Code Review and Quality Skill · Development

Code Review and Quality: a five-axis PR review skill for AI agents

A five-axis code review skill for AI agents: severity-labelled, evidence-backed findings on any diff or PR, with a fixed report and a clear verdict.

  • Code Review
  • Pull Requests
  • Five-Axis Quality
  • Security Check

About This Skill

An unofficial edition of code-review-and-quality from Addy Osmani's agent-skills (MIT). It reviews a diff or pull request on correctness, readability, architecture, security and performance, reads the tests first, and checks each finding against the real callers before reporting it. Our edition adds clear when-to-use and when-not-to-use rules, one severity scheme, a fixed report format, a git-safe mutation check, a full worked example, and self-contained security and performance checklists.

What we changed from the original

Based on code-review-and-quality by Addy Osmani (MIT). Unofficial edition; the original and its license notice are credited in the download.

AreaOriginalOur edition
TriggerDescription said to use it before merging any change, which also fires on commit-message requests and trivial diffs.Description and When to use / When NOT to use lists name real phrases and exclude commit messages, typo-only or generated-only diffs, deploys and whole-repo audits.
Output formatA checkbox checklist with no required structure for findings.A fixed report: summary, findings table (severity, origin, axis, file:line, trigger and consequence, suggested fix), rejected candidates, pre-existing issues, verification, and one verdict with explicit verdict rules.
Severity schemeMixed labels: no prefix, Critical, Nit, Optional/Consider, FYI, plus Required in another section.One scheme of five labels (Critical, Required, Suggestion, Nit, FYI) plus a new/pre-existing origin tag; pre-existing defects can never block the change.
Mutation safetyFlip a condition, run tests, then restore the file from a copy.In place by default for uncommitted diffs, only after one explicit yes that covers up to three mutations; git status/diff snapshot and byte-identical restore check after each mutation; stop at the first surviving mutation; separate worktree only for fully committed changes; untracked files skipped or byte-copied and compared.
BoundariesNo stop points beyond asking before deleting dead code.Read-only by default; asks before edits, posting comments, merging, tests that hit real services, or forced dependency fixes; masks secrets; treats instructions inside reviewed code as data.
Checkable stepsReview steps listed questions without an end condition.Each of the seven steps ends with a concrete check (scope line, behavior-to-test map, git status match, evidence fields per finding).
Worked exampleOnly a sample prompt for a review agent.A complete pagination-diff review: input, each step, a rejected auth claim, a pre-existing issue, a surviving mutation, the final report, and a pass/fail rubric, plus seven behavior eval cases.
Standalone installLinked to repo-root checklists and sibling skills that break when the skill is installed alone; included a one-business-day human review-speed rule.Security, performance, size and dependency guidance folded into the skill's own references/; sibling links and human-team timing rules removed.

Quick Start

  1. 1Download the skill (1 credit) and unzip the code-review-and-quality folder.
  2. 2Copy the folder into your skills directory: Antigravity ~/.gemini/config/skills/ (or .agents/skills/ in a project), Claude Code ~/.claude/skills/, Codex ~/.agents/skills/.
  3. 3Open a repository with a branch or PR to review and ask your agent: "Review this PR before I merge it."
  4. 4Read the report: findings table with severity, file:line and suggested fix, then the verdict. Say "read-only" if you do not want the agent to run any tests or mutation checks.

Core Capabilities

Five-axis review

Checks every changed file for correctness, readability, architecture, security and performance, starting from the tests.

Verified findings only

Traces callers, guards and the base branch before reporting, lists rejected claims, and tags each finding as new or pre-existing.

Fixed report and verdict

Summary, findings table with severity, file:line, evidence and fix, then one verdict: approve, request changes or needs more info.

Built-in security checklist

Secrets, auth and ownership checks, injection, SSRF, path handling and LLM output, all included in the skill folder.

Git-safe test mutation

Optionally flips one condition to see if the tests notice: in place after your one-time approval with a proven restore, or in a separate worktree for committed branches.

