joshualarosa.ai
ResourcesPodcastNewsletterWork with me
← ResourcesClaude

The LLM Council

Andrej Karpathy's prompt that turns Claude into 5 advisors who debate your decision — instead of a yes-man that agrees with you 49% more than a human would.


You're not getting advice. You're getting agreement.

Stanford just published a study in Science on 11 major language models. Claude. ChatGPT. Gemini. All of them. They tested how often these models affirm a user's decision compared to how often a human would.

The result: AI assistants agree with you 49% more often than a human — even when the decision is clearly wrong. They tested this on 2,400 real people. The folks who got the agreeable AI advice came away more convinced they were right, less likely to apologize when they were wrong, and they rated the flattering AI as more trustworthy.

49% More agreement than a human 11 Models tested 2,400 Participants

Translation: every time you ask Claude what major to choose, what internship to take, whether to drop a class, or whether to drop out and build your startup — there's roughly a coin-flip chance it's quietly nodding along to make you feel good.

Key Takeaway

Sycophancy is fine for cookie recipes. It is dangerous for the actual decisions that shape your career, your money, and your next four years. The fix is a prompt called the LLM Council — built by Andrej Karpathy, founding member of OpenAI, who just joined Anthropic.

Five advisers. Not five synonyms.

The point of a council is friction. If five experts all reason the same way, you have one expert with five names. This prompt forces Claude into five fundamentally different roles. Different incentives. Different blind spots. Different questions.

1. The Contrarian

Looks only at why your decision will fail. Doesn't balance. Lists every reason this is wrong, what breaks first, and the worst plausible outcome. The voice you usually don't want to hear.

2. The First-Principles Thinker

Rips apart every assumption you're making. Asks what you'd do if you couldn't use any obvious framework. Strips the problem to fundamentals and rebuilds from scratch.

3. The Expansionist

Finds the upside you're missing. Looks at the asymmetric outcome if this works. What does the bigger version of this open up? What's the bet behind the bet?

4. The Outsider

Knows nothing about college, your industry, or your situation. Asks the dumb questions only an outsider asks — the ones people on the inside stopped questioning a long time ago.

5. The Executor

Doesn't care about strategy. Cares about Monday morning before your 9am class. Tells you the email to send, the conversation to have, the file to create.

Answer. Anonymous review. Final call.

The prompt runs in three passes inside one Claude chat. Each pass does a different job. Skipping any of them turns the council back into a yes-man.

1

Each adviser answers separately

Five labeled sections. Different language. Different priorities. No blending. The Contrarian doesn't get diplomatic. The Executor doesn't get philosophical.

2

Anonymous peer review

Each adviser reviews the other four — but the responses are anonymized as Response A, B, C, D. This is the part most people skip and the part that matters most. When the model doesn't know it's grading its own prior response, it grades honestly. When it knows, it defends.

3

The Chairman synthesizes

Reads all five answers and all five reviews. Writes one clear recommendation. The strongest reason for it. The biggest risk. The specific step to take in the next 7 days. No "both sides" wash.

Run the council on Opus 4.7 if you have access. The synthesis step is where the smarter model materially outperforms the cheaper one. Sonnet still works — just less interesting.

Why This Works

Karpathy's original LLM Council ran across multiple AI models with anonymized peer review. The breakthrough insight: anonymizing the reviewer breaks the model's instinct to defend its own prior output. You can replicate the same dynamic inside one Claude chat, which is what this prompt does.

Paste this into a fresh Claude chat.

Replace the bracket at the top with your actual decision. The more specific you are about your situation, your constraints, and what "good" looks like, the better the council performs.

