LLM Council: A Claude Skill for Multi-Perspective Decision Reviews

The Problem: One AI Answer Has No Dissenting Views

When you ask a single AI model a decision question, you get one answer shaped by that model's training, its current context, and whatever framing you gave it. You have no way to know whether a different framing would have produced a different recommendation, or whether the answer reflects a systematic bias in the model's training data.

LLM Council addresses this by spinning up five separate advisor roles within a single Claude session, each instructed to analyze the problem from a fundamentally different angle. They review each other's reasoning rather than working in isolation. A chairman persona synthesizes the results into a final recommendation that explicitly notes where advisors agreed, where they clashed, and what to actually do.

The project is built by Ole Lehmann and adapts the methodology from Andrej Karpathy's LLM Council project. The README describes the chain: you ask a hard question, five advisors each give their take, they peer-review each other, and a chairman synthesizes everything into a final recommendation with a clear view of the disagreements.

What the Five Advisors and Peer Review Actually Produce

The README gives specific examples of questions the council is designed for: choosing between a workshop and a course at different price points, evaluating positioning angles for a product, deciding whether to pivot from one business model to another, getting a critique of landing page copy, and deciding whether to hire a person or build automation first.

These examples share a structure: they involve a binary or multi-option choice, the cost of getting it wrong is meaningful, and different mental models produce genuinely different recommendations. The README calls this the category of questions where there is genuine uncertainty and the cost of a bad call is high.

The README is equally explicit about what the council is not for. Factual questions with a known answer do not benefit from five perspectives. Creation tasks, like writing a tweet, produce noise rather than insight from this process. Summarization tasks are not judgment problems. The README summarizes this as: if you already know the answer and just want validation, the council will likely tell you things you do not want to hear.

After the peer review step, the chairman's verdict names where advisors agreed and where they diverged. This structure surfaces the specific disagreements rather than averaging them out into a vague recommendation.

Installing the Skill: Two Paths, No Terminal Required

The README documents two installation methods, both designed to avoid command-line interaction.

The first path is to open Claude and paste this instruction directly: ask Claude to install the skill from the GitHub repository at github.com/aiwithremy/claude-skills-llm-council, with the SKILL.md file, and walk you through anything it needs. Claude fetches the file and places it in the correct location. If the setup requires a manual step due to the local configuration, Claude describes exactly what to do.

The second path is to download SKILL.md directly from the repository by clicking the file and using the download button, then ask Claude to install it by pasting a short instruction. Claude guides the remaining steps.

After installation, triggering the council in any Claude conversation requires one of five phrases documented in the README: 'council this', 'run the council on [your question]', 'pressure-test this', 'stress-test this', or 'war room this'. Claude spins up the five advisors, runs the peer review, and delivers the chairman's verdict.

The skill works in Claude Code and Claude Cowork, which is the claude.ai interface. The README does not document installation for tools outside these two environments.

The Methodology: Adapted from Andrej Karpathy's LLM Council

The SKILL.md file contains the instructions that define the five advisor roles and the chairman persona. The README credits the methodology to Andrej Karpathy's original LLM Council project and describes this repository as making that methodology easier to use for people who want to share it with friends.

Karpathy's original LLM Council is a Python script that runs multiple LLM calls in a sequence to simulate a council deliberation. That approach requires a Python environment, API keys, and familiarity with running scripts from a terminal. This repository packages the same concept as a skill file that Claude Code or Claude Cowork can load, removing all technical setup for users who already have Claude access.

The trade-off is that this implementation runs everything within a single Claude session rather than making independent API calls to multiple distinct model providers. The five advisor perspectives are generated by the same underlying model. Whether this produces meaningfully different perspectives than asking Claude once with a broader prompt depends on how the SKILL.md instructions structure the advisor roles, which is the content of the SKILL.md file itself.

Practical Limits on What the Council Delivers

Several constraints shape what this tool can and cannot produce. First, all five advisors and the chairman run within a single Claude session. The diversity of perspective depends entirely on how well the SKILL.md instructions force the model into genuinely distinct reasoning modes. There is no structural guarantee that advisor three and advisor five will disagree if the question has a widely agreed-upon answer.

Second, the README's peer review step means that later advisors can see earlier advisors' reasoning before giving their own. This sequential structure can cause later advisors to anchor on earlier positions rather than developing independent analyses. The README does not address this.

Third, the council produces no code, no files, and no verification of claims. Its output is reasoning and recommendations in natural language. If the underlying question depends on data that was not provided in the prompt, the advisors speculate. The README addresses this implicitly by restricting the recommended use to decision questions with clear framing rather than open research questions.

Finally, as a skill file rather than a software package, there is no test suite and no programmatic way to evaluate whether the council's recommendations are calibrated.

Repository Status: Two Files, No Dependencies

The repository contains exactly two files: README.md and SKILL.md. There are no build files, no package.json, no requirements.txt, and no CI configuration. The license is listed as unknown, which means GitHub's automated license detection found no recognized license file. The README does not address licensing or terms of use for the skill.

The last push to the repository was on 2026-04-26. There are no GitHub releases. The two-file structure means updates would appear as changes to the README or SKILL.md contents directly.

For users evaluating the skill, the most relevant artifact is SKILL.md itself, which contains the full set of advisor and chairman instructions. The README serves as a non-technical user guide explaining the methodology and installation process for users who are not familiar with Claude Code skills.

Editorial conclusion

Anyone with an existing Claude subscription who faces a decision with genuine uncertainty and high cost of failure can install this skill in one paste command. The README explicitly says the council will tell you things you do not want to hear if you already know the answer; that is the point. The repository contains only README.md and SKILL.md, with no build steps and no dependencies. The last push was on 2026-04-26. If the SKILL.md file is what you need, it is there now and not likely to change.

Frequently asked questions

Official sources

1. aiwithremy/claude-skills-llm-council on GitHub
2. Issues
3. README