigger on mention only.
Variations
- Trigger only when someone comments @claude.
- Return JSON outputs to drive labels.