requesting-code-review · git:20260324.691be9c · 2026-03-24 · sha256 a8daee7c29493949
requesting-code-review git:20260324.691be9cA
Immutable. This exact content never changes and is served at /api/v1/blob/a8daee7c29493949.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_z7wv2os74zg2ejwl/a8daee7c29493949b337d39969585c7d247bb7e6fea7d4e223884b3db1dbb22c.json
Then /plugin install requesting-code-review@markdownregistry-z7wv2os74zg2ejwl in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: requesting-code-review description: >- Reviews the agent's own work against the original plan and design before presenting it to the user or submitting a PR. Use this skill when a task or batch of tasks is complete, between implementation phases, before creating a pull request, or when the agent is about to declare "done." Critical issues block progress — no exceptions. --- # Requesting Code Review Review your own work against the plan. Flag issues by severity. Critical issues block progress. ## When to Use - A task or batch of tasks is complete - Between major implementation phases - Before creating a pull request - Before telling the user "done" on any non-trivial change - After a subagent completes delegated work ## How to Use ### Step 1: Compare Output to Plan Pull up the original plan from `tasks/todo.md` and the design (if one exists): 1. **Does the implementation match the spec?** — Compare actual behavior to planned behavior 2. **Are all acceptance criteria met?** — Check each criterion explicitly 3. **Were any shortcuts taken?** — Note where the implementation deviated from the plan ### Step 2: Self-Review Checklist Run through this checklist for every completed task: | Check | Status | |-------|--------| | Implementation matches the plan spec | ✅/❌ | | All tests pass (including new ones) | ✅/❌ | | No debug artifacts left (console.log, TODO hacks) | ✅/❌ | | Error handling covers edge cases | ✅/❌ | | No security issues introduced | ✅/❌ | | Code follows project conventions | ✅/❌ | | Diff scope matches task scope (no extra changes) | ✅/❌ | ### Step 3: Flag Issues by Severity ```markdown ## Self-Review Summary ### 🔴 Critical (Blocks Progress) - [issue] — Must fix before moving forward **Action:** [what needs to change] ### 🟡 Major (Should Fix) - [issue] — Important but doesn't block the current task **Action:** [what needs to change] ### 🟢 Minor (Note for Later) - [issue] — Nice to have, not urgent **Action:** [log for future] ### ✅ Passing - Implementation matches spec - Tests cover the new behavior - Code follows existing patterns ``` ### Step 4: Act on Findings - **🔴 Critical** → STOP. Fix immediately. Do not proceed to the next task. - **🟡 Major** → Fix before the PR, but can continue to the next task - **🟢 Minor** → Log in `tasks/todo.md` as a follow-up item - **✅ All clear** → Proceed to the next task or submit ### Step 5: Present Summary Always show the user a brief review summary before declaring completion: ```markdown **Review: Step 3 — Implement API endpoints** - ✅ 4/4 endpoints implemented per spec - ✅ Input validation on all routes - 🟡 Missing rate limit on POST /users (logged for Step 6) - ✅ 12 new tests, all passing - Proceeding to Step 4. ``` ## Examples **Bad — No Self-Review:** > Agent: "Done! I implemented all the endpoints." > User: *finds missing error handling, wrong response codes, and a SQL injection* **Good — Self-Review Catches Issues:** > Agent: "Implementation complete. Self-review: > - 🔴 The DELETE endpoint doesn't check authorization — fixing now. > - ✅ All other endpoints match spec, 12 tests passing. > - 🟢 Could add pagination to GET /users, logging for future. > > Fixing the auth check before proceeding..." ## Guidelines - Self-review is mandatory between tasks, not optional - Critical issues are non-negotiable blockers — fix before moving forward - The review checks the plan, not just the code — did we build what we said we'd build? - Be honest in self-review — hiding issues from yourself helps nobody - Use the `code-review` skill format for detailed line-by-line review when needed ## Anti-Patterns - **Skipping self-review** — "It works, ship it" is not a review - **Downgrading severity** — A missing auth check is critical, not minor - **Reviewing only happy paths** — Check edge cases and error paths too - **Proceeding past critical issues** — The whole point is that criticals block - **Rubber-stamp reviews** — "✅ All good" without actually checking is worse than no review