Council: When One AI Opinion Isn’t Enough

Publié le par

How I built a system that makes three AI models debate before answering your questions

The Problem with Single-Model Answers

Last month, I asked Claude whether one startup should adopt microservices. The answer was confident and well-reasoned: “Yes, microservices will give you flexibility and scalability.”

Then I asked Gemini the same question. Equally confident: “No, stick with your monolith—microservices add complexity you don’t need yet.”

Two AI models. Two opposite recommendations. Both completely sure of themselves.

This is the dirty secret of AI assistants: they’re trained to sound confident, even when the answer genuinely depends on context they don’t have. There’s no built-in mechanism to say “actually, this is debatable.”

So I built one.

Introducing Council

Council is a plugin for Claude Code that orchestrates three AI models—Claude, Gemini, and Codex—to debate your questions before giving you an answer.

Instead of getting one model’s opinion, you get:

  • Multiple perspectives from models with different training and strengths
  • Structured disagreement when the models don’t agree (which is valuable data)
  • A confidence score based on how quickly they converged
  • A full audit trail of the reasoning, saved as markdown

Think of it as a board of advisors that must reach consensus before advising you—except these advisors respond in minutes, not days.

How It Works

When you ask Council a question, here’s what happens:

  1. Persona Assignment: Each model gets a relevant expert persona (e.g., “Security Architect”, “Performance Engineer”, “System Designer”)
  2. Round 1 – Initial Positions: All three models provide their analysis independently
  3. Round 2+ – Rebuttals: Each model sees the others’ arguments (anonymized) and responds with counter-arguments or concessions
  4. Convergence Detection: The system measures agreement. If models converge, it stops early. If they don’t, it continues or escalates to Devil’s Advocate mode.
  5. Peer Review: The “chairman” model scores each response for accuracy, completeness, reasoning, and clarity
  6. Synthesis: A final answer combines the strongest arguments, notes any dissenting views, and provides a confidence score

Four Deliberation Modes

Consensus (default): Models discuss until they agree. Best for technical questions and design decisions.

Debate: One model argues FOR, one argues AGAINST. Best for controversial topics or binary choices.

Devil’s Advocate: Red Team attacks your idea, Blue Team defends it, Purple Team synthesizes. Best for stress-testing proposals.

Vote: Each model votes with justification. Best for multiple-choice decisions.

Real Example

I asked Council: “Python async scraper hitting rate limits—backoff, semaphore, or queue?”

One model pushed for exponential backoff. Another advocated for semaphores. The third suggested queues.

Their synthesized answer? “You need all three in layers.”

They had debated themselves into a more complete solution than any single model would have proposed:

  1. Queue-based foundation
  2. Per-host semaphores (not global)
  3. Token bucket rate limiting
  4. Exponential backoff with jitter
  5. Adaptive tuning

Total time: ~3 minutes. The answer came with a 0.91 confidence score and a full reasoning trail.

Getting Started

If you use Claude Code, installation takes 30 seconds:

# Add the marketplace
claude plugin marketplace add bacoco/Council-board-skill

# Install the plugin
claude plugin install council@council-board

Then just ask naturally:

  • “Ask the council: should we use PostgreSQL or MongoDB?”
  • “Debate this: React vs Vue for our new project”
  • “Challenge my design for the authentication system”
  • “What does Claude think about this?” (direct mode, skips deliberation)

When to Use Council

Use Council when:

  • The decision has real consequences
  • You want to surface tradeoffs, not hide them
  • You suspect there might be angles you haven’t considered
  • You need to justify a decision to stakeholders (the audit trail helps)

Skip Council when:

  • You need a quick factual answer
  • The question has an objectively correct answer
  • Speed matters more than thoroughness

The Philosophy

Council isn’t about replacing human judgment. It’s about giving you better inputs for that judgment.

When three AI models agree, you can move forward with confidence. When they disagree, that disagreement is shown clearly—and often reveals the genuine complexity of a decision.

The goal is to keep you in the loop as the decision-maker, while ensuring you’ve heard from multiple perspectives before you commit.

Try It

GitHub: github.com/bacoco/Council-board-skill

The decisions that keep you up at night deserve more than one opinion—even if that opinion comes from AI.