Review Skills
Seven review skills that make any SKILL.md-compatible agent file the same report every time: findings ranked P0–P3, each one anchored to a path:line you can open.
Claude Code already ships strong first-party review — a built-in /security-review, an official code-review plugin, and a security-review GitHub Action — and this does not try to out-detect any of them. What it adds is the shape around them: one consistent, evidence-first output format across seven distinct review dimensions, five of which no first-party review tool covers at all. The same seven skills run unchanged in Codex, Cursor, and Copilot CLI, where the Claude-only tooling does not run, and each ships an agents/openai.yaml so it appears as a properly named action in Codex rather than an untitled prompt.
Every audit skill is read-only by default. It reports; it does not edit, until you explicitly ask for fixes.
npx skills add specialone0007/review-skills --skill '*'
Then just ask:
audit my repo for launch risks
security audit this repo
what tests am I missing
Given no scope, every audit surveys the whole repository, ranks surfaces by risk, inspects the riskiest in depth, and tells you what it only skimmed. Name a feature, route, PR, or branch and it narrows to that instead.
Example output
**Security Audit: exports**
No code changed. I reviewed the export API and its permission checks. P0s found: 0.
1. **P1: Team members can request another team's export by ID.**
Abuse path: an authenticated user guesses or increments an export ID.
Impact: cross-tenant data disclosure.
Evidence: `app/api/exports.ts:88`.
Suggested mitigation: scope the lookup by the caller's team before returning the row.
A complete run is in examples/.
Skills
| Skill | Use it when | Not for |
|---|---|---|
feature-audit |
Launch-readiness or bug-risk review of a feature, route, workflow, PR — or the whole repo. | Narrower asks that a specialist below covers. |
security-audit |
AppSec review: auth, authorization, injection, secrets, data exposure, dependencies, abuse paths. | Broad product readiness or non-security defects. |
test-gap-audit |
Which tests are missing, weak, stale, or insufficient. | Finding product bugs rather than evaluating coverage. |
docs-sync-audit |
Comparing code, APIs, config, commands, or examples against the docs. | Implementation review or code-quality feedback. |
repo-health-audit |
Structure, naming, dead code, duplication, reuse, module boundaries. | Runtime behavior, readiness, security, or coverage. |
feature-brainstorm |
Evidence-grounded product, UX, workflow, or technical improvement ideas. | Defects, blockers, or launch risks reported as findings. |
pr-branch-summary |
PR titles, descriptions, release comparisons, team summaries from branch diffs. | Review findings or implementation changes. |
Each skill names its nearest neighbours in its own ## Related Skills section, so the agent can route itself if you pick the wrong one.
Compared to first-party tooling
| First-party (Claude Code) | Review Skills | |
|---|---|---|
| Bug and security detection | /security-review, the official code-review plugin, security-review Action. Use them. |
Not a replacement. Same class of finding. |
| Report shape | Varies by tool and run | One contract: P0–P3 severity plus path:line evidence, no finding without both |
| Dimensions covered | Security, correctness, performance, maintainability | Those, plus test-coverage gaps, documentation drift, repo structure and duplication, improvement ideation, and PR communication |
| Runs in Codex, Cursor, Copilot CLI | No | Yes, unchanged, with Codex interface metadata |
| Cost | Built in on paid plans | MIT, no server, no account |
If you only use Claude Code and only want security and correctness findings, use the first-party tools. This exists for the other five dimensions, and for teams whose agents are not all Claude Code.
Install
With the skills CLI:
npx skills add specialone0007/review-skills --skill '*' # all seven
npx skills add specialone0007/review-skills --skill feature-audit # just one
As a Claude Code plugin:
/plugin marketplace add specialone0007/review-skills
/plugin install review-skills@review-skills
Or copy a skill folder into your agent's skills directory:
.claude/skills/<skill-name>/
.agents/skills/<skill-name>/
.github/skills/<skill-name>/
.cursor/skills/<skill-name>/
~/.codex/skills/<skill-name>/
~/.cursor/skills/<skill-name>/
Each folder holds a SKILL.md, an agents/openai.yaml with Codex interface metadata, and, for most of them, a bundled read-only Python script that uses only the standard library. Scripts are accelerators; every skill still works without them.
Safety
All seven skills are read-only on your files by default. They instruct the agent not to edit, stage, or commit anything unless you explicitly ask for fixes or implementation.
One precise exception: pr-branch-summary may run git fetch origin <base> when the base branch you asked to compare against is missing locally. That updates a single remote-tracking ref and never touches your working tree, index, or local branches. It is skipped when the ref already exists.
To authorize implementation, ask explicitly — "fix these findings", "add the suggested tests", "update the stale docs".
As with any third-party agent skill, read the skill contents before enabling it in a trusted environment. Everything here is plain Markdown plus a handful of small standard-library Python scripts, so it is quick to audit.
Contributing
See CONTRIBUTING.md. Before opening a PR:
python tools/validate_skills.py # spec limits and repo conventions
python tools/validate_evals.py # eval cases and script snapshots
CI runs both, on Ubuntu and Windows. Each skill has trigger, anti-trigger, and behavior cases in evals/, exercised against a deliberately defective fixture.
License
MIT — see LICENSE.
No comments yet
Be the first to share your take.