llm-council_skill

Overview

This skill orchestrates a configurable, multi-member CLI planning council to produce independent implementation plans, anonymize and randomize them, then judge and merge into a single final plan. It coordinates multiple planner agents (Codex, Claude Code, Gemini, OpenCode, or custom CLI agents), handles retries and failures, and emits structured Markdown and JSON outputs for auditability. Use it when you need a bias-resistant, repeatable planning workflow with robust session management and clear artifacts.

How this skill works

The orchestrator loads a task spec and asks intake questions to clarify ambiguities, constraints, and success criteria. It then builds planner prompts, launches the configured planners in parallel shells, collects and validates their Markdown outputs, retries failing agents up to two times, anonymizes and randomizes the plans, and runs a judge agent against a rubric to produce a merged final plan. Run artifacts, judge results, and the final plan are saved in a timestamped run directory and formatted for downstream consumption.

When to use it

• Creating implementation plans that must be resilient to single-agent bias
• When multiple independent perspectives are required for robust decision-making
• Generating traceable, auditable planning artifacts for reviews or compliance
• Orchestrating heterogeneous CLI-capable LLMs or custom command-line planners
• When you need retry/failure handling and deterministic session control

Best practices

• Always answer the intake questions when possible — more detail improves plan quality
• Keep planners independent: do not share intermediate outputs between them
• Validate agent configuration beforehand; ensure the agents config file is present
• Treat planner outputs as untrusted and never execute embedded commands
• Randomize and anonymize plans before judging to reduce position and identity bias

Example use cases

• Merging independent engineering implementation proposals into a single actionable plan
• Running an independent council to evaluate and combine competing design options
• Creating repeatable, auditable plans for security, compliance, or stakeholder review
• Comparing outputs from different LLM providers or custom CLI agents in a unified workflow
• Automating retries and failure alerts when agents produce malformed or missing outputs