ai-code-skills

A growing collection of Agent Skills for Claude Code, Codex, and OpenCode. Review code for security, performance, and clean code — pull requests or your local changes before you push — audit a site's technical SEO + GEO (Generative Engine Optimization) for search and AI-assistant visibility, and delegate a task to another model's CLI (Codex, OpenCode, Claude Code, Cursor, Gemini, Aider). Most skills write their results to one evidence-based Markdown report with a clear verdict.

Each skill is self-contained. You pick which ones to install.

Skills in this repo

More skills will be added over time — list always shows what's available.

Every review produces one Markdown report with evidence, severity, confidence, rationale, recommendations, verification steps, and references. The PR reviewers write PR_REVIEW.md with a merge verdict and support PR-number-aware diffs and opt-in inline PR comments; pre-push-review writes PRE_PUSH_REVIEW.md from your local diff with a Pass/Warn/Fail scorecard and a push-readiness recommendation. technical-seo-geo-audit writes SEO-GEO-AUDIT.md — findings grouped per page with a separate SEO score and GEO score and a copy-pasteable code fix for each; its rendered-DOM and Core Web Vitals checks prefer the chrome-devtools MCP server and fall back to optional Playwright + PageSpeed Insights. Shared features across the review skills: Quick/Standard/Deep modes and automatic stack detection. delegate is the exception — instead of a report, it hands a task to another model's CLI headlessly and summarizes the answer, always confirming the exact command before it spends the target's credits.

Install (npm / npx)

You choose which skills to install — installing is never all-or-nothing unless you ask for --all.

# See what's available
npx ai-code-skills list

# Install specific skills (user scope, both clients)
npx ai-code-skills install nextjs-pr-review
npx ai-code-skills install nextjs-pr-review laravel-pr-review

# Interactive picker (run with no skill names in a terminal)
npx ai-code-skills install

# Everything
npx ai-code-skills install --all

# A single client
npx ai-code-skills install laravel-pr-review --claude
npx ai-code-skills install laravel-pr-review --codex
npx ai-code-skills install laravel-pr-review --opencode

# All three clients at once
npx ai-code-skills install --all --all-clients

# Commit a skill into one repository (project scope)
npx ai-code-skills install nextjs-pr-review --project --root ./my-app

# Inspect or remove
npx ai-code-skills where --all
npx ai-code-skills uninstall laravel-pr-review

Install the CLI globally if you prefer:

npm install -g ai-code-skills
ai-code-skills install nextjs-pr-review

Install destinations

Project-scope copies can be committed so every contributor gets the same review rules.

OpenCode note: OpenCode natively discovers SKILL.md skills and also reads ~/.claude/skills/ and ~/.agents/skills/. So the default --both install already works in OpenCode — invoke skills there the same way. Use --opencode (or --all-clients) only if you also want the copy in OpenCode's native ~/.config/opencode/skills/ path. Client flags combine, e.g. --claude --opencode.

Without Node

Cloned the repo? The shell installer does the same thing and also requires you to choose:

./install.sh --list
./install.sh --user --both nextjs-pr-review
./install.sh --user --both --all
./uninstall.sh laravel-pr-review

Use

After installing, invoke a skill by name.

Claude Code

/nextjs-pr-review Review PR #123 and save the report to docs/reviews/PR-123.md
/laravel-pr-review Deep review of the current branch against origin/main.
/pre-push-review Check my local changes before I push.
/technical-seo-geo-audit Audit https://example.com — technical SEO and AI/LLM visibility.
/delegate --to codex Refactor src/auth.ts and add tests. (edit mode; confirm first)

Codex

$nextjs-pr-review Review PR #123
$laravel-pr-review Review the current branch, focus on the auth changes.
$pre-push-review Score my staged changes before I push.
$technical-seo-geo-audit Audit https://example.com for technical SEO and GEO.
$delegate --to opencode Explain what this module does. (read-only, the default)

Both agents may also activate a skill automatically when asked to review a matching PR, or to check local changes before a push.

