llm-councillisted
Install: claude install-skill ssubbotin/kaggle-research-skills
# LLM Council
## Core principle
One model answering alone has one set of blind spots. A **council** dispatches the same question to several independent models, has them **anonymously peer-review and rank** each other's answers, then a **chairman** synthesizes a single final response from the collected answers and rankings.
The value comes from three things a single call can't give you: **independence** (answers formed without seeing each other), **anonymized cross-critique** (each model judges the substance, not the brand), and **synthesis** (the final answer is built from the best of all, not any one).
Source pattern: Andrej Karpathy's `llm-council` (github.com/karpathy/llm-council) — a local app with the identical three-stage flow. This skill adapts it to run inside Claude Code using real agents and real model diversity.
## When to use / when NOT to use
**Use for:** hard technical judgment calls, design trade-offs with no clear winner, factual questions where being wrong is costly, "which of these approaches is best", forecasts/estimates, anything where you'd want a second and third opinion. Also whenever the user explicitly asks for a council/panel/multiple models.
**Do NOT use for:** simple lookups, mechanical edits, anything with one correct answer you can verify directly, or latency/cost-sensitive tasks. A council is 3–5× the tokens and wall-clock of a single answer — spend it only when the extra reliability is worth it.
## The three stages (do not skip or merge