Here's the uncomfortable thing about asking AI for advice. It mostly tells you what you want to hear.

Stanford researchers just put numbers on it. They tested 11 of the big models, including ChatGPT, Claude, and Gemini, across more than 11,000 real interpersonal dilemmas. The models endorsed the user's actions about 50% more often than a human would, and kept doing it even when the user described something manipulative or clearly in the wrong. It was peer reviewed and published in Science in March 2026. You can read the paper or Stanford's writeup.

So when you ask AI "is this a good idea?" you're not really getting a second opinion. You're getting a yes man with a great vocabulary.

The LLM Council skill fixes that. Instead of one agreeable voice, it runs your question past five advisors who each have a different job, makes them critique each other, and ends with a chairman who gives you the actual call. This post is the whole thing: what each advisor does, how the argument works, and exactly how to set it up in a couple of minutes.

What the LLM Council actually is

It's a free, open skill for Claude, specifically Claude Code, the version that can install and run skills. You set it up once. After that, whenever you have a real decision, you type something like "council this" and Claude stops answering as itself and spins up five separate advisors instead.

Each advisor only sees your question and writes their own take. Then they read everyone else's answers blind, with no names attached, and push back on each other. A chairman reads the whole debate and hands you one final recommendation. You walk away with a clean report and a transcript of the argument that got you there.

One AI agreeing with you is a yes man. Five AIs arguing in front of a chairman is a second opinion.

The five advisors

1. The Contrarian

Only looks at what could go wrong. Its entire job is to find the hole in your plan, the risk you're glossing over, the way this quietly blows up. If your idea survives the Contrarian, it's a real idea.

2. The First-Principles Thinker

Throws out how things are "normally" done and rebuilds from scratch. It asks whether you're even solving the right problem, then takes your assumptions apart one by one to see if the answer still holds.

3. The Expansionist

The optimist with a point. It hunts for the upside you're leaving on the table: the bigger version of the plan, the opportunity sitting right next to the one you asked about.

4. The Outsider

Gets almost no context on purpose. Because it isn't deep in the weeds with you, it asks the obvious question everyone too close to the problem forgot to ask. This is the one that kills groupthink.

5. The Executor

Doesn't care about theory. It only wants to know what you actually do Monday morning. First step, this week, in the real world. It turns all the debate into a plan you can act on.

How they argue, and why that's the whole point

The magic isn't the five personalities. It's that they review each other blind. Each advisor writes first, with no idea what the others said, so you get five genuinely independent reads instead of four models nodding along to the first one. Then they see the other answers stripped of names and argue the weak points.

The chairman comes last. It weighs the debate, calls out where the advisors disagree, and gives you a clear recommendation instead of a mushy "it depends." You're not reading five essays and guessing. You're getting a verdict with the reasoning attached.

How to set it up

Two ways, pick whichever fits you. Either takes a couple of minutes, and the skill is free and open source under an MIT license.

The beginner way: just ask Claude

Open Claude Code and paste this in:

Please install this Claude skill for me. The SKILL.md file is in this
GitHub repo: https://github.com/tenfoldmarc/llm-council-skill

Claude reads the repo and sets it up for you. That's the entire install.

The one-line way: Terminal

If you're comfortable in the terminal, clone it straight into your Claude skills folder:

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

Then open Claude Code and the skill is ready to go.

How to actually run it

Put any of these in front of your decision and Claude will trigger the council: "council this," "run the council on [your decision]," "pressure-test this," "stress-test this," or "debate this." Then hand it a real question, something like "should I quit my job to do content full time?" and let the five of them go at it.

When to actually use it

Don't council your grocery list. Save it for the decisions that are expensive to get wrong: a big hire, quitting something, a pricing change, a launch, a hard conversation you keep replaying in your head. The entire value is that it won't just tell you you're right. For everything else, normal Claude is faster.

I started running my own bigger calls through it, and the part that gets me every time is the Contrarian. It says the thing I was hoping nobody would bring up. That's exactly why it's worth having.

Run your next big decision through all five before you trust your gut. It's the cheapest second opinion you'll ever get.
– Anir

If this is the kind of AI thinking you want working on your own business, apply to work with me. And to catch the next breakdown, join the newsletter.