The LLM Council is a decision making method that sends one question to five AI models, each assigned a different thinking role, then synthesizes their disagreements into a single verdict.
Ask a single model a hard question and it hands you one fluent, agreeable paragraph.
So I stopped asking one AI and started convening five, even for a discussion with my wife about a handbag.
It is called LLM Council: a structured way to run a decision past five distinct AI “advisors,” let them disagree, then synthesize a real answer.
The concept traces to Andrej Karpathy, former Tesla AI director and OpenAI founding member. His open source llm council project fires your question at several models, has them anonymously peer-review each other, then appoints a “chairman” to synthesize the verdict. The version I run needs no special stack: just a codified, role-defined procedure you invoke the same way every time.
The case: should a family buy a $3,500 Louis Vuitton?
Why one AI answer keeps failing you
A single prompt gives you one model’s best guess, stated like a fact. It tells you what you already believed, and you call that confirmation.
That’s how you automate your own confirmation bias.
The five advisors on the council
The fix is to force disagreement on purpose. You assign each AI a fixed thinking style, then make them argue. Five roles, five lenses, one problem:
- The Contrarian. Its only job is to attack the assumptions. “What if the opposite is true?”
- The First Principles Thinker. It gets to the heart of the question. “Why must it be done this way at all?” No industry norms, no conventional wisdom just the core truth underneath.
- The Expansionist. Optimistic and future facing. It generates several answers, stretching the problem into perspectives you hadn't considered.
- The Outsider. It answers as a total newcomer, free of your jargon and your institutional habits. “How would someone unconnected to this solve it?” That’s where the disruptive read comes from.
- The Executor. It doesn’t care about theory. It cares about what you can do this week with the resources you already have.
Run your prompt through all five at once. Read each one. Log exactly where they contradict each other.
The case: a $3,500 bag on trial
Here’s the real prompt I gave the council.
An Australian family, two parents and one child. High monthly income, no consumer debt, 6+ months of emergency savings, on track with long-term investing. They’re weighing a $3,500 Louis Vuitton Neverfull as a daily work bag. There’s also a $7k holiday in the same six-month window. The husband has no strong opinion but wants any spend from the family budget to be equitable. What’s the right call?
On paper, it’s an easy yes. The money is clearly there. But watch what happens when five advisors refuse to give the easy answer.
What each advisor actually said
The Contrarian refused to look at the bag in isolation. “That’s $10,500 of discretionary spend above your run rate, and you’re treating two linked decisions as separate ones.” Then it caught the real signal: the husband says “no opinion” but immediately raises fairness. If she gets to spend $3500 what about him. That grievance will fester if the equity talk never happens.
The First Principles Thinker split the question in two. Can they afford it? Obviously yes. Do they have a shared system for personal spending? No and that’s the actual problem.
The Expansionist flipped the frame entirely. A Neverfull used daily by a client-facing professional isn’t a cost, it’s a signal status, taste, competence. It also holds most of its resale value, while the $7,000 holiday is gone the moment it's spent.
The Outsider called it in one line. $3,500 barely dents a month of this family’s income. The hesitation was never about money. The husband is asking a stranger to validate his wife’s purchase.
The Executor skipped the philosophy. Have one conversation this week. Set a number — say $500/month each, guilt-free, no justification. Write it down. If the bag fits, she buys it. Done.
The chairman moment: where the value lives
You step back into the role of chairman. You look for where the answers overlap, then combine them into one synthesis.
And the overlaps were clear:
- The Contrarian and First Principles Thinker agreed but didn’t say who should raise it. The Expansionist alone caught the resale angle. No one checked whether the wife knew this was even being debated..
The verdict: The financials support it, the use case justifies it, and the resale value even beats the holiday.
My wife’s verdict: All she could recall was expansionist opinion, “it basically free, it provides status, I’ll use it every day and it has resale value”.
How to run your own council this week
Five prompts and the discipline to let them disagree is all it takes.
- Write one sharp, ambitious problem statement. Vague in, vague out.
- Paste it into five fresh chats, each with one role assigned: Contrarian, First Principles, Expansionist, Outsider, Executor.
- Read them sequentially and note the clashes, not the agreements.
- Play chairman. Find the intersection points. Write the synthesis yourself.
Facing a raise negotiation or a career pivot? Run it through the same panel, half AI, half you.
I’ve put the exact LLM Council prompt into a two-page kit you can save and paste. Grab it here, free.
I’ve put the exact llm council prompt into a two-page kit you can save and paste. Grab it here, free:
Download: Quiet Tools 02 Llm Council (PDF)
Thank you for reading. I build with AI and share what learn as I go. Subscribe. Its Free. Subscribe on Substack.