site2skill

Turn any documentation website into a Claude Agent Skill.

site2skill is a tool that scrapes a documentation website, converts it to Markdown, and packages it as a Claude Agent Skill (ZIP format) with a proper SKILL.md entry point.

Agent Skills are dynamically loaded knowledge modules that Claude uses on demand. They work across Claude Code, Claude apps, and the API.

Installation

Install from PyPI (stable)

pip install site2skill

This installs the stable Python version (0.1.x), which requires Python 3.10+ and wget.

Try the Rust beta

A rewrite in Rust is available as a beta. It's ~4x faster, has no wget dependency, and ships as a single binary.

pip install --pre site2skill

Or pin the version:

pip install 'site2skill==0.2.0b1'

Verify you're running the Rust version:

site2skill --version
# site2skill 0.2.0-beta.1

The stable version has no --version flag, so if you see a version string, you're on the Rust beta.

To go back to stable:

pip install 'site2skill<0.2'

Run without Installation

# Beta (Rust)
uvx --pre site2skill <URL> <SKILL_NAME>

# Stable (Python, from legacy branch)
uvx --from 'git+https://github.com/laiso/site2skill@legacy' site2skill <URL> <SKILL_NAME>

Build from Source

git clone https://github.com/laiso/site2skill.git
cd site2skill
cargo build --release
./target/release/site2skill --help

Usage

# Basic usage
site2skill <URL> <SKILL_NAME>

# Example: Create a skill for PAY.JP
site2skill https://docs.pay.jp/v1/ payjp

# Example: Create a skill for uv documentation
site2skill https://docs.astral.sh/uv/ uv-docs

# Target specific agent (sets default output directory)
site2skill <URL> <SKILL_NAME> --target claude-desktop

CLI Options

site2skill <URL> <SKILL_NAME> [options]

Options:
  --target           Target agent (claude|claude-desktop|cursor|gemini|codex). Sets default output directory
  --output, -o       Base output directory for skill structure (overrides target default)
  --skill-output     Output directory for .skill file (default: .)
  --temp-dir         Temporary directory for processing (default: build)
  --max-concurrency  Maximum number of HTTP requests to run at once (default: 10)
  --skip-fetch       Skip the download step (use existing files in temp dir)
  --clean            Clean up temporary directory after completion
  --version          Print version (Rust beta only)

How it works

1. Fetch: Crawls the documentation site with a built-in async HTTP crawler (robots.txt compliant, concurrent).
2. Convert: Converts HTML pages to Markdown using scraper and htmd.
3. Normalize: Resolves relative links to absolute URLs.
4. Validate: Checks the skill structure and size limits.
5. Package: Generates SKILL.md and zips everything into a .skill file.

Output

The tool generates a skill directory in .claude/skills/<skill_name>/ containing:

<skill_name>/
├── SKILL.md           # Entry point with usage instructions
├── references/        # Markdown documentation files (preferred)
└── scripts/
    └── search_docs.py # Search tool for documentation

Additionally, a <skill_name>.skill file (ZIP archive) is created in the current directory when targeting claude-desktop.

Legacy note: older skills may use docs/ instead of references/. The search tool and validator now support both, preferring references/ when present.

Search Tool

Each generated skill includes a Python search script (requires Python 3 at runtime):

python scripts/search_docs.py "<query>"
python scripts/search_docs.py "<query>" --json --max-results 5

License

MIT