| name | code-review-checklist |
| description | Walks a diff against a 5-point review checklist. Use when the user asks "review this diff", "what should I check in this PR", or pastes code with a question like "anything wrong with this?". Reads files via shell; never edits. |
Code review checklist
Walk this diff against the five questions below. Stop after the first failure;
the human can re-run after fixing.
When to invoke
- User says "review this diff", "check this PR", "anything wrong here?"
- User pastes a chunk of code with no explicit question (default to review)
- User asks for a verdict — "ship?", "merge?", "is this safe?"
Procedure
-
Read the full diff first via bash (git diff HEAD or read the file).
Don't comment on individual hunks before seeing the whole change.
-
Five questions, in order. Stop at the first failure.
a. Does it pass the type-check / lint? If a CI signal exists, cite it.
If not, scan for obvious type errors yourself.
b. Are there tests? New behavior without a test is a smell. Existing
behavior changed without a test update is a smell.
c. What's the riskiest line? Pick ONE — the line that, if wrong, breaks
production. Verify it explicitly.
d. Backwards compat? Did any public API change shape (params, return,
schema)? Find callers via grep -rn or rg.
e. Performance / blast radius? Does it add a loop in a hot path? A new
DB query in a request handler? A retry without backoff?
-
Output format.
## Review verdict — <ship | hold | request-changes>
**Riskiest line:** <path:line> — <one sentence why>
**Notes:**
- <path:line> — <issue>
- <path:line> — <issue>
What this skill does NOT do
- Doesn't fix the issues — propose changes as a diff in chat. The hook will
block
apply_patch even if you try; the human applies any fix.
- Doesn't approve or block in source control — that's the human's call.
- Doesn't engage in style debates.
Cross-runtime note
This skill is portable — the same SKILL.md works in .claude/skills/ for
Claude Code and .codex/skills/ for Codex CLI. The Agent Skills standard
(https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills)
means one file, many harnesses.
