Claude Code Skills

What It Does

Claude Code Skills (alirezarezvani/claude-skills) is a multi-domain skill framework containing 330+ SKILL.md files across 16 domains: engineering, marketing, C-level advisory, product, project management, regulatory/quality, finance, business growth, research, productivity, and more. Each skill follows a consistent package pattern: a SKILL.md master doc, a scripts/ directory with Python CLI tools (stdlib only), a references/ directory with expert knowledge bases, and optional assets/ for templates. The repo ships platform adapters for 12 coding agents including Claude Code, OpenAI Codex, Gemini CLI, Cursor, Aider, Windsurf, OpenClaw, Hermes Agent, Kilo Code, OpenCode, Augment, and Antigravity.

The Good

Massive scale with consistent structure. 330+ SKILL.md files, 531 Python scripts, 656 reference documents, 210 scripts/ directories, and 58 assets/ directories. The median SKILL.md is 204 lines. Every skill we inspected uses proper YAML frontmatter with name, description, license, and metadata fields. The package pattern (SKILL.md + scripts/ + references/ + assets/) is uniform across all 16 domains.

Zero-dependency Python tools that run. Every Python script we tested uses only the standard library. We ran 6 scripts across marketing, research, and engineering domains. All produced real output with no pip installs needed. The URL pattern generator produced a scored report with warnings. The fiscal year calculator output NIH-compatible query parameters. The SEO checker has a proper CLI with --file, --url, and --json flags.

12-platform support with conversion tooling. The repo ships pre-built platform trees (.claude-plugin/, .codex/, .codex-plugin/, .gemini/, .hermes/) plus scripts/convert.sh that generates integrations for Cursor, Aider, Windsurf, Kilo Code, OpenCode, Augment, and Antigravity. The marketplace.json declares 59 plugin entries with version numbers, keywords, and per-domain descriptions. Each platform gets its own sync script (sync-codex-skills.py, sync-gemini-skills.py, sync-hermes-skills.py).

Strong CI and quality infrastructure. 11 GitHub Actions workflows run on PRs: quality gate, skill quality review (Tessl), security audit, VirusTotal scan, PR target enforcement, smart sync, and release automation. The CI detects changed skill directories and runs targeted quality checks. There is a SECURITY.md with a responsible disclosure policy. The branch protection enforces conventional commits and PR-based merges.

Attribution of derived work. Skills derived from external sources carry proper attribution. The handoff skill credits Matt Pocock with a link to the original. The research skills link to their megaprompts source specs. Metadata fields include original_author, original_license, and derived_from URLs.

The Bad

Quality variance between skills is wide. The thinnest SKILL.md is 41 lines (handoff). The thickest is 740 lines (terraform-patterns). Skills under 60 lines exist in finance (55), marketing (55), business-growth (56), engineering (58), project-management (59), and product-team (61). These thin skills are basically stubs: a short description, a quick reference table, and an install command. Meanwhile, the top-end skills like terraform-patterns (740 lines), senior-frontend (572), and code-to-prd (507) are detailed enough to be useful. This gap means the "328 production-ready skills" claim needs an asterisk.

25 dead links to gitignored megaprompts/ directory. The research skills (all 8 of them: litreview, grants, dossier, patent, pulse, syllabus, notebooklm, research) contain links like megaprompts/08-grants-megaprompt.md that point to a directory excluded from the repo via CLAUDE.md ("maintainer-local, gitignored"). Anyone cloning this repo gets broken links. The documentation explicitly acknowledges this is intentional, but it is still a poor experience for users.

No automated tests for the skill content itself. There are 531 Python scripts but no test suite in the public repo. The tests/ directory is gitignored as "maintainer-local." The CI quality gate runs linting and doc generation but does not execute the Python tools. We found 1 test file (engineering/skills/skill-tester/tests/test_security_scorer.py) in the tree, but no pytest or unittest runner configuration. For a repo with 531 scripts, the absence of a public test suite is a gap.

Agent and persona definitions are YAML-free stubs. The agents/ directory has 11 subdirectories but no SKILL.md files. Agent definitions live as plain markdown. The commands/ directory has 38 slash commands as bare .md files with frontmatter. The orchestration/ directory has a single ORCHESTRATION.md. These are lighter weight than the full skill packages and rely on Claude Code's plugin system to function, making them harder to evaluate or port.

README and marketplace.json counts are inconsistent. The README headline says "313 production-ready skills." The marketplace.json description says "328." The CLAUDE.md says "328" with a v2.8.0 claim. The actual SKILL.md count (including platform trees) is 721, or roughly 330 if you exclude .gemini/, .codex/, .hermes/ duplicates. The star badge in the README still says "5,200+ GitHub stars" when the repo has 15,166. These discrepancies suggest marketing copy is not keeping up with the actual repo state.

