← All guidesThe full lessons are inside AI Monetizers.Try it free for 7 days
Full community guide

LLM Council — the complete breakdown

Every advisor role, how to read a split verdict, follow-up prompts to push it harder, and the real install method that actually works.

1.7KGitHub stars
Built byOle Lehmann
Freeopen source
Install Triggers The 5 advisors Reading the verdict Pushing it harder Good vs bad questions Troubleshooting

Install it

The repo's "just paste this and Claude installs it" method doesn't reliably work. This is the method that actually does:

01
Go to the GitHub repo

github.com/aiwithremy/claude-skills-llm-council

02
Download it as a ZIP

Click the green "Code" button on the repo page, then "Download ZIP."

03
Open Claude settings

Go to Settings → Capabilities (or Skills, depending on your plan).

04
Upload the skill

Find the option to upload a skill, and select the ZIP file you downloaded.

How to trigger it

/llm-council "council this" "run the council on [question]" "pressure-test this" "stress-test this" "war room this"

The 5 advisor roles

Each advisor is deliberately built to think from a different, sometimes opposing angle — that's what stops the whole thing from collapsing back into agreement.

The Contrarian
Actively looks for what kills your idea. Its job is to find the weakest point, not to be balanced — treat its objections as the sharpest test your plan has to survive.
First Principles
Ignores the specific question you asked and instead asks what you're actually trying to solve. Often the most useful advisor when you've gotten attached to one specific solution.
The Expansionist
Hunts for upside you haven't considered — bigger versions of the idea, adjacent opportunities, what you'd do if resources weren't a constraint.
The Outsider
Gets none of your context or backstory. This is deliberate — it catches the obvious issues you've stopped noticing because you're too close to the problem.
The Executor
Doesn't care about theory. Only cares what you actually do tomorrow morning. Use this one to convert the whole discussion into a real next action.

How to read the verdict

The chairman's job isn't to average the five opinions into a safe middle answer — it's structured to show you exactly where they actually landed:

01
Where they agree

If all five advisors land in the same place despite thinking from different angles, that's a strong signal — treat it as high-confidence.

02
Where they clash

This is the most valuable part of the output. A clash usually means there's a real tradeoff you need to consciously decide on, not a bug in the process.

03
Blind spots caught

Pay special attention to anything the Outsider or Contrarian flagged that you hadn't considered — that's the whole reason to run a council instead of asking once.

04
The one thing to do first

This is your actual action item. If the verdict doesn't give you one clear next step, ask it directly: "what's the single next action based on this?"

Pushing it harder

If the first verdict feels too safe or too obvious, use these follow-ups instead of just accepting it:

Force disagreement

Which advisor do you think is most wrong here, and why? Argue their position as strongly as possible before telling me if they're actually right.

Test the verdict against a real constraint

Assume I only have half the budget/time I described. Does the verdict change?

Check for confirmation bias

Did any advisor just tell me what I probably wanted to hear? Re-run their section assuming they have to disagree with my framing.

Good vs bad questions

✓ Good questions
  • "Should I launch a $97 workshop or a $497 course?"
  • "Which of these 3 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 a VA or build an automation first?"
✗ Bad questions
  • "What's the capital of France?" — one right answer
  • "Write me a tweet" — that's creation, not a decision
  • "Summarize this article" — that's processing, not judgment

Be specific. "Should I launch a course?" is vague. "Should I launch a $299 AI workflow course targeted at senior freelancers before September?" gives every advisor something real to argue over.

Troubleshooting

Skill doesn't show up after uploadConfirm you uploaded the ZIP directly, not just the SKILL.md file on its own — some setups require the full folder structure.
Trigger phrase doesn't do anythingStart a fresh conversation after installing — skills sometimes don't apply retroactively to a chat that was already open.
Verdict feels too agreeableYour question was probably too vague. Add real numbers, names, or constraints, then re-run it.
Advisors all say the same thingUse the "force disagreement" follow-up prompt above to make the model actually separate the perspectives.

This is adapted from Andrej Karpathy's original LLM Council methodology. Credit for this Claude-specific version goes to Ole Lehmann.