How I Built a Multi-Agent Voting System Using NotebookLM and Claude Code

May 04, 2026

∙ Paid

TL;DR: I’ve built a multi-agent system in Claude Code, fed by NotebookLM, that argues every idea before I see it.

Anthropic published a study on April 30, 2026.

They sampled 1 million claude conversations.

6% of them were people asking Claude what to do with their lives.

Should I take the job. Should I leave. Should I move.

Claude was sycophantic in 9% of those conversations.

In relationship questions, that number jumped to 25%.

One in four times, the model agreed with the user without pushing back.

That is what happens when you ask one AI for one answer.

The model wants to help. The model wants to keep you happy.

The model tells you what you want to hear.

I do not want that when I am making decisions.

I want a fight.

So I built one.

Layer 0: The spark is mine

I do not start with AI. I start with walking.

Before any prompt, I leave the desk and brainstorm in my head.

I am the first flame.

The starter.

The rest of the system runs only after I lit something first.

This matters. Most “AI brainstorming” workflows skip this step.

They open Claude. They type “give me 10 ideas.”

The model returns 10 generic ideas. The user picks one.

That is not brainstorming. That is procurement.

My system has me as the seed.

1 spark. 5 filters. Then I see the result.

I start the spark. I type the idea.

Claude asks me 10 questions before it generates anything. After my 10 answers, it produces 16 ideas.

The Champions League bracket runs and four ideas survive.

A team of three Claude Code agents, Pro, Negative, Reporter, argues each surviving idea against a NotebookLM trained on the document that I want.

(Anthropic research and my own Substack performance data.)

The Reporter sends me two finalists and saves two backups.

I read both memos. I pick.

That is the whole system.

Every layer below the spark is automated.

Every layer above the finalists is mine.

Here is how each layer works.