/llm-council: make Claude argue with you, not agree

How it works

1. 1Neutral brief. Your question is rewritten without your own opinion ("I'm sure", "Good idea?"). You see the brief and what was removed. No advisor sees your original wording.
2. 2Five advisors, independently. Contrarian, First-principles thinker, Expansionist, Outsider and Executor each answer as a separate subagent. None can see the others. Each picks GO, NO-GO or CHANGE and gives a confidence.
3. 3Anonymous review. The five answers are shuffled and labelled A–E. Three reviewers critique them without knowing who wrote what, and are told to reward the strongest dissent, not the majority.
4. 4Chairman. A verdict (go / no-go / change it), the strongest objection still standing, what would change the decision, and 3 next steps.

Or copy the skill folder (in a terminal)

Example

/llm-council I'm raising my app's price from $9 to $19 a month next week. I'm sure users will pay. Good idea?

VERDICT: CHANGE IT — Charge $19 to new signups next week; existing users stay at $9 for now.
...
NEXT 3 STEPS:
1. Pull subscriber count, monthly churn and cost per user — you'll know whether $9 loses
   money.
2. Put $19 on new signups only and compare 2 weeks against the prior 4 — it works if
   revenue per visitor is at or above baseline.
3. Survey existing users on what price they'd accept (a price-sensitivity survey) — this
   works if at least 60% say $19 or less is acceptable.

Chairman output from a real run, trimmed. Full run in examples/03-demo-price-increase.md.

Good to know

• A run takes about 3 minutes and makes 9 subagent calls. It counts toward your Claude plan's usage.
• It's a thinking aid. It doesn't eliminate sycophancy: all advisors run on the same model family and can share a blind spot.

Why

A Stanford study published in Science (Cheng et al., March 26, 2026) tested 11 AI models: on average, they affirmed users' actions 49% more often than humans did. The idea comes from Andrej Karpathy's llm-council, where several models answer, rank each other anonymously and a chairman writes the final answer. This version runs inside Claude Code with subagents.