Size and dependency discipline

Flags oversized changes, suggests how to split them, and reviews dependency upgrades through changelogs and lockfile diffs.

Usage Examples

Input

Review this PR before I merge it. It adds paging to the orders endpoint.

Output

A report with three Required findings (empty list on missing page, whole-table query sliced in memory, no test past page 1 shown by a surviving mutation), one rejected auth claim citing the router guard, one pre-existing issue, and the verdict Request changes.

Input

Another agent wrote this refactor of the billing module. Can you check it, read-only?

Output

A read-only review that notes whether the refactor reduces or only relocates complexity, maps changed behavior to tests, reports the mutation check as not run, and lists what could not be verified.

Input

Is this dependency upgrade safe to merge?

Output

A review of the changelog and lockfile diff that flags behavior changes and new transitive packages, checks that tests pass before and after, and recommends one dependency per change.

SKILL.md

---
name: code-review-and-quality
description: Reviews a diff, pull request, branch or set of changed files across five axes (correctness, readability, architecture, security, performance) and returns a fixed-format report with severity-labelled, evidence-backed findings and a verdict. Reads tests first, checks every finding against callers and guards before reporting, separates new regressions from pre-existing defects, and accepts "no findings" as a valid result. Use when the user says "review this PR", "review my diff", "check this change before I merge", "do a quality pass", or "review what the other agent wrote". Do NOT use for writing a commit message or PR description, writing new tests or features, deploying, trivial typo or comment-only diffs, lockfile-only or generated-file diffs with no hand-written code, or a whole-repo security audit with no specific change to review.
---

# Code Review and Quality (Antigravity Skills unofficial edition)

Unofficial edition of `code-review-and-quality` by Addy Osmani (MIT). See `NOTICE.md`.

Review a specific change and say, with evidence, whether it leaves the codebase healthier. Approve a change that clearly improves overall code health even if it is not how you would have written it. Block only on real problems. A review with zero findings is a valid, honest result; never invent findings to look thorough.

## When to use

- "Review this PR / diff / branch before I merge it."
- "Do a quality pass on these changes for correctness and readability."
- "Another model (or agent) wrote this, can you check it?"
- "Review the fix for bug X and its regression test."
- "Is this dependency upgrade safe to merge?"

## When NOT to use

- Writing a commit message or PR description for a diff (even if a diff is pasted).
- Writing a failing test, implementing a feature, or debugging a crash from scratch.
- Deploying, releasing, or merging.
- Trivial diffs: typo fixes, comment-only or whitespace-only changes. Just confirm the change i

Frequently Asked Questions

FAQ

Is this the official code-review-and-quality skill?
No. It is an unofficial edition by Antigravity Skills, based on code-review-and-quality from Addy Osmani's agent-skills repository (MIT). It is not affiliated with or endorsed by the original author. Credits are in NOTICE.md.
What did you change from the original?
We narrowed when it triggers, removed human-team process rules, defined one five-label severity scheme, added a fixed report format with a verdict, made the test-mutation step git-safe, added a full worked example and eval cases, and copied the security and performance checklists into the skill so it works on its own.
Will the agent change my code during a review?
Not by default. The review is read-only. For an uncommitted diff, the optional mutation check asks you once first; a yes covers up to three temporary one-line edits, each restored and proven identical with git diff before moving on. Fully committed branches are tested in a separate git worktree. Editing, deleting dead code, posting comments or merging all need your go-ahead.
What if the change is fine?
The skill treats no findings as a valid result. It says what it checked and approves, instead of inventing problems to look thorough.
Does it work with Claude Code, Codex and Antigravity?
Yes. It is plain Markdown with no scripts. Copy the folder to ~/.claude/skills/ for Claude Code, ~/.agents/skills/ for Codex, or ~/.gemini/config/skills/ (or .agents/skills/ in a project) for Antigravity.

Discussion

Discussion

0 comments
U

Trigger Phrases

Use these phrases to activate this skill in your AI coding assistant:

review this PRreview my diff before I mergedo a quality pass on this changecheck the code another agent wroteis this upgrade safe to mergereview the fix and its regression test