LLM Council (Claude Council)
Claude Code skill that convenes a council of several personas (Architect, Skeptic, Pragmatist, Critic) on an ambiguous decision, each argues their case, then they peer-review one another before producing a synthesis. Instead of a smooth single answer, you get a structured debate that surfaces blind spots and disagreements. Inspired by Andrej Karpathy's LLM Council. It's a community pattern with several implementations: council-review (ngmeyer), claude-skills-llm-council (aiwithremy), design-council (sjsyrek).
Strengths
- Breaks single-model sycophancy bias: personas contradict each other instead of validating you
- Internal peer-review surfaces blind spots before you commit to the choice
- Open-source, inspired by Karpathy's LLM Council: a documented, re-implementable pattern
Limitations
- Community multi-implementation pattern: no canonical official version, quality varies by repo
- Overkill and token-heavy for a trivial decision: reserve it for real arbitrations
- The implementation cited here, ngmeyer/council-review, is archived on GitHub with no push since May 2026: still cloneable, but frozen. The pattern lives on elsewhere (aiwithremy, sjsyrek)
Best for
- Tech leads facing an architecture trade-off where a single opinion isn't enough to decide calmly
- PM/PO wanting to pressure-test an ambiguous product decision from several angles before committing
- Solo devs without a peer on hand to challenge a risky choice