Is the LLM Council skill official Anthropic software?

No. It is a community skill built by Ole Lehmann, adapted from Andrej Karpathy's LLM Council methodology, and published on GitHub by a third account (aiwithremy). Anthropic has no involvement.

Does it actually call GPT, Gemini, or other models?

No. All five advisors are Claude sub-agents given different thinking-style prompts, so the diversity is in the prompting, not the models. If you want a genuine multi-model council, hex/claude-council or gcpdev/llm-council-skill query external models but require their API keys.

What does a council run cost?

Around 11 agent invocations per question — five advisors, five peer reviewers, and a chairman — so expect several times the tokens and latency of a normal answer. The skill's own docs say to reserve it for decisions where being wrong is expensive, not factual or creative tasks.

How do I trigger it once installed?

Say 'council this', 'run the council on X', 'pressure-test this', 'stress-test this', 'war room this', or 'debate this' in any conversation. The skill also self-triggers on genuine either-or decisions with stakes, and is instructed not to fire on trivial questions.

Skill review · by Ole Lehmann (repo published by GitHub user aiwithremy)

Claude Council Skill

The Claude council skill that search traffic points to is LLM Council, a single-file skill built by Ole Lehmann and published at aiwithremy/claude-skills-llm-council on GitHub, where it holds 1,791 stars. It adapts Andrej Karpathy's LLM Council method to run entirely inside Claude: your question goes to five parallel sub-agent advisors — the Contrarian, the First Principles Thinker, the Expansionist, the Outsider, and the Executor — who answer independently, peer-review each other anonymously, then hand everything to a chairman that writes a verdict with one recommendation and one next step. The thing to know before installing: despite the name, no other LLMs are involved. All five advisors are Claude prompted with different thinking styles, so you get prompt diversity, not the model diversity Karpathy's original used. It is unofficial, ships no license, and the repo has not been touched since the day it went up in April 2026 — one Markdown file, so little to rot, but nothing maintained either.

View the sourcelicense unclear · 1,791★ · facts checked September 1, 2026

By Robert MacKelfresh

Founder, Automation Squad · September 1, 2026

Install

1Claude Code (per-user): run mkdir -p ~/.claude/skills/llm-council then curl -L https://raw.githubusercontent.com/aiwithremy/claude-skills-llm-council/main/SKILL.md -o ~/.claude/skills/llm-council/SKILL.md
2Project-scoped alternative: put the same file at .claude/skills/llm-council/SKILL.md inside the repo you want it available in.
3Start a new Claude Code session and say 'council this: <your decision>' — the skill's trigger phrases handle invocation; there are no dependencies or API keys.
4Claude.ai / Cowork: zip a folder containing SKILL.md and upload it under Settings > Capabilities > Skills; the README confirms the skill works on both surfaces.
5Lazy route (the README's own suggestion): paste the repo URL into Claude and ask it to fetch SKILL.md and install it for you.

What it does

Five advisors in parallel: Spawns five sub-agents simultaneously, each locked to one thinking style — Contrarian, First Principles Thinker, Expansionist, Outsider, Executor — and each instructed to lean fully into its angle, 150-300 words, no hedging.
Anonymous peer review: Advisor responses are anonymized as Response A-E with randomized mapping, then five reviewer sub-agents each name the strongest response, the biggest blind spot, and what all five missed — the step the skill calls the core of Karpathy's insight.
Chairman synthesis: A final agent de-anonymizes everything and writes a fixed-format verdict: where the council agrees, where it clashes, blind spots caught in review, one clear recommendation, and one concrete first step — explicitly allowed to side with a lone dissenter.
Workspace context enrichment: Before framing the question it scans your setup — CLAUDE.md, memory/ folders, attached files, prior council transcripts — for the two or three files that ground the advice, capped at about 30 seconds of looking.
Trigger phrases with guardrails: Fires on 'council this', 'run the council', 'pressure-test this', 'war room this', 'stress-test this', 'debate this'; the description explicitly blocks trivial yes/no and factual questions from convening a council.

Common questions

Is the LLM Council skill official Anthropic software?: No. It is a community skill built by Ole Lehmann, adapted from Andrej Karpathy's LLM Council methodology, and published on GitHub by a third account (aiwithremy). Anthropic has no involvement.
Does it actually call GPT, Gemini, or other models?: No. All five advisors are Claude sub-agents given different thinking-style prompts, so the diversity is in the prompting, not the models. If you want a genuine multi-model council, hex/claude-council or gcpdev/llm-council-skill query external models but require their API keys.
What does a council run cost?: Around 11 agent invocations per question — five advisors, five peer reviewers, and a chairman — so expect several times the tokens and latency of a normal answer. The skill's own docs say to reserve it for decisions where being wrong is expensive, not factual or creative tasks.
How do I trigger it once installed?: Say 'council this', 'run the council on X', 'pressure-test this', 'stress-test this', 'war room this', or 'debate this' in any conversation. The skill also self-triggers on genuine either-or decisions with stakes, and is instructed not to fire on trivial questions.

Before you install

Unofficial. This is not an Anthropic skill; it is a personal project shared via an X article, and the GitHub repo is a courtesy mirror by a third party (aiwithremy) of Ole Lehmann's file, not a maintained open-source project.
Effectively frozen: created 2026-04-25, last pushed 2026-04-26, untouched for four months as of this review. One Markdown file rots slowly, but there is no issue triage and no updates.
The SKILL.md contradicts itself: step 5 says do NOT generate an HTML report, while the closing notes still say the visual report matters and to make the HTML output clean — leftover text from an earlier version. The optional transcript step also writes to an active/ directory most projects do not have.
Each run costs roughly 11 sub-agent invocations (5 advisors, 5 reviewers, 1 chairman), so a council answer consumes several times the tokens and wall-clock time of a normal reply.
The name oversells the mechanism: no external LLMs are queried. All five advisors are the same Claude model with different prompts, so correlated blind spots are still possible — Karpathy's original polled genuinely different models.
Name collision risk: hex/claude-council is a different, unrelated project (a plugin that queries Gemini, OpenAI, Grok, Perplexity, and Kimi side by side). Check which one you are installing.

Facts come from the skill's own repository and docs on the date above, via the GitHub API — not from memory, and never from a sponsor.

Worth comparing

hex/claude-council — Claude Code plugin that asks Gemini, OpenAI, Grok, Perplexity, Kimi, and local ollama the same question side by side — real multi-model, MIT-licensed, 682 stars, but needs API keys or CLI subscriptions.
gcpdev/llm-council-skill — Claude Code skill that brainstorms with ChatGPT and Gemini before presenting an implementation plan — 425 stars, MIT, requires access to those external models.
karpathy/llm-council — Andrej Karpathy's original — a standalone web app that dispatches one query to multiple models with anonymous peer review; the methodology this skill borrows, not a Claude skill itself.

← All MCP servers & skills