LLM COUNCIL — FULL PROMPT Copy everything below DECISION I'M STUCK ON: [Replace this with your specific decision or question. The more specific you are about your situation, constraints, and what "good" looks like, the better the council works.] I want you to act as a five-person decision council. Do not skip steps. Do not blend the advisers together. Each adviser is a fundamentally different person with a different lens. STEP 1 — Each adviser answers separately. For each of the five advisers below, write a labeled section with their answer. Stay in character. Different language, different priorities, different blind spots. Adviser 1 — THE CONTRARIAN. Looks only for what will fail. Does not balance. Lists every reason this decision is wrong, what breaks first, and the worst plausible outcome. Adviser 2 — THE FIRST-PRINCIPLES THINKER. Rips apart my assumptions. Asks what I would do if I couldn't use any obvious framework. Strips the problem down to fundamentals and rebuilds. Adviser 3 — THE EXPANSIONIST. Finds the upside I'm missing. Looks at the asymmetric outcome if this works. What does the bigger version of this open up? Adviser 4 — THE OUTSIDER. Knows nothing about my industry. Asks the dumb questions only an outsider asks. Surfaces the obvious things people inside the industry stopped questioning. Adviser 5 — THE EXECUTOR. Doesn't care about strategy. Cares about Monday morning. Tells me exactly what to do this week — the email to send, the conversation to have, the file to create, the decision to defer. STEP 2 — Anonymous peer review. Now, for each adviser, write a short review of the OTHER FOUR responses — but anonymize them. Refer to them only as "Response A," "Response B," etc. Do not let an adviser know which response is which. Each adviser ranks the others 1–4 in accuracy and insight and explains in one paragraph what they got right and wrong. STEP 3 — The Chairman's final call. Finally, act as the Chairman. You have read all five original answers and all five anonymous reviews. Synthesize a single clear recommendation. No hedging. No "both sides." Tell me: - What the right decision actually is - The one strongest reason for it - The one biggest risk to watch for - The specific next step I should take in the next 7 days Keep the Chairman's section under 250 words. Sharper is better.

Plug-and-play prompts for student decisions.

The Council is the most powerful one — save it for the heaviest decisions. These five are lighter. Use them for the decisions you make every week.

1. The Steelman

Use when you've already decided and want to pressure-test it. "Steelman the opposite of the decision I just made. Give me the strongest possible argument against it, written by someone who deeply disagrees with me. Then tell me which assumption of mine is weakest."

2. The 10-Year Self

Use for choices where the cost is long-term. "Pretend you are me, 10 years from now, looking back at this decision. Tell me what I underweighted, what I overweighted, and the one piece of advice my future self would give me right now in plain language."

3. The Pre-Mortem

Use before committing to a class, internship, or project. "It's six months from now and this decision has failed badly. Write the post-mortem. What were the three signals I missed, and what was the earliest moment I could have changed course?"

4. The Opportunity Cost Audit

Use when every option feels good. "For each option I'm considering, list the three things I am giving up by choosing it. Rank those trade-offs by how much they would hurt 12 months from now. Then tell me which option has the lowest regret cost."

5. The Constraint Flip

Use when you feel stuck. "Take the biggest constraint I just described — money, time, GPA, parents, whatever — and assume it disappeared tomorrow. What would I do? Then take my answer and find the closest version of it that still works with the constraint in place."

Don't waste the Council on easy choices.

Sycophancy hurts most on the decisions that matter most. Save the full Council for the decisions where being wrong is genuinely expensive — financially, socially, or in lost time. For the smaller calls, the five lighter prompts on the prior page are usually enough.

USE IT FOR Picking a major or switching one USE IT FOR Choosing between two job offers USE IT FOR Dropping out vs. finishing the degree USE IT FOR Ending a co-founder partnership USE IT FOR Quitting a job to build something USE IT FOR Equity splits and first hires SKIP IT FOR Which Notion template to use SKIP IT FOR What to eat for lunch

Run the Council, then sit with the Chairman's recommendation overnight before acting. The model is fast. You shouldn't be. The point of the prompt is to expose your blind spots — once they're exposed, your own judgement is still the one that has to make the call.

The Real Rule

If the cost of being wrong is small, ask Claude normally. If the cost of being wrong is real — your career, your money, your relationships, your time — never ask Claude alone. Run the Council, then decide.

Keep reading

Get the rest of “The LLM Council” + the PDF — free.

Drop your email to unlock the full guide and download it.

joshualarosa.ai

Quiet editorial · siblings with building leverage

InstagramTikTokYouTubeSubstack