What it does

This is a Claude Code skill that automates Google NotebookLM through a local browser. You upload documents to NotebookLM, share the notebook publicly, then ask Claude questions about them. Claude fires up a headless Chrome instance, queries NotebookLM, and gets back answers that are strictly grounded in your sources — with citations — rather than inventing APIs or specs from its weights.

The interesting bit

The whole thing is basically sanctioned prompt injection. The skill scripts append “Is that ALL you need to know?” to every NotebookLM answer, nudging Claude to ask follow-up questions automatically. It’s a hack to compensate for the fact that each query opens a fresh browser with zero session memory. The README is admirably frank about this being stateless glue code, not a protocol integration.

Key highlights

• Uses Patchright (stealth-Playwright) with “human-like typing” to avoid Google bot detection
• Self-contained: drops a .venv and Chrome setup inside ~/.claude/skills/notebooklm/ on first run
• Maintains a local library.json of notebooks with tags so Claude can pick the right one for a query
• One-time Google auth via popped Chrome window; cookies persist in local browser state
• Explicitly not an MCP server — simpler, but only works with local Claude Code (the web UI sandbox blocks network access)

Caveats

• Notebooks must be shared publicly; there is no private API access
• Each question spins up a fresh browser, so no conversational context within a NotebookLM session
• Free tier rate limits apply, and you’re at the mercy of Google’s UI not changing

Verdict

Worth a look if you already live in Claude Code and want citation-backed answers from your own docs without building a RAG pipeline. Skip it if you need private notebooks, persistent multi-turn NotebookLM chats, or anything beyond Google’s free tier limits.