🤔 What are agent skills?
Agent skills are instruction files that teach AI coding agents how to do specific tasks. Think of them like recipes — each SKILL.md file contains step-by-step procedures, best practices, and pitfalls for a particular workflow (debugging, code review, deploying, etc.).
When you give your AI agent access to these files, it can:
- Follow proven workflows instead of improvising
- Avoid common mistakes documented in the skill
- Use the right tools and commands for each task
- Work more consistently across sessions
No code to install, no dependencies. Skills are just markdown files your agent reads. Drop them in a folder, point your agent at it, done.
Any agent that reads markdown instruction files:
- Hermes Agent —
~/.hermes/skills/
- Claude Code —
~/.claude/skills/
- Cursor —
~/.cursor/skills/
- OpenCode, Windsurf, Cline — configure your skills directory
- Any custom agent — just point it at the folder
🎬 Demo
⚡ The 30-second pitch
┌─ obra/superpowers ─────┐
├─ mattpocock/skills ────┤ 191 SKILL.md files
├─ addyosmani/agent-… ──→│ → one-command installer → ~/.hermes/skills/
├─ kepano/obsidian-skills┤ validation CI ~/.claude/skills/
├─ original (Hermes/Nous)┤ REST API (:5555) ~/.cursor/skills/
└─ community + adapted ──┘ attribution metadata
Derivative work, full attribution. Most skill content was authored by other people in their public repos. The value-add here is the aggregation, categorization, conflict detection, validation tooling, attribution metadata, and a portable installer. The named upstream authors did not contribute to or endorse this repo — see NOTICE and per-skill source_url: frontmatter.
[!IMPORTANT]
Want the canonical version of one author's skills? Install directly from their repo. Links in NOTICE.
📦 Quick start
🜲 Hermes Agent
git clone --recurse-submodules https://github.com/kevinnft/ai-agent-skills.git
cd ai-agent-skills && ./scripts/install.sh
🟧 Claude Code
git clone --recurse-submodules https://github.com/kevinnft/ai-agent-skills.git
cd ai-agent-skills && ./scripts/install.sh --agent claude --preset developer
🟦 Cursor
git clone --recurse-submodules https://github.com/kevinnft/ai-agent-skills.git
cd ai-agent-skills && ./scripts/install.sh --agent cursor --preset developer
🟪 Anything else
./scripts/install.sh --agent custom --target /path/to/your/agent/skills --preset minimal
🐍 Windows (no WSL)
git clone --recurse-submodules https://github.com/kevinnft/ai-agent-skills.git
cd ai-agent-skills && python scripts/install.py
Installer flags
./scripts/install.sh --list # browse all categories
./scripts/install.sh --list-presets # browse curated starter packs
./scripts/install.sh --preset developer # 5-category dev bundle (~92 skills)
./scripts/install.sh --preset minimal # just superpowers (14 skills)
./scripts/install.sh --category superpowers # install one category
./scripts/install.sh --dry-run --preset full # preview without copying
./scripts/install.sh --validate # lint frontmatter
./scripts/install.sh --help
Curated starter packs
| Preset | Categories | Approx size |
|---|
developer | superpowers, software-development, addyosmani, mattpocock, github | ~92 |
researcher | research, mlops, data-science, note-taking | ~29 |
content-creator | creative, media, social-media, productivity | ~37 |
devops | devops, github, mcp, autonomous-ai-agents | ~24 |
agentic | superpowers, autonomous-ai-agents, mcp, red-teaming | ~20 |
minimal | superpowers | 14 |
Uninstalling
./scripts/uninstall.sh --preset minimal # remove a preset
./scripts/uninstall.sh --category creative # remove one category
./scripts/uninstall.sh --all --dry-run # preview a full wipe
./scripts/uninstall.sh --help
A timestamped backup is created before each removal unless you pass --no-backup.
Tip: start with one preset. --preset minimal is the smallest meaningful slice. If you like the way the agentic loop feels, scale up with --preset developer or another category. Prefer the installer over raw cp -r so duplicate names and submodules are handled correctly.
