/council

Overview

The council mode runs multiple AI agents in parallel to gather diverse perspectives on a task, then synthesizes their opinions into a final answer.

Stages:

1. Stage 1: All agents provide independent opinions (parallel)
2. Stage 2: Each agent reviews and ranks others' opinions (anonymized)
3. Stage 3: Chairman synthesizes all opinions weighted by rankings

Features:

• Dynamic persona suggestion based on question analysis
• Anonymized peer ranking to avoid bias
• Weighted synthesis based on peer rankings

Pre-Execution: Analyze Question & Gather Configuration

Before running the council:

1. Analyze the Question

First, analyze the user's question to determine appropriate perspectives/personas. Consider:

• Technical questions (code, architecture, security) → technical personas (Security Analyst, Systems Architect, Code Reviewer, etc.)
• Business questions (strategy, market, product) → business personas (Business Strategist, Product Manager, Financial Analyst, etc.)
• General knowledge questions (recommendations, advice, comparisons) → generalist personas (Research Analyst, Local Expert, Practical Advisor, Critical Thinker)
• Creative questions → creative personas (Creative Director, User Advocate, Trend Analyst)

2. Suggest Personas via AskUserQuestion

Use AskUserQuestion to suggest relevant personas based on your analysis:

Use AskUserQuestion with:
- question: "I've analyzed your question. Which perspectives would be most valuable?"
- header: "Perspectives"
- multiSelect: true
- options: [Generate 3-4 relevant personas based on the question type]

Example for "best hiking in Seattle":
  - label: "Local Expert", description: "Deep knowledge of Seattle area trails and conditions"
  - label: "Outdoor Enthusiast", description: "Practical hiking experience and recommendations"
  - label: "Research Analyst", description: "Comprehensive data on trail ratings and reviews"
  - label: "Critical Thinker", description: "Questions assumptions about 'best' and considers trade-offs"

Example for "should we use microservices":
  - label: "Systems Architect", description: "Scalability, infrastructure, distributed systems"
  - label: "DevOps Engineer", description: "Deployment, monitoring, operational complexity"
  - label: "Developer Experience", description: "Team productivity, learning curve, tooling"
  - label: "Devil's Advocate", description: "Challenge assumptions, identify hidden risks"

3. Number of Agents

Also ask about number of agents if not specified:

Use AskUserQuestion with:
- question: "How many agents should participate?"
- header: "Agents"
- options:
  - label: "2 agents", description: "Quick analysis with two perspectives"
  - label: "3 agents (Recommended)", description: "Balanced coverage"
  - label: "All available", description: "Maximum perspectives"

Skip questions for options explicitly provided in $ARGUMENTS.

Arguments

Parse these from $ARGUMENTS:

• Task/question (required) - the main argument
• --agents N - Number of agents to use (default: all available)
• --personas JSON - Custom personas as JSON array (see below)
• --no-personas - Disable automatic persona assignment
• --verbose or -v - Show detailed progress

Available Agents

Check which agents are available:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/list_agents.py"

Execution

Run the council with custom personas (based on user selections):

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/run_council.py" --task "YOUR_TASK_HERE" --personas '[{"name":"Local Expert","focus":"Seattle area knowledge"},{"name":"Outdoor Enthusiast","focus":"hiking experience"},{"name":"Critical Thinker","focus":"trade-offs and nuance"}]' [--agents N] [-v]

Or run with automatic persona inference:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/run_council.py" --task "YOUR_TASK_HERE" [--agents N] [-v]

Output Formatting

Parse the JSON result and present it as:

Council Result - N agents deliberated

Opinions:

Agent	Persona	Opinion Summary	Confidence
claude	Security Analyst	Key finding...	85%
codex	Systems Architect	Key finding...	90%
gemini	Devil's Advocate	Key finding...	70%

Peer Rankings (1=best):

• Agent A: 1.5 avg rank
• Agent B: 2.0 avg rank

Chairman's Synthesis:

[The synthesized final answer]

Consensus Level: X%

Key Insights:

• Insight 1
• Insight 2
• Insight 3

Dissenting Views:

• [Any notable disagreements]

Example Usage

/tower:council "Should we use TypeScript or JavaScript for the frontend?"
/tower:council "Review the security of this authentication flow" --agents 3
/tower:council "Evaluate this startup idea: AI meal planning" --verbose