agent-skills
Universal skills, prompts, and markdown resources for AI coding agents — designed to work across providers and tools.
Overview
A centralized, version-controlled collection of reusable agent skills and reference documents shared across AI coding agent platforms. Skills are provider-agnostic by default with optional platform-specific adaptations.
Four ways to use this
| Description | Best for |
|---|
| 🧩 VS Code extension | Browse, scan, install, and semantically search skills inside VS Code | Day-to-day discovery and installation |
| ⚡ TypeScript CLI | npx agent-skills binary that wraps the same install pipeline | CI, scripting, automation |
| 🔌 MCP server | Cross-vendor find_skill tool for Claude Code, Codex, Cursor, Cline, and any MCP host | Letting external agents discover skills semantically |
| 🐚 Bash CLI | Original scripts/install-skill.sh reference implementation | Minimal environments, audit/reference |
All four read and write the same ~/.agent-skills/skills.json manifest. Mix and match freely. See the migration guide for details on switching between them.
Semantic skill discovery (Phase 7)
Once your library grows past a few dozen skills, finding the right one for a task becomes a problem. Phase 7 ships a RAG-powered retrieval system that solves this:
- CLI:
agent-skills find "review my typescript pull request" returns ranked, semantically matching skills in milliseconds
- VS Code extension: open the Find panel from the Installed view's nav bar
- MCP server: any MCP-capable agent (Claude Code, Codex, Cursor, Cline) can call
find_skill directly via stdio
Embeddings run entirely on your machine with all-MiniLM-L6-v2 (22 MB local model) by default. Power users can drop an OPENAI_API_KEY or VOYAGE_API_KEY into the SecretStorage backend for higher-quality cloud embeddings (~$0.01 to index 500 skills).
The local index lives at ~/.agent-skills/index/ and is auto-updated on every install. See the MCP server README for cross-vendor setup snippets.
Supported Providers
Repository Structure
agent-skills/
├── skills/ # Executable agent skills (verb-oriented)
│ ├── code-quality/ # Review, linting, simplification
│ ├── git-workflow/ # Commits, PRs, changelogs
│ ├── testing/ # Unit tests, coverage, test generation
│ ├── documentation/ # Docstrings, API docs
│ └── devops/ # Docker, CI/CD, deployment
├── references/ # Non-executable context documents (noun-oriented)
│ ├── coding-standards/
│ ├── testing-patterns/
│ ├── architecture/
│ └── security/
├── providers/ # Integration guides per platform
├── scripts/ # Validation and index generation
├── skills.json # Machine-readable skill registry
├── CONTRIBUTING.md # How to add skills and contribute
└── CHANGELOG.md # Release history
Skills vs. References
- Skills (
skills/) contain instructions that an agent executes. They are verb-oriented: code-review, lint-fix, generate-tests.
- References (
references/) contain knowledge that an agent reads for context. They are noun-oriented: typescript, rest-api-design.
Quick Start
1. Add to Your Project
Git submodule (recommended):
git submodule add https://github.com/jakesterns/agent-skills.git .agent-skills
cd .agent-skills && git checkout v0.1.0 && cd ..
git add .agent-skills .gitmodules
git commit -m "Add agent-skills submodule at v0.1.0"
Clone to a shared location:
git clone https://github.com/jakesterns/agent-skills.git ~/.agent-skills
cd ~/.agent-skills && git checkout v0.1.0
Raw URL (single file):
https://raw.githubusercontent.com/jakesterns/agent-skills/v0.1.0/skills/code-quality/code-review/skill.md
2. Use a Skill
Each skill lives at skills/<category>/<skill-name>/skill.md. Load the file and follow its instructions. If your provider has a specific adaptation, also load the corresponding file from providers/ inside the skill directory.
Example — code review with Claude Code:
- Load
skills/code-quality/code-review/skill.md (base skill)
- Load
skills/code-quality/code-review/providers/claude-code.md (Claude Code adaptation)
- Apply to your target diff or files
secondsky
dean0x
Jeffallan
Yuan1z0825
Randroids-Dojo
anthropics