NotebookLM + Claude Code as an expert knowledge cloning and querying system. Clone any expert from public content. Use parallel sub-agents to query multiple expert notebooks simultaneously. Token economics: NotebookLM uses Google's tokens; Claude only synthesizes.

Watch on YouTube

The Concept

NotebookLM processes your source documents on Google's infrastructure. When you query an expert notebook, Google's tokens do the retrieval. Claude uses tokens only for the final synthesis step. For large corpora, this significantly reduces Claude spend vs loading all documents into Claude's context window.

Setting Up an Expert Clone

1. Create account at notebooklm.google.com (use Google account)
2. Create a new notebook for each expert
3. Add sources: YouTube video URLs, uploaded files, copy-pasted text, Google Drive links
4. More sources = more accurate, less hallucination (model is grounded in source material)

Best formats: plain text, markdown, YouTube videos. PDFs with images have inconsistent results. YouTube videos work exceptionally well.

Persona prompt pattern: "Answer as [expert name]. Give specific advice based only on what [expert] has said in the source material." The model adopts their communication style and specificity.

Note: Personal and business Google accounts cannot share notebooks. Keep everything within one workspace if collaboration is needed.

Integration with Claude Code

No native NotebookLM integration exists. Use the open-source notebooklm-py library (11,000+ GitHub stars) as a bridge.

Setup:

1. Give install instructions to Claude Code; it installs the library
2. Run notebooklm login - Playwright browser window opens for Google auth
3. Complete auth - cookies cached for subsequent requests
4. Install the notebooklm Claude Code skill

Usage:

"Use the notebooklm skill to query which notebooks I have."
"Query Matt and ask how he would do link building in 2026."

Parallel Expert Council Pattern

"Use parallel agents and send out three sub-agents that will simultaneously
query [Expert 1], [Expert 2], and [Expert 3]. Ask what are the best 
[topic] methods in 2026."

Claude spawns three sub-agents. Each queries a different notebook. All return results simultaneously. Main agent synthesizes.

Result: cross-expert consensus in 10-15 seconds, instead of asking each one serially. Can query 5-10 experts simultaneously for complex strategic questions.

Token Economics

• NotebookLM does the heavy retrieval work on Google's infrastructure
• Claude uses tokens only for synthesis
• For a corpus of 50 documents, this saves significant context spend vs loading all 50 into Claude on every query

This validates the Master Brain RAG approach: chunked expert knowledge + query-time retrieval + synthesis is more efficient than context-window stuffing.

Fleet Note

The Master Brain RAG system (gmgxxiqgshbbgzhqzngq) already provides more sophisticated functionality for the Carlos fleet. This video is useful as architectural validation: external practitioner independently reaching the same conclusion about chunked retrieval vs context stuffing.