Smoke Test Results

Tested on macOS (host, aarch64) against a shallow clone of v2.8.0 (commit f398e16). No sandbox available. All commands run directly from the cloned repo.

Structural validation

$ git clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills-r4
✅ Cloned in 8s. 42 MB on disk.

$ find . -name "SKILL.md" -not -path "./.git/*" | wc -l
✅ 721 SKILL.md files total (including platform trees)

$ find . -name "SKILL.md" -not -path "./.git/*" -not -path "./.gemini/*" -not -path "./.codex*" -not -path "./.hermes/*" | wc -l
✅ 330 unique SKILL.md files (excluding platform duplicates)

$ find . -name "*.py" -not -path "./.git/*" | wc -l
✅ 531 Python scripts

$ head -5 LICENSE
✅ MIT License, Copyright 2025 Alireza Rezvani

$ grep -rn "megaprompts/" --include="SKILL.md" . | wc -l
❌ 25 dead links to gitignored megaprompts/ directory

$ for f in $(find . -name "SKILL.md" -not -path "./.git/*" -not -path "./.gemini/*" | head -50); do head -10 "$f" | grep -c "^name:" ; done | sort | uniq -c
✅ 50 of 50 sampled SKILL.md files have `name:` frontmatter

$ for f in $(find . -name "SKILL.md" -not -path "./.git/*" -not -path "./.gemini/*" | head -50); do head -10 "$f" | grep -c "^description:" ; done | sort | uniq -c
✅ 50 of 50 sampled SKILL.md files have `description:` frontmatter

$ find . -name "scripts" -type d -not -path "./.git/*" | wc -l
✅ 210 scripts/ directories

$ find . -path "*/references/*.md" -not -path "./.git/*" | wc -l
✅ 656 reference documents

$ ls .github/workflows/
✅ 11 CI workflow files (quality gate, security audit, skill review, VirusTotal, etc.)

Pass rate: 10 of 11. Only failure is the 25 dead megaprompts/ links.

Run B. Functional verification of Python tools

$ python3 marketing-skill/skills/seo-audit/scripts/seo_checker.py --help
✅ usage: seo_checker.py [-h] [--file FILE] [--url URL] [--domain DOMAIN] [--json]

$ python3 marketing-skill/skills/programmatic-seo/scripts/url_pattern_generator.py
✅ PATTERN SCORE: 95/100, generated comparison URLs with warnings

$ python3 research/grants/skills/grants/scripts/fiscal_year_calculator.py --help
✅ usage: fiscal_year_calculator.py [-h] [--reference-date ...] [--window ...] [--output ...]

$ python3 research/grants/skills/grants/scripts/fiscal_year_calculator.py
✅ Output: FY 2026, fiscal_years: [2023, 2024, 2025, 2026]

$ ls .claude-plugin/marketplace.json && python3 -c "import json,sys; d=json.load(open('.claude-plugin/marketplace.json')); print(len(d.get('plugins',[])))"
✅ 59 plugin entries in marketplace.json

$ bash scripts/convert.sh --help 2>&1 | head -5
✅ Usage: ./scripts/convert.sh [--tool <name>] [--out <dir>] [--help]

Functional pass rate: 6 of 6. All tested scripts run with zero external dependencies and produce real output.

What the runs tell you

The structural validation confirms consistent SKILL.md formatting and package layout across 330+ skills. The Python tools are zero-dependency and produce useful output in demo mode. The main weakness is the 25 dead links in research skills and the quality variance between stub skills (41-60 lines) and substantial ones (500+ lines). The CI infrastructure is strong but the test suite is not public.

Setup Walkthrough

1. Clone: git clone --depth 1 https://github.com/alirezarezvani/claude-skills (42 MB shallow).
2. For Claude Code: /plugin marketplace add alirezarezvani/claude-skills then /plugin install @claude-code-skills.
3. For Gemini CLI: ./scripts/gemini-install.sh.
4. For Codex: ./scripts/codex-install.sh --skill .
5. For other platforms: ./scripts/convert.sh --tool .
6. Python tools work with bare python3 (stdlib only, no pip installs needed).

Post-install gotcha: research skills reference megaprompts/ which is not in the repo. These links will 404.

Alternatives

1. anthropics/skills -- Anthropic's official skill pack. Fewer skills but consistent quality and official backing.
2. vercel-labs/agent-skills -- Vendor skill pack for Vercel ecosystem. Deeper in its niche, narrower overall.
3. github/awesome-copilot -- Curated list of Copilot skills and prompts. Community-maintained, no Python tools.