council

# Council ## Overview Claude is agreeable. Ask "should I do X?" and it finds reasons for X. Ask "is X a bad idea?" and it finds reasons against. Same decision, different framing, opposite answers. That's fine for drafting prose. It's dangerous for engineering decisions with stakes. The council fixes this through structured adversarial review: 1. **Frame** the question with workspace context. 2. **Convene** 5 advisors in parallel — each with a fundamentally different thinking style. 3. **Anonymize** responses and run a peer review round (no advisor knows who said what). 4. **Synthesize** into a chairman verdict that names where the council agrees, where it clashes, blind spots only the peer review surfaced, and one concrete next step. 5. **Persist** the full transcript to `.devt/state/` for downstream workflows. Adapted from [Karpathy's LLM Council](https://github.com/karpathy/llm-council). Persona-based Claude Code adaptation pioneered by [@tenfoldmarc](https://github.com/tenfoldmarc/llm-council-skill); this devt version retunes the personas for engineering decisions, auto-pulls `.devt/rules/` context, writes markdown to `.devt/state/`, and supports optional model diversity. ## When To Run It **Run the council on:** - Architecture choices: "monolith vs services", "Postgres vs Mongo for this workload" - Refactor strategies: "rewrite vs strangle", "is this abstraction worth introducing" - API design: "REST vs GraphQL", "synchronous request vs event-driven" - Library/f...