LLM Council
Five AI advisors. One verdict.
Ask one AI a question and you get one answer. It might be great. It might be mid. You have no way to tell, because you only saw one opinion. The LLM Council runs your question past five advisors who think in completely different ways, makes them tear each other's answers apart, then hands you a single verdict.
Watch the Short
The 60-second version.
Saw the Short and want the actual setup? It's below. Install takes about two minutes and you never touch a terminal — you paste one prompt and Claude does the rest.
Why this matters
One answer feels like the answer.
AI is confident by default. When it hands you a single response, there's nothing to compare it to — you can't see what it ignored, what it assumed, or which angle it never considered. Five advisors with conflicting jobs make the gaps visible, and the peer-review round is what separates this from just asking the same question five times.
One AI, one answer
1 opinion
You get a confident-sounding response with no way to judge it. No pushback, no second angle, no idea what it skipped.
The council
5 + peer review
Five conflicting angles, each one tearing into the others, then a chairman verdict telling you where they agree and what to actually do.
Same question. A verdict instead of a guess.
Meet the council
Five thinking styles, built to disagree.
These aren't job titles or personas. They're angles that naturally create tension: Contrarian versus Expansionist (downside versus upside), First Principles versus Executor (rethink it versus just ship it), with the Outsider in the middle keeping everyone honest.
The Contrarian
Finds what's wrong
Assumes your idea has a fatal flaw and goes hunting for it. Not a pessimist — the friend who asks the question you've been avoiding.
The First Principles Thinker
Asks what you're really solving
Ignores the surface question and rebuilds the problem from scratch. Sometimes the most valuable output is "you're asking the wrong question."
The Expansionist
Hunts the upside
Doesn't care about risk — that's the Contrarian's job. Looks for the bigger version, the adjacent opportunity, the thing you're undervaluing.
The Outsider
Has zero context
Responds purely to what's in front of it. Catches the curse of knowledge — the stuff that's obvious to you and confusing to everyone else.
The Executor
Only cares if it can be done
Ignores theory. Looks at everything through "OK, but what do you do Monday morning?" If there's no clear first step, it says so.
Easiest Path — 3 Steps
Let Claude install it for you.
The skill is one open-source file on GitHub. You don't have to understand it or download it — you just point Claude at it.
- 1Open a brand-new Claude chat
- 2Paste the install prompt with the GitHub link (Step 01 below)
- 3Type "council this" on your next real decision
Setup
Four prompts, start to verdict.
Works in Claude Code and in the Claude app. Copy each one in order — you won't write a single line of code.
Step 01
Ask Claude to install the skill
No terminal, no downloads. Open a fresh chat in Claude (Code or the desktop/web app) and paste this. Claude fetches the file and puts it where it belongs.
Please install this Claude skill for me. The SKILL.md file lives in this GitHub repo: https://github.com/aiwithremy/claude-skills-llm-council
Set it up so I can start using it. Walk me through anything you need from me.Step 02
Confirm it's actually installed
Skills are silent until they trigger, so verify before you rely on it. If Claude can't find it, re-run step 1 and let it walk you through manual placement.
Do you have the llm-council skill installed? Show me the trigger phrases.Step 03
Bring it a real decision
Give it stakes and context — options, numbers, constraints. Vague input gets you five generic takes. Specific input gets you a verdict you can act on.
Council this: I have 8 hours a week outside my job. Should I put them into a paid newsletter, a YouTube channel, or freelance client work? I have 400 email subs, no audience on YouTube, and I need $1k/mo within 6 months.Step 04
Read the verdict, not just the advisors
The five takes are the raw material. The chairman section is the output — where they agree, where they clash, the recommendation, and the one thing to do first.
Which part of the council's verdict has the weakest evidence behind it? Push back on your own recommendation.What happens
Answer, argue, decide.
- 1Claude reframes your question so all five advisors get the same context — including anything useful sitting in your project files.
- 2The five advisors answer in parallel, each leaning all the way into their angle. No hedging, no balance.
- 3The answers get anonymized and handed back around: strongest take, biggest blind spot, what everyone missed.
- 4A chairman reads all of it and writes the verdict — agreements, clashes, blind spots, recommendation, and one first step.
The peer-review round is the part that makes it work. Skip it and you've just asked one AI the same thing five times.
What you get back
The verdict, in five parts.
- Where the council agrees — the high-confidence stuff multiple advisors landed on independently
- Where the council clashes — the real disagreements, presented as disagreements instead of smoothed over
- Blind spots the council caught — things that only surfaced during peer review
- The recommendation — an actual answer, not "it depends"
- The one thing to do first — a single next step, not a list of ten
Phrases that trigger it
- "council this"
- "run the council on ___"
- "pressure-test this"
- "stress-test this"
- "war room this"
- "debate this"
When to use it
Save it for the expensive calls.
The council shines when there's real uncertainty and being wrong costs you. If you already know the answer and just want validation, it'll probably tell you something you don't want to hear. That's the point.
Worth a council
- Should I launch a $97 workshop or a $497 course?
- Which of these three positioning angles is strongest?
- I'm thinking of pivoting from X to Y — am I crazy?
- Here's my landing page copy. What's weak?
- Should I hire help or build an automation first?
Just ask normally
- — What's the capital of France? (one right answer)
- — Write me a tweet. (creation, not judgment)
- — Summarize this article. (processing, not a decision)
- — Should I use markdown? (no real tradeoff)
Credit where it's due: the skill was built by Ole Lehmann, adapted from Andrej Karpathy's LLM Council. I just use it and wanted more people to know it exists.
Questions people ask
What is the LLM Council?
It's a free Claude skill that runs your question through five advisors with deliberately different thinking styles — Contrarian, First Principles, Expansionist, Outsider, Executor. They answer independently, peer-review each other's answers anonymously, and a chairman synthesizes one verdict. It's adapted from Andrej Karpathy's LLM Council, but it runs inside Claude with sub-agents instead of five separate models.
Do I need to know how to code to install it?
No. Open a new Claude chat, paste the install prompt with the GitHub link, and Claude fetches the skill and sets it up. If your setup needs manual placement, Claude tells you exactly where to put the file.
When should I not use the council?
Skip it for factual lookups, writing tasks, and summaries. The council is for decisions where being wrong is expensive and there's a genuine tradeoff — pricing, positioning, pivots, hiring, what to build next.
Does it cost anything?
The skill is free and open source. You're only using your normal Claude usage. It does burn more than a regular chat, though — five advisor passes plus five reviews plus a synthesis. Save it for calls that matter.
Why does the peer review step matter?
Without it you've just asked the same AI five times in five costumes. The review round is where advisors get anonymized responses and have to name the strongest take, the biggest blind spot, and what everyone missed. That's where the non-obvious stuff comes out.