name: council
description: Convene an LLM Council for deliberative multi-model reasoning. Uses 3-stage process (Query, Peer Review, Chairman Synthesis) for deeper analysis than simple consensus polling.
allowed-tools: Bash, Read, AskUserQuestion

LLM Council

Convene a council of AI models for deliberative reasoning on technical decisions or research synthesis. Based on Karpathy's LLM Council concept.

Overview

The council uses a 3-stage process:
1. Query: Each council member responds independently
2. Peer Review: Members review anonymized responses (Response A, B, C)
3. Synthesis: Chairman combines insights into final recommendation

Process

Step 1: Validate Question

Ensure the question is appropriate for council deliberation:
- Technical decisions: Architecture, library selection, design patterns
- Research synthesis: Gathering perspectives on a technical topic
- Complex trade-offs: Questions with multiple valid approaches

If too simple, suggest /cc-unleashed:consensus instead.

Step 2: Offer Configuration

Use AskUserQuestion to let user customize (or accept defaults):

Question: "How would you like to configure the council?"
Options:
1. Use defaults (Recommended) - GPT-4o, Gemini 2.5 Flash, Grok 4, Claude Sonnet as chairman
2. Customize council members - Select which models participate
3. Customize chairman - Choose who synthesizes the final answer

If user selects "Customize council members" or "Customize chairman":

First, fetch available models from OpenRouter:

$HOME/.claude/plugins/marketplaces/rlg-unleashed-marketplace/skills/council/scripts/council.sh --discover --format table

This returns models from major providers (OpenAI, Anthropic, Google, xAI, Meta, Mistral) with context length and pricing.

Then use AskUserQuestion to let them pick from the available models:
- For council: Pick 3-5 models (recommend diversity across providers)
- For chairman: Pick one model (Claude or GPT recommended for synthesis)

Step 3: Run Council Script

Execute the shell script with configuration:

$HOME/.claude/plugins/marketplaces/rlg-unleashed-marketplace/skills/council/scripts/council.sh \
  --question "USER'S QUESTION" \
  --council "model1,model2,model3" \
  --chairman "chairman-model"

Default invocation:

$HOME/.claude/plugins/marketplaces/rlg-unleashed-marketplace/skills/council/scripts/council.sh \
  --question "USER'S QUESTION"

Step 4: Present Results

The script outputs structured results. Present to user:

1. Individual Perspectives (Stage 1 summary)
2. Peer Review Highlights (Stage 2 - key agreements/disagreements)
3. Chairman's Synthesis (Stage 3 - final recommendation)

When to Use Council vs Consensus

Prerequisites

• OPENROUTER_API_KEY environment variable
• curl and jq installed (standard on most systems)
• Optional: ~/.claude/config/council.json for default configuration

Error Handling

• Missing API key: Tell user to set OPENROUTER_API_KEY
• Model failures: Script continues if 2+ council members succeed
• Missing curl/jq: Tell user to install these tools

Configuration File

Optional ~/.claude/config/council.json:

{
  "council": [
    "openai/gpt-4o",
    "google/gemini-2.5-flash",
    "x-ai/grok-4-fast"
  ],
  "chairman": "anthropic/claude-sonnet-4",
  "max_tokens": 1000,
  "timeout_seconds": 90
}

Model Discovery

To see all available models from major providers:

# JSON format (for parsing)
$HOME/.claude/plugins/marketplaces/rlg-unleashed-marketplace/skills/council/scripts/council.sh --discover

# Human-readable table
$HOME/.claude/plugins/marketplaces/rlg-unleashed-marketplace/skills/council/scripts/council.sh --discover --format table

Major providers included: OpenAI, Anthropic, Google, xAI, Meta, Mistral

References

See reference.md for:
- Default council and chairman configuration
- Recommended council compositions
- Prompt templates for each stage