Dual-use / red-team skills
A small number of skills document security research or dual-use workflows (notably red-teaming/godmode). They are opt-in:
- Not required for normal coding/research presets (
minimal, developer, researcher, content-creator, devops)
- Included only via
--category red-teaming or --preset agentic
- Review the skill body before enabling, and only use against systems you own or are authorized to test
See SECURITY.md for reporting and dual-use guidance.
🎯 Categories
Counts verified at every release with find skills -name SKILL.md after git submodule update --init.
🛠️ Engineering — 80
🎨 Creative — 22
Diagrams · ASCII art · p5.js · ComfyUI · DESIGN.md tokens · pixel art · Manim · TouchDesigner · Excalidraw
🤖 MLOps — 15
llama.cpp · Axolotl · Unsloth · vLLM · TRL · DSPy · lm-eval-harness · Hugging Face Hub
🔬 Research — 11
arXiv · web scraping · blogwatcher · polymarket · NFT/crypto analysis · trending repos
🐙 GitHub — 10
PR workflow · code review · repo mgmt · auth · issues · visual assets
⚙️ DevOps — 9
Docker compose · VPS hardening · browser automation · kanban · webhooks
📋 Productivity — 8
Notion · Google Workspace · Linear · Airtable · PowerPoint · OCR · maps
Plus media (5) · obsidian-skills (5, submodule) · apple (4) · autonomous-ai-agents (4) · note-taking, social-media, gaming, software-copyright (2 each) · data-science, dogfood, email, mcp, red-teaming, smart-home, yuanbao, patent-disclosure-skill (1 each)
📚 Full per-skill catalog → docs/categories.md
📊 Provenance
Every SKILL.md declares where it came from. Run grep -h "^origin:" skills/**/SKILL.md | sort | uniq -c to verify.
┌─────────────────────────────────────────────────────────────┐
89 │ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ │ aggregated
63 │ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ │ original
22 │ ▓▓▓▓▓▓▓▓▓ │ unknown ⚠
11 │ ▓▓▓▓ │ adapted
└─────────────────────────────────────────────────────────────┘
191 total*
*191 SKILL.md files total. 6 are inside the obsidian-skills / patent-disclosure-skill git submodules — those keep upstream-owned frontmatter and are excluded from the origin: provenance contract.
origin: | Count | Meaning |
|---|
🟦 aggregated | 89 | Top-level frontmatter copied from a named upstream repo with full attribution (addyosmani/, mattpocock/, superpowers/, NousResearch/hermes-agent, JimLiu/baoyu-skills) |
🟩 original | 63 | Authored for this repo's ecosystem (Hermes / Nous Research / Orchestra / community) |
🟧 adapted | 11 | Materially edited from a named upstream — see source_url |
🟥 unknown | 22 | Author present in frontmatter but upstream repo not yet identified — help wanted! |
🤝 Help wanted on the remaining unknown-origin skills. Down from 53 in v1.6.0 — community-traceable upstream identified for 31 skills (25 from NousResearch/hermes-agent, 6 from named author repos). If you recognize one of the remaining 22, open a PR adding source_repo / source_url / source_license to the frontmatter.
⚠️ Conflicts & duplicates
Same-name skills exist under multiple categories. Pick one when installing — duplicates confuse the loader (last-write-wins).
| Skill | Locations | Recommended pick |
|---|
test-driven-development | addyosmani/, superpowers/, software-development/ | 🥇 superpowers/ — most explicit RED-GREEN-REFACTOR loop |
systematic-debugging | superpowers/, software-development/ | 🥇 superpowers/ — 4-phase |
requesting-code-review | superpowers/, software-development/ | 🥇 superpowers/ |
subagent-driven-development | superpowers/, software-development/ | 🥇 superpowers/ |
writing-plans | superpowers/, software-development/ | 🥇 superpowers/ |
The installer auto-resolves these: by default it keeps the superpowers/ copy and skips duplicates from other categories. Override with --prefer <category> if you want a different pick. See docs/sync.md for the upstream-drift strategy.
