LLM Council: a Claude skill that turns one decision prompt into five advisors and a chairman verdict

The problem is not missing answers, it is missing disagreement

Ask one model a hard question and you get one answer with no calibration. The README frames this bluntly: you get one perspective and no way to tell whether it is good or merely confident. The skill targets that gap by forcing the same prompt through several angles before any synthesis happens. It is aimed at people making business or positioning calls, and the README's own examples are commercial: whether to launch a $97 workshop or a $497 course, which of three positioning angles is strongest, whether to pivot from X to Y, whether to hire a VA or build an automation first. The common thread is a decision where the cost of a bad call is high and the correct answer is not written down anywhere. The README also names the inverse case. Capital-of-France questions, tweet drafting and article summarisation are listed as bad council questions, because one has a single right answer and the other two are creation or processing tasks rather than judgement. That self-limitation is the most useful sentence in the document, and it is the one most likely to be ignored by people who install first and read second.

Five advisors, a peer review pass, and a chairman who reports the disagreements

The mechanism described is a three-stage pipeline. Stage one spins up five advisors, each instructed to think about the question from a fundamentally different angle. Stage two has those advisors review each other's takes. Stage three hands everything to a chairman, who produces the final recommendation and, according to the README, tells you where the advisors agree, where they clash, and what you should do. The output format matters more than the count. A synthesis that only reports the majority view throws away the useful part, and the README explicitly promises the clashes are surfaced. What the README does not do is name the five angles. There is no list of roles, no description of how the angles are chosen, and no explanation of whether the advisors run as separate model calls or as sequential reasoning steps inside one conversation. The methodology is credited to Andrej Karpathy's LLM Council project, and this repository is described as a packaging of that methodology rather than a reimplementation. Anyone who needs to know whether the five opinions are genuinely independent samples or one model role-playing five times has to read SKILL.md itself, because the README does not answer it.

Installation is a file placement problem, not a build

There is no package manager step and no service to start. The README offers two routes. The first is to paste an instruction into a fresh Claude chat telling it to install the skill from the repository URL and to walk you through anything it needs. The second is to open SKILL.md in the repository, use the download button on the file view, save it locally, then paste a message telling Claude you downloaded a file called SKILL.md and asking it to install it. Both routes end with the same artifact in the same place: a SKILL.md file that Claude Code or Claude Cowork reads as a job description. The README is explicit that some setups require manual upload and that Claude will say so rather than failing silently. There are no configuration keys, no environment variables and no CLI flags documented here, which is consistent with the skill being a prompt artifact. The practical consequence is that versioning is manual. Nothing in the material describes an update command, a version field or a migration path, so replacing the skill means replacing the file.

Trigger phrases are the interface, and they are fixed

You do not call a function or pass arguments. You say a phrase in a normal conversation and the skill activates. The README lists five: "council this", "run the council on [your question]", "pressure-test this", "stress-test this", and "war room this". The bracketed form is the only one that carries the question inline; the others rely on context from the surrounding conversation. This is a real constraint. Two of the phrases, "pressure-test this" and "stress-test this", are generic enough that they could plausibly be used by other skills or by the model's own default behaviour, and the README gives no guidance on precedence or on how to tell whether the council actually ran. A user who says "stress-test this" and gets a single flowing answer has no documented way to distinguish a council run whose verdict happens to be concise from a skill that never fired. The mitigation is to use "run the council on ..." and to check that the response separates agreement from disagreement, since that structure is what the README says the chairman produces.

Where the council is the wrong tool, and what to use instead

The README's own bad-question list is the honest limitation: single-answer questions, drafting, and summarisation. Beyond that list, the material does not describe cost, latency or token behaviour, so anyone running this on a long document should expect the pipeline to be heavier than a single prompt without being told by how much. There is also no described mechanism for grounding the advisors in external evidence. They reason over what you paste in. If the decision depends on current pricing, a competitor's changelog or a contract clause, the council will reason about your summary of those facts rather than the facts themselves. For that class of problem, a retrieval approach is the different tool: something like a plain Claude Code session with file access and a web search tool, or a RAG setup that pulls the source documents into context before any reasoning happens. The difference in approach is not quality of reasoning, it is where the evidence enters. The council deliberates over a fixed prompt; a retrieval workflow expands the prompt with retrieved material first. If your question is "is this clause risky", you want the clause in context, not five opinions about your paraphrase of it.

Maintenance, provenance and the missing licence

The repository is a distribution wrapper. The README says as much: the skill was built by Ole Lehmann, the methodology is adapted from Karpathy's LLM Council, and this repo exists to make sharing easier. That framing has two consequences. First, the maintenance surface is small and entirely in SKILL.md. When the underlying model's behaviour shifts, or when the prompt stops producing the promised structure, the fix is an edit to that one file, and there is no upstream sync described that would bring improvements back automatically. Second, attribution obligations sit with the methodology, not the packaging, and the README credits both the original author and the adapter. The material supplied does not state a licence for this repository. No LICENSE file is mentioned, no SPDX identifier appears, and the description field is empty. That is not a legal problem in itself, but it means the default position applies: absent an explicit grant, you do not have clear permission to redistribute or modify. If you plan to fork it, embed it in an internal template, or ship it to a team, resolve the licence question with the repository owner first. If you only install it for personal use, the question is less pressing but still unanswered.

Who should install it, and what to check first

Install it if you already work inside Claude Code or Claude Cowork, you make recurring judgement calls that have no single correct answer, and you want a second through sixth opinion without leaving the chat. The install cost is one file and the trigger vocabulary is five phrases, so the downside of trying it is small. Do not install it if your questions are lookups, drafts or summaries, if you need the advisors grounded in documents you have not pasted in, or if you need to know exactly how many model calls each council run costs. The README does not answer that last question and the repository does not appear to contain anything that would. Two checks before you rely on it. Open SKILL.md and confirm the five advisor angles and the peer review step are specified in the file, because the README describes the behaviour without documenting the instructions that produce it. Then check whether the repository carries a licence file, and if it does not, decide whether an unlicensed prompt artifact is something you are comfortable depending on.

Editorial conclusion

Adopt this skill if you routinely face judgement calls with no single correct answer and you already pay for Claude Code or claude.ai, since installation is a single SKILL.md file and the trigger phrases are fixed. Do not adopt it for retrieval questions, drafting tasks or summarisation, because the README states those are the wrong shape for the council and the skill will not add signal. Before relying on it, open SKILL.md and confirm the five advisor roles and the peer review step are actually defined in the file rather than implied by the README, and check the repository for a licence file, since none is stated.

Official sources

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