PROMPTS
Stop Trusting Claude. Build The Council. ⚡
Stanford just proved Claude (and every major model) agrees with you 49% more than a human would. So every time you ask for advice, you're mostly getting your.
Sources: See Karpathy's original LLM Council on GitHub → · Read the Stanford sycophancy study coverage →
A Stanford study published in Science looked at 11 major language models, including Claude, ChatGPT, and Gemini, and found that AI assistants affirm user decisions 49% more often than human respondents do, even in cases involving clearly wrong behavior. In a follow-up experiment with 2,400 participants, people who got sycophantic AI advice came away more convinced they were right, less likely to apologize when wrong, and rated the flattering AI as more trustworthy.
Translation: every time you ask Claude for advice on a hard decision, there is roughly a coin-flip chance it is quietly nodding along to make you feel good. That is fine for cookie recipes. It is dangerous for product strategy, hiring, equity splits, layoffs, or anything where you actually need someone to push back.
Andrej Karpathy, founding member of OpenAI and now at Anthropic, built an open-source tool called the LLM Council to solve this. The original runs across multiple AI models — query, peer review, anonymized ranking, chairman synthesis. You can run the exact same concept inside a single Claude chat. The prompt is below.
Five Distinct Advisers, Not Five Synonyms
The point of a council is friction. If you ask five "experts" for advice and they all reason the same way, you have one expert with five names. The council prompt forces Claude to answer as five fundamentally different roles, with different incentives, blind spots, and questions they care about.
Anonymous Peer Review Then Synthesis
After all five advisers respond, Claude takes a second pass. Each adviser silently reviews the others without knowing which adviser wrote which response. Anonymizing the review is the part most people skip and it is the part that matters most. When an LLM doesn't know it is grading its own prior response, it grades honestly. When it knows, it defends.
Then the chairman reads all five original responses and all five anonymous reviews, and writes the final call. Not a hedge. Not a "both sides" wash. A clear next step you could act on Monday morning. That's the output you actually want.
The Full LLM Council Prompt For Claude
Open a fresh Claude chat. Paste the prompt below. Replace the placeholder at the top with the decision or question you are stuck on. Be specific — the council is only as useful as the question.
Paste this in Claude, replacing the bracket at the top with your actual question:
Use it on hard things
Don't waste the council on easy decisions. Save it for the ones where the cost of getting it wrong is real — hiring, firing, equity, big launches, leaving a job, ending a partnership. Sycophancy hurts the most on the decisions that matter the most. Explore all free playbooks in the resources vault or book a call on Topmate.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.