How pleaseprompto/notebooklm-skill works
Project type and purpose
• This is an open-source “Claude Code Skill” plugin written in Python.
• Purpose: let Claude Code (local CLI AI agent) query Google NotebookLM via browser automation (Patchright/Playwright) for source-grounded answers based on user-uploaded docs.
File structure
• Top level:
• SKILL.md – Claude’s instructions/manifest for the skill
• scripts/ – all Python automation modules
• requirements.txt – Python dependencies (patchright, python-dotenv)
• AUTHENTICATION.md, SKILL.md, references/ – docs
• data/ (created at runtime) – stores library.json, auth_info.json, browser_state/
• scripts directory:
• run.py – entry point invoked by Claude Code
• auth_manager.py – handles one-time Google login via Chrome
• notebook_manager.py – manages “library” of NotebookLM links and metadata
• ask_question.py – sends queries to NotebookLM, extracts answer+citation
• cleanup_manager.py – clears cookies/local state when needed
• setup_environment.py – bootstraps .venv, installs Chrome, dependencies
• browser_session.py – low-level browser launch and context management
README for architecture insights
• The “Architecture” section already outlines a 3-piece flow:
1. Claude loads the skill → 2. Python script runs → 3. Browser automation interacts with NotebookLM → returns answers
• Data storage is local and stateless per question (fresh browser session per query), but the library persists.
Elements
– Blue for user-facing components (Claude Code CLI, SKILL.md)
– Green for local code (scripts folder, .venv)
– Orange for external dependencies (Chrome, NotebookLM service)
– Gray for data storage (library.json, auth_info.json)
• Notation for “stateless session” on the browser component and “persistent storage” on data store component