council

You ask one model a question and you get one answer with one blind spot — and no way to tell what it missed. The council skill runs the same decision through five independent thinking lenses, has them peer-review each other anonymously, then a chairman synthesises a single verdict. Adapted from Andrej Karpathy’s LLM Council.

What it solves

A single perspective is a single blind spot, and on a high-stakes call the cost of that blind spot is real. Asking the same model again does not help — it tends to agree with itself. The council manufactures genuine disagreement: five lenses chosen for built-in tension (downside vs upside, rethink-everything vs just-ship, plus an outsider keeping everyone honest), so the decision gets attacked from angles a single pass never reaches. The anonymous peer-review round is what makes it more than “ask five times” — reviewers judge the arguments on merit, not on which lens produced them.

How it works

Four phases, each read-only over the workspace:

1. Frame — scan the workspace for context (CLAUDE.md, a memory/ directory, referenced files, and CONTEXT.md / docs/adr/ when present), then reframe the raw question into one neutral prompt every advisor receives. No opinion is injected; if the question is too vague, the skill asks exactly one clarifying question first.
2. Convene — the five advisors (Contrarian, First Principles, Expansionist, Outsider, Executor) answer in parallel, each leaning fully into its lens.
3. Anonymous peer-review — the responses are anonymised as A–E with a randomised mapping, then each advisor reviews the set: strongest response, biggest blind spot, what everyone missed.
4. Chairman synthesis — one agent produces a fixed-structure verdict: where the council agrees, where it clashes, the blind spots peer-review caught, the recommendation (a real answer, not “it depends”), and the one thing to do first.

The advisors are ephemeral prompt-injected lenses, not roles — the council creates no role file and reuses none (unlike delegate, which dispatches persisted roles). The verdict is presented in chat as markdown; the council writes no files by default — a transcript is opt-in.

Pre-mortem mode

--pre-mortem answers a different question with the same machinery: instead of weighing an open decision, it treats a plan that has already been decided as already failed and asks the council to explain why. Phase 1 reframes the question as a post-failure autopsy — the plan shipped, six months passed, it failed — and the failure is stated as settled fact rather than a risk, because a hypothetical failure only ever draws hypothetical answers.

Phases 2 and 3 do not change, and no sixth lens is added. That is the point: each existing lens already fails a plan its own way — the Contrarian finds the fatal flaw, First Principles finds the wrong problem solved, the Expansionist finds the window that closed, the Outsider finds the proposition nobody understood, the Executor finds the first step that never shipped. Five structurally different failure modes rather than five phrasings of one.

Phase 4 returns them ranked by likelihood × impact, each carrying a mitigation and a tripwire — the early signal that this mode is materialising, concrete enough to notice (“week 3 and integration tests still aren’t green”, not “watch for delays”). A mode without a tripwire is a worry, not a finding.

When to use

Convene the council for a genuine decision with stakes and a tradeoff — “launch X or Y first”, “is this pivot sane”, “what’s weak in this plan”. Do not convene for a factual lookup, a creation task (“write me a tweet”), or a trivial yes/no with no real tradeoff — just answer those directly. Run interview first when the decision is still fuzzy and needs scoping; reach for delegate when you need persisted specialist roles to do work rather than five lenses to weigh one call.