Most AI answers are fine, but where you want to really push the AI to the limits of its thinking powers and reasoning, use this prompt, which forces it to take a multi-faceted approach.

This prompt makes the model answer your question five separate times, from five very different points of view, then argue with itself before giving you a final answer. It is a simplified version of an idea from Andrej Karpathy, a founding member of OpenAI and former director of AI at Tesla.

Here is the prompt, where it came from, what it is good for and where it falls down.

The prompt

Copy this into ChatGPT, Claude or Gemini. Replace the bracketed parts with your own question.

Prompt 01
You are running a Council of Geniuses on the question below.

QUESTION: [your question, problem or decision]

CONTEXT: [constraints, budget, timeline, what you have already tried]

Convene five advisors. Each one reasons in the documented style of a
historical figure, applied to a modern problem. Use their thinking
method only. No archaic language, no impressions, no costume drama.

1. Albert Einstein: thought experiments and first principles. Strips
the problem back to what is actually happening underneath, tests it at
the extremes, distrusts received wisdom.

2. Leonardo da Vinci: observation and cross-disciplinary analogy. Asks
where this problem has already been solved in nature, engineering or
art. Wants to sketch, prototype and test rather than theorise.

3. Marie Curie: evidence and method. Asks what has actually been
measured, how good the data is, and what would have to be true for this
to work. Refuses to move faster than the evidence allows.

4. Ada Lovelace: systems and second-order effects. Looks at the process
behind the answer, what generalises, what can be automated, and what
breaks at scale.

5. Confucius: people, incentives and long-term consequence. Asks who is
affected, what the right thing to do is, and whether this still looks
sensible in five years.

Stage 1. Independent answers. Each advisor answers the question in 150
words or fewer, in their own reasoning style. No advisor sees another
answer yet.

Stage 2. Blind review. Show all five answers back to the council with
the names removed. Each advisor ranks the other four from strongest to
weakest and gives one sentence of criticism for each. Disagree where
there is genuine disagreement.

Stage 3. Debate. Identify the two or three points where the council
actually conflicts. Have the relevant advisors argue those points
directly until one position wins on merit, or state plainly that the
disagreement is unresolved.

Stage 4. Chair's verdict. Acting as chair, produce the final answer:

- the recommendation, stated plainly in the first two sentences
- the reasoning that survived the debate
- what the council disagreed on, and why
- the strongest argument against the recommendation
- the first three things I should do

Do not flatter me. If I have asked the wrong question, say so and
answer the better one.

Where the idea came from

In November 2025, Karpathy published a small project on GitHub called llm-council. Rather than asking your preferred model, you group several of them together. A query goes to each model, the models review and rank each other’s work with the identities hidden, and a designated chairman model writes the final response. He built it as a weekend hack while reading books with AI.

The mechanism matters more than the code. Anonymising the answers before review stops the models favouring their own output, so the criticism is about the ideas rather than the source.

The version above changes one thing. Karpathy’s council runs across different models from different companies, which gives you genuinely different training data and genuinely different blind spots. This one runs inside a single chat with a single model, using personas instead. You lose real model diversity. You gain something you can paste into any chat window in ten seconds without an API key.

Why the personas are historical figures

Any set of contrasting perspectives would work. Karpathy-style councils often use roles like contrarian, executor or outsider, and those are perfectly good.

Historical geniuses have one practical advantage: their thinking methods are documented in enormous detail, and models have been trained on that material heavily. Ask for “a sceptical perspective” and you tend to get generic scepticism. Ask for Curie’s approach to evidence and you get something more specific, because there is a large body of writing describing exactly how she worked.

The five in this prompt were chosen because their methods conflict. Einstein reasons from theory, Curie refuses to move ahead of the data, Da Vinci wants to build something and see, Lovelace thinks in systems, Confucius asks whether the answer is right rather than clever. That conflict is the whole point. Five advisors who agree with each other are just one advisor.

Why it produces better answers

There is no magic here. A few ordinary things are happening at once.

  • It samples more of the answer space. A single pass gives you the most probable answer. Five constrained passes reach ideas the first one skipped.
  • It puts criticism into the context before the final answer. The model writes the objections itself, then has to respond to them. Standard answers rarely contain their own counterargument.
  • It reduces agreeableness. Models tend to go along with the framing you give them. A council that is explicitly told to disagree is harder to please.
  • It plays to what models are genuinely good at. Synthesising well-established knowledge from different fields and holding several voices at once are strengths. Being right first time is not.

It is also worth being accurate about the default. Models are not deliberately made stupid. Fast, short, single-pass answers are cheaper to serve and are what most questions need, so that is the default behaviour. This prompt opts out of it. That costs more compute and more of your time, which is exactly why you should not use it for everything.

What to use it for

The council earns its keep on judgement calls, where there is no single correct answer and the cost of being wrong is real.

  • Business decisions: pricing, hiring, whether to take on a client, whether to kill a project
  • Strategy where you suspect you are too close to the problem
  • Career decisions and job offers
  • Large purchases and financial trade-offs, as a way to surface questions rather than as advice
  • Pressure-testing a plan you have already written
  • Difficult conversations, where the Confucius seat tends to earn its place
  • Working out why something is not working when you have run out of theories

What not to use it for

  • Factual lookups. Five personas will not know something one model does not.
  • Anything where speed matters more than depth.
  • Creative first drafts. Committees flatten writing.
  • Legal, medical or financial questions in place of a qualified professional. Use it to work out what to ask them.

Two honest limitations. It is one model wearing five hats, so a blind spot in the underlying model stays a blind spot in all five seats. And the personas are reasoning styles, not the actual people. Einstein does not know your business, and the council’s confidence is not evidence. For everyday tasks that need less deliberation, try these five reusable prompt structures.

How to get the most from it

Turn on the reasoning or thinking mode if your model has one, then give it far more context than feels necessary. The council is only as good as what you tell it. Numbers, constraints, what you have already tried and what you are afraid of are all worth including.

If the verdict feels thin, ask the chair to reconvene the two advisors who disagreed most and go a round deeper. If it feels too agreeable, add a line telling the council to assume you are wrong.

Finally, treat it as a way of thinking rather than an oracle. The most useful output is often not the recommendation. It is the objection you had not considered.