/onex:council

A Claude Code skill that pressure-tests a decision through a 5-advisor council, anonymous peer review, and a chairman synthesis — adapted from Andrej Karpathy's LLM Council methodology. Saves a visual HTML briefing report and a full markdown transcript to docs/council/.

Ships in the onex plugin — invoked as /onex:council.

What it does

Trigger it with /onex:council (or "council this", "war room this", "pressure-test this", "stress-test this", "debate this", or a real decision phrased as "should I X or Y / I'm torn between / validate this / which option"). Claude will:

1. Scan the workspace for relevant context (CLAUDE.md, memory/, prior transcripts, files referenced in the question).
2. Frame the question as a neutral prompt embedded with that context.
3. Convene 5 advisors in parallel — each leans fully into one thinking style and produces a 150–300 word response.
4. Run an anonymous peer review — 5 reviewers see the 5 responses as A–E (randomized) and answer: strongest? biggest blind spot? what did all five miss?
5. Synthesize a chairman's verdict — five-section output: agreement / clash / blind spots / recommendation / one thing to do first.
6. Generate a visual HTML report at docs/council/council-report-<ts>.html and open it.
7. Save the full transcript at docs/council/council-transcript-<ts>.md.
8. Reply briefly in chat — quote only the Recommendation and One-thing-to-do-first; link the artifacts.

The five advisors

Advisor	Thinking style	Natural tension
The Contrarian	Looks for what's wrong, what's missing, what will fail.	⇄ Expansionist
The First Principles Thinker	Strips the question to its essentials and rebuilds from scratch.	⇄ Executor
The Expansionist	Looks for upside everyone else is missing.	⇄ Contrarian
The Outsider	Has zero context; responds purely to what's in front of them — catches the curse of knowledge.	(sits in the middle)
The Executor	Only cares whether it can be done, and what to do Monday morning.	⇄ First Principles

Good council 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 council questions

• "What's the capital of France?" (lookup)
• "Write me a tweet" (creation)
• "Summarize this article" (processing)
• "Should I use markdown?" (no real tradeoff)

The council is for genuine uncertainty with real stakes.

Output files

docs/council/council-report-YYYYMMDD-HHMMSS.html    # visual briefing — what the user reads
docs/council/council-transcript-YYYYMMDD-HHMMSS.md  # full transcript — what the next session reads

The HTML report is a self-contained single file (inline CSS, no JS) — clean briefing-document styling, an alignment grid showing where advisors aligned/diverged, the chairman's verdict at the top, and collapsible sections for the full advisor responses and peer reviews.

Install

This skill ships in the onex Claude Code plugin:

/plugin marketplace add onextech/skills
/plugin install onex@skills

Then invoke it with /onex:council.

Credits

Adapted from Andrej Karpathy's LLM Council methodology — dispatch the question to multiple models, have them peer-review anonymously, then a chairman synthesizes the final answer. This skill runs the same loop inside Claude using parallel sub-agents with different thinking lenses instead of different models.

License

MIT