Extension of the NotebookLM expert clone pattern: multiple expert notebooks combined into a cross-grilling advisory council. Experts debate each other's recommendations using the grill-me skill before synthesis. Demonstrates the sub-agent council pattern for strategic decision-making.
The Problem with Generic LLM Councils
Generic personality-prompt councils (contrarian, executor, expansionist, outsider) give generic advice. They have no actual domain knowledge. They cannot give specific advice on cold email outbound strategy because their prompts have no firsthand context on that subject.
The fix: ground each council member in real expert content from NotebookLM notebooks. Each "advisor" becomes an LLM with verifiable domain knowledge from a specific person's actual content.
The Combined Skill Stack
Three skills combined into notebooklm-council:
- grill-me: Each expert challenges and grills the others' recommendations
- LLM council: Multi-expert structure with distinct roles per advisor
- notebooklm: Grounds each expert in real source material
Execution
Basic prompt structure:
Use the notebooklm-council skill. Build me a beautifully formatted HTML output
with synthesis and individual responses from experts.
[Question with full context]Add user context: Create a user.md file describing yourself (age, skills, location, goals, business context). Reference it in the prompt: "Use user.md for context about me and my business." Council responses become personalized.
Specify experts: You can name which notebooks to use. If you don't, the skill browses your notebooks and picks the most relevant.
The Grilling Session
After each expert responds individually, grill-me runs between them:
- Expert A challenges Expert B's plan
- Expert B challenges Expert A's plan
- Cross-examination reveals contradictions and refines recommendations
Then Claude synthesizes the cross-grilled responses into a unified recommendation.
Skipping the grilling: Tell it "don't do the grilling session, give me individual responses." Faster and cheaper; useful when you want independent opinions without the debate.
Output Format Options
- HTML report (shown in video - styled, section-organized)
- Plain chat response
- Any format Claude can produce
Time and Cost
Approximately 3-5 minutes for the full council including grilling session. Multiple NotebookLM queries + Claude synthesis. Cost is primarily in Claude synthesis tokens; NotebookLM queries run on Google's infrastructure.
Practical Fleet Application
Lower priority for the Carlos fleet given Master Brain RAG capabilities. The pattern's value: as a template for decision-support workflows where multiple knowledge sources need to be synthesized with structured debate before a recommendation emerges. The grill-me-between-agents pattern is transferable to other contexts: have multiple research agents challenge each other's findings before Carlos synthesizes for Oliver.