🛠️ Tooling (original to this repo)
| Script | Purpose | Notes |
|---|
install.sh | Copy skills to your agent's skills dir (Linux/macOS) | --preset, --dry-run, --validate-only, --prefer, --agent |
install.py | Cross-platform installer (Windows/macOS/Linux) | Same flags as install.sh, Python 3.7+ stdlib only |
validate.sh | Lint frontmatter + broken links | Non-zero exit on failure |
update.sh | Pull latest, back up existing, apply | Safe re-run |
release.py | Generate changelog + tag semver | Tested ✅ |
quality-dashboard.py | Self-scored repo health metric | Heuristic — not third-party validation |
skills-api.py | Local REST server (:5555) | GET /api/skills, GET /api/skills/search?q=... |
generate_site.py | Generate searchable HTML catalog | Outputs docs/index.html |
.github/workflows/ci.yml | Validation + pytest on push/PR | 28 unit tests (ubuntu + macos) |
|
All tooling MIT-licensed under LICENSE.
📈 Stats at a glance
191 SKILL.md files ✅ verified each release
28 populated categories ✅
185 non-submodule skills with origin: 89 aggregated · 63 original · 11 adapted · 22 unknown
22 skills awaiting upstream identification 🤝 help wanted
11 scripts under scripts/
4 GitHub Actions workflows CI · auto-label · upstream-sync-check · pages
28 pytest unit tests scripts/release · skills-api · quality-dashboard
2 git submodules obsidian-skills · patent-disclosure-skill
🤝 Contributing
PRs welcome — see CONTRIBUTING.md. Two contribution types with different rules:
| Type | Required frontmatter | What we check |
|---|
| 🟢 Original skill | name, description, author, tags | Tested in at least one real session — describe how in the PR |
| 🟡 Aggregated skill | source_repo, source_url, source_license (in addition to the above) | Upstream license permits redistribution; you didn't strip credit |
We reject PRs that strip attribution or misrepresent upstream authorship. Reviewers see the attribution checklist on every PR.
❓ FAQ
No. They have not contributed to or endorsed this repo. See NOTICE for the non-affiliation disclaimer.
You can — and should, if you only want one author's skills. The win here is one command for 191 skills across 28 categories, plus categorization, conflict detection, validation, and a REST API. We diff against upstream weekly (see docs/sync.md) and surface drift via upstream-sync-check.yml.
It's a heuristic computed by scripts/quality-dashboard.py (this repo's own script) based on README presence, LICENSE, CI status, broken-link check, and a few other surface signals. Not third-party validation. Don't read it as endorsement.
191 as of v1.7.0 (verified each release via find skills -name SKILL.md). v1.6.x was 189; v1.7.0 promoted domain/domain-intel and inference-sh/inference-sh from descriptor-only stubs to real skills. Open a PR if you find a stale figure.
A weekly GitHub Action (Mondays 06:17 UTC) opens or updates a tracking issue with the upstream HEAD SHA of each source repo. If you see drift you care about, open a sync PR — full strategy in docs/sync.md.
This repo's tooling: yes (MIT). Skill content from upstream: yes if their license permits — see per-skill source_license frontmatter or NOTICE.
Yes — the searchable catalog is deployed via GitHub Pages. Filter by category, search by name/tag, and preview any skill before installing.
Yes. Use python scripts/install.py — it's a cross-platform installer with the same flags as install.sh, requires only Python 3.7+ stdlib (no pip dependencies).
Raw copy works for a quick try, but the installer is safer: it resolves same-name duplicates via --prefer, warns on empty git submodules, supports --dry-run / --validate-only, and maps --agent claude|cursor|hermes to the right target path.
Yes — a small set (for example red-teaming/godmode) documents security-research workflows. They are not in the default coding presets; install only if you need them and only for authorized testing. See SECURITY.md.
📄 License & attribution
- This repo's tooling, layout, aggregation work: MIT —
LICENSE
- Skill content from upstream repos: each skill keeps its upstream license — see per-skill frontmatter and
NOTICE
If you are an upstream author and would like a skill removed, renamed, or attributed differently, please open an issue.
Made for the agent ecosystem · powered by community work · attribution-first.
Hermes Agent · docs · repo
NOTICE · CHANGELOG · SECURITY · Sync strategy · Issues
