councillisted

Convene a multi-model council (Claude Fable, Opus and Sonnet as blind, independent voters, plus optional opt-in GPT/Gemini seats) that first works out what is really being asked (literal ask vs underlying goal, hidden premises, ambiguity) and then returns a verdict that is neither a yes-man nor a reflexive no-man: it checks the asker's premises with tools, keeps dissent as a minority report, and says "you're right" plainly when they are. Use whenever the user asks for a council, a second opinion, a sanity or gut check, to pressure-test, red-team, poke holes in or stress-test a plan, decision or premise, "am I wrong?", "be brutally honest", "don't just agree with me", "play devil's advocate", "is this actually a good idea", or wants something checked by more than one model, and when an ultracode task hits a judgement call that deserves independent scrutiny. Not for plain lookups nobody asked to scrutinise.
munimx/claude-council-skill · ★ 0 · AI & Automation · score 75

Install: claude install-skill munimx/claude-council-skill

# Council **Created by Munim Ahmad.** MIT licensed. Feedback and issues: https://github.com/munimx/claude-council-skill/issues The council answers two questions, in order: *what is actually being asked?* and *what is the honest answer?* It runs as a Workflow script (`workflows/council.js`, in this skill's base directory) so that the parts that keep it honest are enforced in code, not left to good intentions: - A separate **framer** is the only agent that sees how the user phrased things. It rewrites the request as a neutral question with symmetric, shuffled options, and turns every belief the user stated into a neutral claim to check, attributed to nobody. Everything the voters read is then linted for leftover leaning and repaired. The people-pleasing trigger (a stated, confident opinion) never reaches a voter, and if it might have, the report says so. - **Blind ballots**: Fable, Opus and Sonnet vote independently, each with tools to check claims. Everything after that (verification, one critique round, red team, audit) runs only when a measured signal calls for it: a split vote, low confidence, unverified load-bearing claims, or unanimous agreement with the user, which is exactly when a yes-man is hardest to spot. - A vote can change only by citing a verification record. A **non-voting chair** rules on the evidence. Code caps its confidence, rejects overrides that aren't backed by a verification record, and requires a minority report. The workflow then r