Run structured debates among multiple AI models to handle complex decisions
When a single perspective is not enough to answer a critical question, imagine a room of analysts with different temperaments arguing their cases. This tool implements that scene with actual AI agents, organizing diverse viewpoints and a final judgment for your query. In short, it is a system where multiple AI models take on roles, debate, and return the results as a structured report.
Collision and verification of 18 perspectives
The tool manages the decision-making process in stages. It assigns 18 virtual members with different analytical viewpoints and has each analyze the problem independently. They then undergo a cross-examination phase to rebut each other's claims, ultimately producing a consensus conclusion along with remaining uncertainties, dissenting opinions, and execution criteria. The process focuses on reducing unfounded confidence by clearly distinguishing facts, reasoning, assumptions, and unknown information.
Choosing debate scale for the situation
You do not need to convene the full committee for every decision.
- Full mode Goes through independent analysis, cross-examination, final stance, and synthesis stages. Suitable when stakes are high and different frameworks conflict.
- Quick mode Performs problem restatement, rapid analysis, and final stance only. Used when broad review is needed but full confrontation rounds are unnecessary.
- Duo mode Consists of open stance, direct response, and final statement. A format where two members face off when only one core tension matters.
Platform-specific installation and model assignment
It is designed to work with multiple AI clients, including Claude Code, Codex, Gemini CLI, and OpenCode. The installation script automatically detects available providers and assigns seats. It aims to reduce single-model bias by assigning opposing sides to different model families whenever possible. The --dry-route option allows you to preview routing without executing the actual committee, while --no-auto-route lets you maintain default settings.
Post-decision tracking and verification
The process does not end once a conclusion is reached. Before execution, you must record predictions, owners, review dates, and evidence that could change the recommendation. Later, at checkpoints, results are classified as confirmed, revised, reversed, or uncertain, allowing you to track them without rewriting the original rationale. This helps ensure decisions do not remain merely as text but become verifiable data for later review.
Pre-application checks
This tool is not suitable for cheap choices where fact-checking or experimentation is possible. It should not be used to manufacture support for answers that are already decided. It is released under the MIT license and supports Shell-script-based installation and verification. Before use, check the protocol in SKILL.md and the member contracts in the agents/ folder, and test with --dry-run to ensure your environment’s supported AI providers are detected correctly.
