LLM Council Skill for Claude: 5 AI Advisors That Pressure-Test Your Decisions (2026 Guide)

The LLM Council is a free Claude skill that runs your decision past five AI advisors — a Contrarian, a First Principles Thinker, an Expansionist, an Outsider and an Executor — then has them review each other anonymously, and finishes with a chairman's verdict: where they agree, where they clash, what everyone missed, a clear recommendation and the one thing to do first. You trigger it by typing "council this" in front of your question. I installed it today, and it's the most useful antidote I've found to Claude's habit of agreeing with whatever you already think.

What is the LLM Council skill, in plain English?

Ask Claude "should I launch this?" and it will find reasons to say yes. Ask "is this a bad idea?" and it will find reasons to say no. Same idea, opposite answers — because a single answer reflects the way you framed the question. That's fine for writing an email. It's risky when you're about to set a price, hire someone or rewrite your homepage.

The council fixes this by forcing disagreement on purpose. Each advisor is a thinking style, not a job title:

• ✓The Contrarian assumes there's a fatal flaw and goes looking for it.
• ✓The First Principles Thinker asks whether you're even solving the right problem.
• ✓The Expansionist looks for the upside you're not seeing.
• ✓The Outsider gets zero background, so it catches what's obvious to you and confusing to everyone else.
• ✓The Executor only cares what you actually do on Monday morning.

Where does the idea come from?

It's adapted from Andrej Karpathy's LLM Council — a small local web app he published as a weekend project. His version sends your question to several different AI models (through OpenRouter), has each model review and rank the others' anonymized answers, and then a "chairman" model writes the final response. The skill keeps that three-stage structure but runs it inside Claude: one model, five different thinking lenses, each running as its own sub-agent. The skill credits Karpathy for the method and Ole Lehmann for the Claude Code adaptation, and it's published by tenfoldmarc. As of October 2026 the skill's repo has close to 1,000 stars on GitHub.

How does a council session actually work?

• ✓It gathers context. Claude quickly looks for relevant files in your folder — like CLAUDE.md or a memory folder — so the advisors know your business, not a generic one.
• ✓It frames the question neutrally. If your question is too vague, it asks you one clarifying question, then moves on.
• ✓Five advisors answer in parallel. Each one leans fully into its angle — no hedging, no "on the other hand."
• ✓Five anonymous peer reviews. The answers are shuffled and labeled A–E, and each reviewer says which is strongest, which has the biggest blind spot, and what all five missed.
• ✓The chairman rules. One final verdict, which is allowed to side with a lone dissenter if that reasoning is strongest.
• ✓You get two files: a visual HTML report (which opens automatically) and a markdown transcript of everything.

When should I use the council — and when shouldn't I?

The rule of thumb: use it when being wrong is expensive and there's no single right answer. Here are examples from the kind of work my readers do every day:

How do I install the LLM Council in Claude Code?

git clone https://github.com/tenfoldmarc/llm-council-skill ~/.claude/skills/llm-council

Can I use it in the Claude app instead?

Yes, with one honest caveat. In the Claude app (Free, Pro, Max, Team and Enterprise), you upload skills as a ZIP: put SKILL.md inside a folder named llm-council, zip the folder, then go to Customize → Skills, click +, choose Create skill → Upload a skill and upload the ZIP. Skills need code execution and file creation turned on (for individual plans that's under Settings → Capabilities). Menu names change, so if yours look different, check Anthropic's help article on using skills.

How do I run the council?

Start your message with a trigger — "council this", "run the council", "pressure-test this", "stress-test this", "war room this" or "debate this" — and then give it real context. Numbers, audience, constraints, what you've already tried. The more specific you are, the less generic the advisors get.

How do I read the chairman's verdict?

• ✓Where the council agrees — points several advisors reached on their own. Treat these as your strongest signals.
• ✓Where the council clashes — real disagreements, both sides laid out. This is where your judgment and knowledge of your customers come in.
• ✓Blind spots — things that only surfaced in the peer reviews. In my experience, this is the section worth reading twice.
• ✓The recommendation — a real answer, not "it depends."
• ✓The one thing to do first — a single concrete step. If you do nothing else, do this.

Remember what it is: Claude reasoning from different angles about the information you gave it. It doesn't know your market better than you do, and it can be wrong. Use the verdict to sharpen your thinking, not to replace it.

Want to build skills like this yourself?

In my Zoom course you learn Claude Code hands-on — including installing, reading and writing your own skills. Or browse the skills I've already reviewed.

See the courseBrowse skillsקראו בעברית