council

name	council
description	Multi-model consensus for tough decisions. Fans out questions to frontier AI models via OpenRouter and synthesizes a single answer.

Council

Ask hard questions to multiple frontier models, get back a synthesized answer. Different models have different blindspots and strengths — triangulating across 4 top-tier models and synthesizing beats relying on any single perspective.

The Problem

Every AI model has systematic biases and blindspots. Claude is cautious and thorough but can over-qualify. GPT tends toward confident consensus answers. Gemini brings strong reasoning but different priors. Grok is direct and contrarian.

For genuinely tough decisions — architecture choices, complex tradeoffs, uncertain domains — you want diverse perspectives, not a single model's take.

Triggers

• /council "question" — ask the council directly
• "ask the council about..." or "what would the council say about..."
• Any complex question where model diversity would help

What Happens

1. Context gathering (optional): search your codebase/docs for relevant background
2. Question framing: reframe the raw question with that context so models have what they need
3. Parallel query: send to all models simultaneously via OpenRouter
4. Synthesis: read all responses and synthesize into a single answer

You only see the synthesis. Raw model outputs stay behind the scenes unless you ask.

The Models

Top-tier only — these are supposed to be tough questions:

Model	OpenRouter ID
Claude Opus 4.6	anthropic/claude-opus-4.6
GPT-5.4	openai/gpt-5.4
Gemini 3.1 Pro	google/gemini-3.1-pro-preview
Grok 4.1	x-ai/grok-4.1-fast

To customize models, edit the MODELS dict in council.py.

Setup

1. Get an OpenRouter API key
2. Set it via environment variable or config file:

# Option A: environment variable
export OPENROUTER_API_KEY="sk-or-..."

# Option B: config file (path is configurable in council.py)
mkdir -p ~/.config/council
echo '{"api_key": "sk-or-..."}' > ~/.config/council/openrouter.json

3. Install the one dependency:

pip install httpx
# or: uv pip install httpx

Usage

In a Claude Code session

/council "should we use sqlite or postgres for the job queue?"

The agent will:

1. Optionally gather context from your codebase
2. Frame the question with that context
3. Query all 4 models in parallel
4. Synthesize: "The consensus is X. GPT and Gemini emphasized Y, while Grok raised Z."

From the command line

python council.py "should we use sqlite or postgres for the job queue?"

From Python

from council import ask_council_sync, format_for_synthesis

result = ask_council_sync("your question here")
print(format_for_synthesis(result))

Options

• Pass include_context=False to skip codebase context gathering (for general questions)
• Pass a custom models list to override the defaults

When to Use

Good for:

• Technical architecture decisions
• Complex tradeoffs with multiple valid approaches
• Questions where model diversity surfaces different angles
• Triangulating on uncertain topics

Not for:

• Simple factual lookups
• Quick questions with obvious answers
• Anything a single model handles fine

Cost

Roughly $0.05–0.15 per query depending on response lengths. Uses frontier models so not cheap — but if the question is worth asking multiple models, it's probably worth the cost.

Adaptation

The script is self-contained (~200 lines, one dependency). To integrate into your project:

1. Copy council.py into your project
2. Optionally implement gather_context() to search your own data for relevant background
3. Reference this skill in your CLAUDE.md so your agent knows about /council

The default gather_context() is a no-op that passes questions through unchanged. Override it to add project-specific context enrichment.