Agent skill · research science · vishalsachdev
llm-council
Convene a 3-model council (Claude + GPT via codex CLI + Gemini CLI) on a high-stakes decision. Forces cross-critique between members and surfaces where they actually disagree, breaking Claude's default agreeableness. Use when the user asks to "convene a council", "get a second opinion", "ask GPT and Gemini", "what would other models say", or has an architecture / strategy / hiring / pricing decision where being wrong is expensive. Skip for factual questions, code with one right answer, or anything premortem-shaped (route to premortem skill instead).
Why this skill is useful
Adds a structured multi-model decision-making process that includes executable CLI commands for cross-critique and synthesis, enhancing the AI's capability to handle complex strategic decisions.
What it needs
Requires bash installed locally. About 6k tokens when loaded. Last updated 2026-08-01. 4 stars on the source repository.
What this skill does
LLM Council Three frontier models, one decision. Output is not three answers side-by-side — it's a structured map of agreement, disagreement, and which disagreements matter for this user's call. Members Member Invocation Bias to expect --- --- --- Claude (this session) direct reasoning Pragmatic, agreeable, code-rooted GPT-5.5 codex exec Skeptical, hedge-prone, edge-case sensitive Gemini gemini -p Structural, taxonomical, categorization-heavy When to run Good fit: Architecture/design with long-lived consequences Strategy calls (pricing, positioning, hiring) where Claude's agreeableness is a real risk "Am I thinking about this right?" — challenge the framing, not the plan Any decision the user explicitly flags as high-stakes Bad fit: Factual lookups Code with one right answer Premortem-shaped questions ("what could go wrong") — route to premortem Speed > depth CLI invocations (verified) Codex (GPT-5.5): --skip-git-repo-check is required outside trusted dirs -o <file> writes just the final message (avoids verbose preamble in stdout) Gemini: Outputs the answer directly to stdout Add -m <model> to pin a model Run in parallel via two simultaneous Bash tool calls in a single message. Don't sequence them. Process (3 rounds) Round 1 — Independent answers Send the question to all three members in parallel. None see the others' answers. Each member gets the same prompt: the user's question + relevant context + this directive: "Answer in under 250 words. Be direct. State your strongest position, not a hedged one. End with the single biggest risk you see in your own answer." The "biggest risk in your own answer" line is load-bearing — it primes the model to flag its own weakness, which makes round 2 more productive. Round 2 — Cross-critique (anonymized) Send each member the other two answers, with peer identities anonymized as "Model A" and "Model B". Each model gets a different random A/B mapping so it can't infer who's who across runs. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills vishalsachdev/llm-council