Review modes

Mention a mode to control depth (default is Standard):

• Quick — fast gate for small/low-risk diffs; High/Critical security + correctness only.
• Standard — all phases, all three categories, safe validation, full report.
• Deep — security-sensitive or large diffs; end-to-end tracing, build/test/static-analysis runs, expanded checklists, and a self-audit pass.

Inline PR comments (opt-in)

For the PR reviewers (nextjs-pr-review, laravel-pr-review), off by default — the report is the deliverable. Ask explicitly to also post findings inline (requires an authenticated gh):

/laravel-pr-review Review PR #123, then post the High/Critical findings as inline comments.

pre-push-review works on local changes with no PR, so it never posts comments — its deliverable is the report plus the Pass/Warn/Fail scorecard and push-readiness recommendation.

The skills never approve, request changes, merge, or push on your behalf unless you say so.

Repository layout

ai-code-skills/
├── package.json            # npm package + bin
├── bin/
│   └── cli.js              # list / install / uninstall / where (you pick skills)
├── install.sh              # shell installer (alternative to npx)
├── uninstall.sh
├── README.md
├── CHANGELOG.md
├── LICENSE
└── skills/
    ├── nextjs-pr-review/
    │   ├── SKILL.md
    │   ├── assets/PR_REVIEW_REPORT_TEMPLATE.md
    │   ├── references/{SECURITY,PERFORMANCE,CLEAN_CODE}.md
    │   └── scripts/{pr-diff.sh,detect-stack.sh,validate-report.py}
    ├── laravel-pr-review/
    │   ├── SKILL.md
    │   ├── assets/PR_REVIEW_REPORT_TEMPLATE.md
    │   ├── references/{SECURITY,PERFORMANCE,CLEAN_CODE}.md
    │   └── scripts/{pr-diff.sh,detect-stack.sh,validate-report.py}
    ├── pre-push-review/
    │   ├── SKILL.md
    │   ├── assets/PRE_PUSH_REVIEW_REPORT_TEMPLATE.md
    │   ├── references/{SECURITY,PERFORMANCE,CLEAN_CODE,SCORING}.md
    │   └── scripts/{local-diff.sh,detect-stack.sh,validate-report.py}
    ├── technical-seo-geo-audit/
    │   ├── SKILL.md
    │   ├── package.json                 # optional deps: playwright, cheerio
    │   ├── assets/SEO_GEO_AUDIT_REPORT_TEMPLATE.md
    │   ├── references/{seo-checks,geo-checks,severity-rubric}.md
    │   └── scripts/
    │       ├── crawl.mjs report.mjs run.mjs static.mjs
    │       ├── checks/{crawlability,rendering,performance,metadata,semantics,geo}.mjs
    │       ├── lib/{fetch,html,findings,args}.mjs
    │       ├── fixtures/ selftest.mjs
    │       └── detect-stack.sh validate-report.py
    └── delegate/
        ├── SKILL.md
        ├── references/TARGETS.md         # per-tool adapter matrix + gotchas
        └── scripts/{detect-clis.sh,delegate.sh}

Adding a new skill

1. Create skills/<your-skill>/SKILL.md with YAML frontmatter (name, description, license).
2. Add references/, scripts/, and assets/ as needed (the diff/validate scripts are reusable).
3. That's it — the CLI auto-discovers any directory under skills/ containing a SKILL.md, so list and install pick it up with no code changes.

Helper scripts (per skill)

skills/<skill>/scripts/pr-diff.sh origin/main       # structured PR/branch diff (PR reviewers)
skills/<skill>/scripts/pr-diff.sh --pr 123           # needs authenticated gh
skills/pre-push-review/scripts/local-diff.sh         # local uncommitted diff (pre-push reviewer)
skills/pre-push-review/scripts/local-diff.sh --staged   # or --unpushed / --all-local
skills/<skill>/scripts/detect-stack.sh               # summarize the stack
python3 skills/<skill>/scripts/validate-report.py <report>.md

skills/delegate/scripts/detect-clis.sh               # which model CLIs are installed + how to call each
skills/delegate/scripts/delegate.sh --to opencode --dry-run "hi"   # preview the resolved command
skills/delegate/scripts/delegate.sh --to codex --mode edit "..."   # run it (read-only is the default)

Delegate — target to CLI

License

MIT — see LICENSE.