LLM Council — AI skill for Claude Code

llm-council

[](https://github.com/sponsors/YonasValentin)

A Claude Code skill that stops you from trusting the first answer you get.

You ask Claude a question. You get one answer. It sounds smart. You move on. Two days later you realize the answer was shaped by how you asked, not by what was true. This skill fixes that. It spins up five advisors with genuinely different thinking styles, has them answer your question independently, makes them peer-review each other anonymously, and gives you a verdict with a single clear next step.

Runs entirely inside Claude Code as a native skill. No external services, no API keys, no extra installs.

What this is

`llm-council` is a Claude Code skill. If you have not used Claude Code skills before, they are bundles of instructions plus prompt templates that sit in your `~/.claude/skills/` directory. Claude Code loads them on demand when your question matches the skill's triggers.

When you type a phrase like "council this" or "pressure-test this", the skill takes over:

1. It scans your workspace for context files (CLAUDE.md, AGENTS.md, memory folders) so the advisors have something real to chew on.
2. It frames your question as a neutral prompt.
3. It spawns five Claude Code sub-agents in parallel, one per advisor thinking style.
4. It anonymizes the five responses, then spawns five reviewers in parallel to peer-review the set.
5. It synthesizes everything into a five-section verdict.
6. It writes a clean HTML briefing and a full markdown transcript into ./council//, then opens the briefing in your browser.

Total time is around 2 to 4 minutes depending on how much context the advisors need to read.

Why I built it

I kept noticing the same pattern. I would ask Claude something, get a confident answer, and then catch myself realizing the answer just matched my framing back at me. Ask the same question with the op