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 ~110-line cap.

When to run

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).

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 → references/advisors.md (read before step 2).

Tier

Dispatched roles — advisors (step 2), peer reviewers (step 3), the research pass (step 1C) — run on cheap-tier sonnet sub-agents (Agent-tool model: override); only the chairman synthesis (step 4) stays on the session model (the high-judgment step). Tier map → .claude/CONTEXT.md.

The 6 steps

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

1. Frame (+ context & research). Scan the workspace for context (CLAUDE.md · memory/ · referenced files — ≤30s). Decide if the call turns on external current facts; if so, run one shared research pass (template in prompts.md) and carry an evidence brief — say so in one line if you skip it. Reframe the raw question into one neutral prompt every advisor receives. Too vague → ask one clarifying question, then proceed.
2. Convene (5 advisors, parallel). Spawn all 5 as sub-agents (identities → advisors.md; template → prompts.md), 150–300 words each, leaning fully into their lens. Each first emits its 1–2 decision-critical questions, then its analysis (a question no advisor resolves becomes a chairman blind-spot). Exception: the Outsider does NOT get the evidence brief — keep it naive (that's the curse-of-knowledge detector).
3. Peer review (5, parallel). Anonymize the responses as A–E, randomizing the mapping independently per reviewer (not once). Each reviewer scores on a rubric — reasoning · evidence · coverage, NOT length/fluency — then answers: strongest? biggest blind spot? what did ALL miss? (template → prompts.md).
4. Chairman synthesis (session model). One agent gets the question + de-anonymized responses + all 5 reviews + the advisor questions → the verdict (template → prompts.md): agree · clash · blind spots · recommendation (written before the score) · pre-mortem · calibrated confidence + dissent · one-thing-first · the single-model ceiling caveat. The chairman MAY side with a lone dissenter if the reasoning is strongest.
5. Write the verdict (lean). Present in chat AND write only the verdict (sections per prompts.md) to verdict-.md — use the slug the ADR will use. Default to the OS temp dir; repo only if asked. Never dump the full transcript.
6. Feed forward. The verdict is a decision input, not a record. Hard-to-reverse + surprising + a real trade-off → fold the recommendation + alternatives into an ADR (docs/adr/, DOCS_Guide §4), then the verdict file can be deleted. Don't accumulate stale verdicts.

Conditional passes (gated — off by default)

Two optional passes fire only when warranted, keeping the base run at ~11 calls (templates → prompts.md). Say in one line whether each fired or was skipped.

• Moderator (unknown-unknowns) — after step 3, before the chairman. Fires when the panel converges fast (groupthink risk). One cheap sub-agent surfaces the most important consideration no advisor/reviewer raised → feeds the chairman as an extra blind spot. +1 call.
• Fact-verify (adversarial) — after step 4. Fires only when the verdict rests on external facts / citations / benchmarks — skip for pure judgment forks (the common case). One refuter per claim cluster (~1–4) extracts load-bearing claims + verifies cited sources (a citation resolving to no real source = FALSE); the chairman corrects/demotes before finalizing and notes it in Confidence & Dissent. +1–4 calls.

Worked example → references/example.md.

Red flags

❌ Spawning advisors sequentially — always parallel; sequential lets earlier responses bleed in. ❌ Skipping anonymization for peer review — reviewers defer to thinking styles instead of judging merit. ❌ Forcing the chairman to the majority — it may side with the strongest reasoning, even a lone dissenter. ❌ Counciling a trivial question — one right answer → just answer it; the council is for genuine uncertainty. ❌ Dumping the full transcript into the verdict file — write only the verdict sections; the sprawl is what makes it un-lean. ❌ Laundering disagreement into false consensus — low advisor agreement is reported with a lowered confidence, never smoothed into a confident verdict. ❌ Over-trusting unanimity — 5 personas on one model share its knowledge gaps; the verdict says so (the ceiling caveat), so agreement isn't read as independent proof. ❌ Running the gated passes on a pure-judgment fork — fact-verify fires only when the verdict rests on external facts; the moderator only when groupthink risk is real. Default off; never a reflex.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: aldianriski
• Source: aldianriski/lean-flow
• License: MIT

Install and usage instructions live in the source repository linked above.