Summary This project tackles LLM agreeableness, which can bias decision-making. It offers a Claude Code skill simulating a council of five AI advisors with distinct thinking styles. These advisors anonymously peer-review each other before a synthesized verdict is delivered. This empowers users, especially those making critical business decisions, with more trustworthy insights beyond a single AI's initial response.

How It Works

The skill scans workspace context, neutralizes the user's question, and spawns five parallel AI advisors: The Contrarian (failure points), First Principles Thinker (core problem), Expansionist (upside), Outsider (context bias), and Executor (action). Responses are anonymized for peer review. A "Chairman" synthesizes findings into a verdict, highlighting agreements, clashes, blind spots, recommendations, and next steps. This multi-perspective, peer-reviewed approach aims for objective output.

Quick Start & Requirements

• Installation: Recommended: git clone https://github.com/tenfoldmarc/llm-council-skill ~/.claude/skills/llm-council. Manual: Create ~/.claude/skills/llm-council/ and place SKILL.md inside. Restart Claude Code.
• Prerequisites: Claude Code environment.
• Usage: Trigger with council this, run the council, etc., followed by your question and context.
• Output: Automatically opening HTML report and saved markdown transcript.
• Runtime: Approx. 4 minutes.

Highlighted Details

• Five distinct AI advisor personas: Contrarian, First Principles Thinker, Expansionist, Outsider, Executor.
• Anonymous peer review among advisors mitigates individual biases.
• Synthesized verdict includes agreement, clashes, blind spots, recommendation, and next step.
• Outputs visual HTML report and full markdown transcript.

Maintenance & Community

Adapted for Claude Code by @olelehmann. No community channels or roadmap links provided in the README.

Licensing & Compatibility

MIT License, permitting broad use including commercial applications and integration into closed-source projects.

Limitations & Caveats

Not for factual queries, creative generation, or simple summarization. Designed to provide critical feedback and surface downsides; expect potentially unwelcome insights. Effectiveness depends heavily on context richness.