llm-council

What This Skill Does

The llm-council skill is a sophisticated orchestration tool designed to minimize bias and maximize technical rigor in AI-generated implementation plans. Instead of relying on a single large language model, this skill spins up a configurable, multi-member planning council—comprised of agents like Codex, Claude Code, Gemini, or OpenCode—to brainstorm solutions independently. By anonymizing and randomizing the outputs before subjecting them to a judge agent, the skill ensures that the final plan is a synthesis of the strongest ideas rather than the output of a single entity. It features built-in retry logic, structured Markdown logging, and comprehensive auditability for complex software development tasks.

Installation

To integrate this skill, run the following command in your terminal: clawhub install openclaw/skills/skills/am-will/llm-council

Ensure you have configured your environment by executing ./setup.sh to populate your $XDG_CONFIG_HOME/llm-council/agents.json file. You can adjust your agent preferences at any time by running python3 scripts/llm_council.py configure.

Use Cases

• Complex Refactoring: Use the council to plan major architectural changes where multiple design patterns need to be weighed against one another.
• Unbiased System Architecture: Deploy when you need to avoid the 'echo chamber' effect of using just one model for high-stakes technical designs.
• Audit-Heavy Projects: Perfect for team environments where every step of the planning process must be documented, timestamped, and saved in a reproducible Markdown format.
• Risk Mitigation: Ideal for mission-critical tasks where failure recovery and structured, iterative improvement are required.

Example Prompts

1. "I need to rewrite our authentication module from OAuth to OIDC. Please start the llm-council planning process for this migration."
2. "Use the llm-council skill to draft a plan for containerizing our legacy monolith; make sure to include constraints about zero-downtime deployment."
3. "Launch a new planning session for the database schema update. I'm ready to answer your intake questions to refine the requirements."

Tips & Limitations

• Patience is Key: The council workflow can take significant time. Do not interrupt the session; the 30-minute window ensures the agents have time to deliberate properly.
• Intake Questions: Always prioritize the initial intake questions. Providing granular, specific constraints during this phase significantly increases the quality of the final plan generated by the council.
• Session Integrity: Never yield or terminate the session until the final-plan.md has been successfully confirmed; exiting prematurely will cause the process to fail.
• Configuration: Always verify your agents.json configuration before starting a large run to ensure you are using the models most appropriate for your specific technical stack.