The Council // Day 20
Your personal LLM Council, inside one Claude chat
Stanford found Claude agrees with you about 49% more than a human would. So when you ask it to pressure-test a real decision, you mostly get your own opinion back in better words. That's sycophancy, and it's quietly wrecking the calls you trust it with.
Andrej Karpathy's fix was an LLM Council that runs across multiple models. This is that idea rebuilt to run inside a single Claude chat: five advisors who answer separately, review each other blind, and a chairman who rules. The upgrade I added is in Step 4, where the chairman names the one place they clashed hardest. That clash is your actual blind spot.
// How to use it: paste the whole block into a fresh Claude chat, then type your decision and context where it says so. Run it before any call that matters.
You are going to act as my personal LLM Council for the decision I give you. Run this in five steps and do not skip any:
STEP 1: Five independent advisors.
Answer my question five separate times, once as each advisor below. Keep them genuinely distinct, with different priorities and different conclusions. Don't converge.
1) The Contrarian: only argues what will go wrong and why this fails.
2) The First-Principles Thinker: ignores convention, rebuilds the answer from base truths, and names the assumptions I'm making.
3) The Expansionist: finds the upside, the bigger version, the option I'm not seeing.
4) The Outsider: knows nothing about my industry, reasons from common sense, and asks the "dumb" questions that expose blind spots.
5) The Executor: ignores theory, only cares what I actually do Monday morning.
STEP 2: Blind peer review.
Now have each advisor review the other four answers WITHOUT knowing who wrote which (label them Answer A to E, shuffled). Each flags the strongest point and the weakest point across the others.
STEP 3: The Chairman.
A chairman reads everything and gives me the final recommendation, the reasoning, and 3 clear next steps.
STEP 4: The clash (most important).
The chairman explicitly flags the ONE place the advisors disagreed the hardest, and explains why that clash is likely my real blind spot. The thing I haven't resolved.
STEP 5: Ask me for anything missing.
If you need more context to do this well, ask me up to 3 questions first.
My decision / question:
[ paste your decision + context here ]
A yes-man tells you what you want to hear. A council tells you what you need to hear. The signal was never the agreement. It's the clash.
Run this before your next big decision and watch where the room splits. That split is where your real work is.
Follow @mikemeansbusiness_ai for the next system.