What it does Feed it a GitHub URL or local directory, and it crawls the codebase, identifies core abstractions, and spits out a structured tutorial with visualizations. Think of it as an intern who actually reads the source before writing documentation. It supports filtering by file patterns, size limits, and output languages including Chinese.

The interesting bit The project is built on Pocket Flow, a 100-line LLM framework, and the author used “agentic coding” — designing the flow while letting AI agents write the implementation. The generated tutorials for well-known projects (FastAPI, Celery, DSPy, even LevelDB) are hosted live and were entirely AI-authored.

Key highlights

• Crawls remote repos or local directories; filters with include/exclude globs and max file size
• Supports multiple LLM providers (Gemini, xAI, Ollama) via environment variables
• LLM response caching enabled by default; disable with --no-cache
• Docker support with mounted volumes for input codebases and output tutorials
• Generated tutorials for 20+ popular repos are already published and linked

Caveats

• Requires a capable LLM with “thinking capabilities” (Claude 3.7 with thinking, O1, or Gemini Pro 2.5); weaker models likely produce weaker tutorials
• Default setup assumes Gemini; other providers need manual URL/model/key configuration
• No explicit evaluation of tutorial accuracy is mentioned — the examples look good, but your mileage may vary

Verdict Worth trying if you’re onboarding to an unfamiliar codebase or maintaining internal docs. Skip it if you need guaranteed correctness without human review, or if you’re hoping for deep architectural critique rather than beginner-friendly explanation.