╔═══════════════════════════════════════════════════════╗
   ║           CLAUDE CODE COUNCIL                        ║
   ║                                                      ║
   ║   5 advisors.  3 model families.                     ║
   ║   Anonymous peer review.  One verdict.               ║
   ╚═══════════════════════════════════════════════════════╝

---

Claude Code Council is a Claude Code skill that spawns 5 AI advisors with different thinking styles — across 3 different model families — to argue a question, anonymously peer-review each other, and produce a synthesized verdict.

Use it when being wrong is expensive.

Why?

One AI gives you one perspective. One perspective has blind spots.

Claude Code Council gives you five perspectives from three model families (Claude, GPT-5.4, Gemini 3.1 Pro), then hides who said what, and has a panel judge the arguments on merit alone.

The result: decisions grounded in adversarial debate, not confirmation bias.

How It Works

  YOU: "/council Should we migrate our database from Postgres to MongoDB?"
   |
   v
  GATHER CONTEXT
   |  Read relevant code, query memory for prior decisions
   |  Build a context brief (max 500 words)
   |
   v
  SPAWN 5 ADVISORS (all in parallel, ~30 seconds)
   |
   |  +---> The Contrarian (Claude)         "Find what will fail."
   |  |
   |  +---> First Principles (Claude)       "What are we ACTUALLY solving?"
   |  |
   |  +---> The Expansionist (GPT-5.4)      "What upside is being missed?"
   |  |
   |  +---> The Outsider (Gemini 3.1 Pro)   "What's obvious to a newcomer?"
   |  |
   |  +---> The Executor (Claude)           "What do you do Monday morning?"
   |
   v
  ANONYMOUS PEER REVIEW
   |  Responses shuffled randomly (A-E)
   |  Identities hidden — no one knows which model wrote what
   |  Panel evaluates:
   |    1. Strongest response and why
   |    2. Biggest blind spot
   |    3. What ALL FIVE missed (most important)
   |
   v
  +------------------------------------------+
  |  COUNCIL VERDICT                         |
  |                                          |
  |  RECOMMENDATION: Stay on Postgres.       |
  |                                          |
  |  KEY INSIGHTS:                           |
  |  - MongoDB loses ACID for zero benefit   |
  |    in this use case (Contrarian)         |
  |  - The real problem is query speed, not  |
  |    schema flexibility (First Principles) |
  |  - Add read replicas instead (Executor)  |
  |                                          |
  |  RISKS: Migration would take 3 months    |
  |  and break 12 downstream services        |
  |                                          |
  |  DISSENT: Expansionist argued for a      |
  |  hybrid approach — Postgres for writes,  |
  |  Redis for hot reads                     |
  +------------------------------------------+

The 5 Advisors

Advisor	Model	Job	Thinks About
The Contrarian	Claude	Find what will fail	Hidden risks, false assumptions, historical failures, scaling problems
First Principles	Claude	Reframe the problem	Wrong variables being optimized, inherited assumptions, what actually matters
The Expansionist	GPT-5.4	Find missed upside	Adjacent opportunities, 10x improvements, things you haven't imagined
The Outsider	Gemini 3.1 Pro	Fresh eyes	Curse of knowledge, unclear value props, what only makes sense to insiders
The Executor	Claude	Monday morning action	Execution path, fastest test, hidden dependencies, reversibility

Why 3 model families? Different training data = different blind spots. Claude excels at reasoning. GPT-5.4 brings a different creative perspective. Gemini sees things both miss. The anonymous shuffle prevents the panel from anchoring to model reputation.

Install

# Add the marketplace
/plugin marketplace add mtarcure/claude-code-council

# Install
/plugin install council@claude-code-council

# Reload
/reload-plugins

Requirements

• Claude Code (CLI, desktop app, or web)
• OpenAI API key (for GPT-5.4 advisor) — set OPENAI_API_KEY in your environment
• Gemini API key (for Gemini advisor) — set GEMINI_API_KEY in your environment

Without API keys, the council still works — it just runs all 5 advisors on Claude (still valuable, just less diverse).

Usage

# Ask the council anything
/council Should we rewrite the auth system or patch it?

# Architecture decisions
/council Monolith vs microservices for our 5-person team?

# Strategy
/council Should we launch on Product Hunt or focus on direct sales?

# Technical trade-offs
/council Redis vs Kafka for our event queue?

When to use it:

• Decisions where being wrong costs >$10K or >1 month
• Architecture choices that are hard to reverse
• Strategy questions with no obvious right answer
• Any time you catch yourself thinking "I'm probably right but..."

When NOT to use it:

• Trivial decisions ("tabs vs spaces")
• Questions with clear, factual answers
• When you need speed over deliberation

How the Anonymous Review Works

After all 5 advisors respond, the system:

1. Shuffles responses randomly and assigns letters A-E
2. Hides which advisor (and which model) wrote each response
3. Evaluates on merit alone — the panel doesn't know if response B came from GPT-5.4 or Claude

This prevents anchoring bias. In testing, reviewers rated responses differently when they knew the model vs when they didn't. Anonymization produces better verdicts.

Output Format

COUNCIL VERDICT
===============

Question: {your question}

RECOMMENDATION: {clear, actionable recommendation in 1-2 sentences}

KEY INSIGHTS:
- {strongest point from advisors, attributed by role}
- {second strongest point}
- {what the peer review caught that no advisor saw}

RISKS TO WATCH:
- {top risk from Contrarian, if valid}
- {any risk the peer review surfaced}

NEXT STEP: {one concrete action from the Executor}

DISSENT: {any advisor who disagreed — included even if overruled}

Configuration

Env Variable	Description	Required
OPENAI_API_KEY	For GPT-5.4 Expansionist advisor	Optional (falls back to Claude)
GEMINI_API_KEY	For Gemini Outsider advisor	Optional (falls back to Claude)

Plugin Structure

ai-council/
  .claude-plugin/
    marketplace.json
  plugins/council/
    .claude-plugin/
      plugin.json
    skills/
      council/
        SKILL.md          # Full council orchestration logic

Inspired By

• Anthropic's multi-agent research (90.2% better than single-agent on complex tasks)
• Adversarial collaboration in science (pre-registration + blinded review)
• Red team / blue team exercises in security

License

MIT

Built By

Mtarcure — part of the WireWork multi-agent system.