Skill Seekers Scrapes Any Documentation Into an Installable Agent Skill

Your agent is only as good as the knowledge you can hand it. Skill Seekers, at 15k stars, bills itself as the data layer for AI systems: point it at a documentation site, a GitHub repo, a PDF, or a video, and it scrapes, categorizes, and AI-enhances the content into a skill package — a comprehensive SKILL.md plus reference files, ready to install.

Why This Skill Matters

The pipeline is five steps: scrape (checking llms.txt first), categorize into topics, enhance with AI-written examples, package, and optionally upload. Input side: 18 source types, from documentation sites and local codebases through PDFs, Word, EPUB, Jupyter notebooks, OpenAPI specs, RSS feeds, and man pages, to YouTube/Vimeo/local videos and Confluence or Notion wikis. Output side: 22 export targets — the README groups them as 12 LLM platforms (claude, gemini, openai, kimi, and more), 8 RAG and vector formats (LangChain, LlamaIndex, Pinecone, Chroma, and more), plus 2 others.

Prepare once, then export everywhere: the same scraped corpus packages for Claude, a LangChain pipeline, or a Cursor context folder without re-scraping.

Installation

Python 3.10+ and Git are the prerequisites:

pip install skill-seekers

Extras layer on more sources and targets — skill-seekers[mcp] adds the MCP server, skill-seekers[all-llms] every LLM platform, skill-seekers[all] everything. Not sure what you need? The README ships a wizard: skill-seekers-setup.

Real Workflow: Turn Django's Docs Into a Claude Skill

1. Create the skill from the documentation site. Auto-detection figures out the source type (use skill-seekers detect <url> --json to preview it without creating anything):

skill-seekers create https://docs.djangoproject.com/

2. Package the result for your platform — this writes output/django-claude.zip:

skill-seekers package output/django --target claude

3. Install it. With an ANTHROPIC_API_KEY set, skill-seekers package output/react/ --upload or skill-seekers upload output/react.zip ships it for you; without one, upload the zip manually at claude.ai/skills.

Real Workflow: Skill-ify a Local Codebase

Point create at a directory and pick a depth preset — quick (1–2 minutes, surface level), standard (balanced, the default), or comprehensive (deep, exhaustive):

skill-seekers create ./my-project --preset standard

For a project you have not mapped yet, the AI-driven scan reads manifests, README, and Dockerfile/CI, then emits one config per detected framework — react.json, vite.json, tailwind.json — plus a <project>-codebase.json for your own code. You can push new presets back to the community registry from the same flow.

Tips

• Generated skills can install themselves into your agents: skill-seekers install-agent output/react/ --agent cursor targets one agent, --agent all every detected one — Claude Code, Cursor, Cline, Windsurf, and the rest of the supported set.
• The MCP server (40 tools) lets an assistant drive the packaging workflow — the README's example prompt is "Package and upload the React skill"; run the server with python -m skill_seekers.mcp.server_fastmcp.
• On JavaScript-heavy doc sites, the scraper tries sitemap.xml, then llms.txt, then headless browser rendering — and detection is 10× faster when an llms.txt is present.
• Before troubleshooting anything by hand, run skill-seekers doctor — it diagnoses the installation and environment.

When Not to Use This

AI enhancement in API mode consumes tokens and needs provider keys; the LOCAL mode delegates enhancement to coding agents like Claude Code instead, but that still spends your agent's time on large doc sets. And if a skill for your framework already exists in a curated collection, installing the ready-made one beats re-scraping the docs yourself.

See the leaderboard for more skills.