A fork of huggingface/skills — a collection of AI skills, agents, and evaluations.

This repository contains:

• Skills: Modular AI capabilities that can be composed into agents
• Agents: Autonomous AI systems built from skills
• Evals: Benchmarks and leaderboards for measuring skill performance
• Marketplace: Discoverable plugins for Claude and Cursor

.
├── .claude-plugin/          # Claude AI plugin configuration
│   ├── plugin.json          # Plugin metadata and entry points
│   └── marketplace.json     # Marketplace listing
├── .cursor-plugin/          # Cursor IDE plugin configuration
│   ├── plugin.json          # Plugin metadata
│   └── marketplace.json     # Marketplace listing
├── .github/
│   └── workflows/
│       ├── generate-agents.yml          # CI: auto-generate agent configs
│       ├── push-evals-leaderboard.yml   # CI: update evals leaderboard
│       └── push-hackers-leaderboard.yml # CI: update hackers leaderboard
└── skills/                  # Core skill implementations

• Python 3.10+
• pip or uv for package management

git clone https://github.com/your-org/skills.git
cd skills
pip install -e .

python -m skills.evals run --skill <skill-name>

Install via the Claude marketplace or load the .claude-plugin/plugin.json directly in your Claude environment.

Install via the Cursor marketplace or load the .cursor-plugin/plugin.json directly in your Cursor IDE.

1. Fork the repository
2. Create a feature branch (git checkout -b feat/my-skill)
3. Add your skill under skills/
4. Add corresponding evals under evals/
5. Open a pull request

See CONTRIBUTING.md for detailed guidelines.

See .github/workflows/SECURITY.md for our security policy.

Apache 2.0 — see LICENSE for details.

Personal fork notes: I'm using this repo to experiment with building custom skills for my own workflows. Main areas of interest: text summarization and code review skills. Not intended for production use.

TODO:

• Build a summarization skill that handles long documents (>10k tokens) by chunking
• Experiment with a code review skill focused on Python style/type hints
• Compare eval results against upstream once I have a baseline
• Look into adding a --dry-run flag to the evals runner so I can test without writing results
• Try running evals against gpt-4o-mini as a cheaper baseline before committing to full runs
• Try chunking strategy: overlap by ~10% between chunks to avoid losing context at boundaries — tested this, works well; settled on 15% overlap since 10% occasionally dropped a sentence at boundaries
• Set up local dev environment with uv instead of pip — noticeably faster for resolving deps