An agent skill that drives the NotebookLM CLI (nlm) to build fully sourced, well-configured notebooks — and optionally Studio artifacts (slides, audio, flashcards, and friends).
The product is not “chat with NotebookLM.” It’s a prompt system: how the skill scopes a topic, writes deep-research queries, interrogates sources, and crafts Studio focus prompts so the notebook is actually usable.
Rebuilt in public from a private prototype. Early commits will look thin on purpose.
Left alone, an agent asks NotebookLM one broad question, accepts the first fluent answer, and leaves you with a mushy notebook and weak artifacts.
- User asks a research question.
- A high-quality LLM scouts the web with Tavily (dynamic search or deep research) and/or Firecrawl.
- That same LLM writes a pointed NotebookLM Deep Research prompt from what it found.
- We run that prompt through NotebookLM (
nlm). Studio artifacts can come later.
Our work is engineering those LLM prompts, not a Python operator framework.
Obsidian / vault export is out of scope. The notebook is the deliverable.
Agents and humans: start at docs/OUTLINE.md. Update it when the plan changes. docs/OLD_FLOW.md is the original skill.
Rebuilding (LLM scout → compose pipeline).
- Living outline + AGENTS entry
-
prompts/scout.mdandprompts/compose-nlm-research.md - Wire Tavily/Firecrawl (after approval)
- Run through real
nlm - Optional Studio
People who want NotebookLM notebooks that are sourced on purpose — especially if they use agent CLIs and care about prompt quality.
MIT (added with the first code commit)