Installs into .claude/skills of the current project.
Are you the author of Requesting Code Review?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/yishan-io-requesting-code-review)
---
name: requesting-code-review
description: Use when finishing substantial code changes, before merging, or at natural checkpoints to request an explicit code review pass.
---
# Requesting Code Review
Use this skill to request a focused review before changes continue or ship.
## Core Principle
Review early enough to catch issues before they compound, and review with concrete scope.
## When to Use This Skill
Use this skill:
- After completing a substantial task or feature
- Before merging or handing work off
- After a complex bug fix or refactor
- At natural checkpoints in multi-step work
- When a fresh set of eyes would reduce risk
For tiny mechanical edits, a separate review pass is optional.
## What To Review
Prepare a tight review scope:
- What changed
- What behavior was intended
- What files or diff range matter
- What risks deserve attention
Good review requests are specific. Avoid asking for a vague review of the entire project when only one change matters.
## How to Request Review
### 1. Identify the review scope
Prefer one of these:
- Current uncommitted diff
- A specific commit range
- A completed task's touched files
If using commits, gather the relevant base and head SHAs.
### 2. Dispatch a review pass
Use the `code-reviewer` agent for read-only code review.
Include:
- A short summary of the change
- The intended behavior or requirements
- Any known risks or areas you are unsure about
- The files or diff range to inspect
Example request contents:
```text
Review this change for bugs, regressions, missing tests, and risky assumptions.
Summary: Added a new retry path for transient API failures.
Expected behavior: Requests retry once on network timeout, but not on validation errors.
Scope: diff between <base> and <head>.
Focus: error handling, duplicate side effects, and test coverage.
```
### 3. Act on the findings
- Fix blocking issues that are within the stated review scope before continuing or shipping
- Record non-blocking or out-of-scope findings as follow-up work instead of extending the current work
- If a finding seems wrong, verify it against the code and push back with evidence
## Bounded Review
Keep every review bounded by its concrete scope and intended behavior. After fixing findings, request at most one follow-up review to verify those fixes.
If the user explicitly requests review rounds beyond the initial review and one follow-up, stop before dispatching more reviews. Create a new review brief with a finite review budget that states the additional scope, intended behavior, and exact number of additional rounds. Do not run an open-ended "review until clean" loop.
Review does not close a tracked task. The user controls task closure.
## Review Cadence
For larger efforts:
- Review after meaningful checkpoints instead of waiting until the very end
- Avoid stacking multiple risky changes without feedback in between
For smaller efforts:
- One review near completion is usually enough
## Red Flags
Do not:
- Skip review just because the change feels simple
- Ask for review without stating the intended behavior
- Ignore serious findings and continue anyway
- Treat review as a formality instead of a real risk check
## Bottom Line
Request review with concrete scope, clear expectations, and enough context for someone to find real issues quickly.