PleasePrompto / notebooklm-skill

Install is a git clone into ~/.claude/skills; the first real use creates its own Python virtual environment and installs actual Google Chrome, not Chromium, then opens it once for you to sign in to Google. After that, asking a question opens a notebook from a small local library file, asks NotebookLM in the browser, reads back the citation-backed answer, and closes the browser — each question is a fresh, stateless browser session, though the skill prompts itself to ask follow-up questions in the same run rather than stopping at the first answer.

This is for someone running Claude Code locally who wants answers grounded in a specific document set — a manual, a spec, a pile of PDFs — without Claude re-reading the same files on every question. It will not run in the Claude web UI at all, because that sandbox has no network access for the browser automation to reach. And any notebook you want it to query has to be shared as "Anyone with the link," so treat that link, and whatever you upload to it, as no longer private to your Google account alone.

READ THE FULL ANALYSIS＋

How it gets past Google. The skill's own documentation says it uses "human-like typing speeds and interaction patterns" and calls this "stealth" — it is automating your real, logged-in Google session to look like a person clicking around NotebookLM, not calling an API. That is the price of there being no NotebookLM API to call.

A sibling worth knowing about. The same author maintains notebooklm-mcp, a TypeScript MCP server version with persistent chat sessions and support for Codex, Cursor, and other MCP clients. This skill is the simpler, Python, Claude-Code-only alternative — no persistent session, but nothing beyond a git clone to set up.