council

LLM Council

Ask one AI, get one answer — and you can't tell if it's great or mid, because you only saw one perspective. The council runs your question through 5 independent advisors (different thinking lenses), has them peer-review each other anonymously, then a chairman synthesizes a final recommendation: where they agree, where they clash, and what to actually do. Adapted from Andrej Karpathy's LLM Council, run inside Claude via sub-agents with different lenses instead of different models.

Cap (ADR-006): this SKILL is the procedure only. The executable artifacts — advisor definitions, prompt templates, the worked example — live in references/ and are read on demand, so they don't count toward the ~140-line cap.

When to invoke

For decisions where being wrong is expensive and there's genuine uncertainty — pricing, positioning, pivots, hard architectural / scope forks, "am I crazy to do X?". NOT for one-right-answer questions, creation tasks, or processing/summarizing tasks. If you already know the answer and just want validation, expect the council to tell you things you don't want to hear — that's the point. Triggers are in the description; reserve it (~11 model calls/run — each spawn re-pays the full base substrate, so cost ≈ 11 × context, not 11 small calls; ADR-010 cost term).

The five advisors

Five thinking styles chosen for the tensions they create — Contrarian ↔ Expansionist (downside ↔ upside), First Principles ↔ Executor (rethink ↔ ship), with the Outsider in the middle keeping everyone honest. Full definitions → ${CLAUDE_SKILL_DIR}/references/advisors.md (read before step 2).

Tier

Dispatched roles — advisors (step 2), peer reviewers (step 3), the research pass (step 1C) — run on the execution role (sonnet default); only the chairman synthesis (step 4) runs on the decision role (session model — the high-judgment step). Role map → .claude/CONTEXT.md (ADR-010).

The 6 steps

All sub-agent prompt templates are in ${CLAUDE_SKILL_DIR}/references/prompts.md — read it before spawning.