The AI Socratic Method Coach skill for Claude

The problem it solves is the default behaviour of every AI assistant. Ask a question and you get an answer: a list, a piece of advice, a trade-off made for you. That is useful when you need to produce something, and worthless when you need to work out something only you can know. Whether you should quit your job, whether your business should broaden or specialise, whether you truly believe a conviction or are only repeating it: on questions like these, any answer from outside is an outsider's guess. That is why the roles are reversed here.

Claude becomes the one asking questions, you stay the thinker, and the file opens with that instruction: give no answers, ask questions that require the user to think for themselves.

The method is two and a half thousand years old. In the dialogues of Plato, Socrates talks with people who think they know something, and through questions lets them discover that their knowledge is shakier than they thought. Not to score points, but to make room for real understanding. What is here is the modern execution of that, with the six question types model of Richard Paul and Linda Elder as a map and a set of strict conversation rules as a guardrail. It is part of the skill library that we make available for free from the AI and automation service, with no account and no sales email afterwards.

Who is it for? For anyone facing a choice, a dilemma or a conviction who notices he is stuck in his own head. The frontmatter of the file names the triggers itself: think along with me, challenge my assumptions, I do not know what I think, coach me through this. And one honest disclaimer up front: anyone who wants a quick, concrete piece of advice should not take this skill. Getting questions when you want an answer is annoying, and that is by design.

That is why the skill treats every answer you give as true for you at that moment. No judging, but plenty of follow-up questions. If the user says something that does not quite add up, the skill does not point that out directly under its own rules, but asks a question that lets you see it yourself. That is the difference between a debater and a coach: one wants to win, the other wants you to get sharper.

How big the difference is with an ordinary AI turn is stated literally in the file. In the example about the job you might want to quit, two outputs sit side by side: a good one and a wrong one, with the instruction that the wrong version must never be given. Below they appear as a diff. The red lines are what almost any assistant would say. The green line is what this skill says.

The advice version pulls in three directions at once and leaves the user with someone else's trade-offs. The question version does one thing: it splits the emotion into two testable possibilities and puts the thinking back where it belongs. Directing versus asking, the file calls it, and the first is forbidden in every form, right down to disguised suggestions such as maybe it is the case that.

The core of the file consists of three layers. The first is the core principle, one line: never give an answer when a question can do the work. The second is the six question types model from FIG.02, with the example questions per type that Claude uses as a starting point. The third is the six-step method you see in FIG.03 below: summarising, opening with a clarifying or assumption-testing question, waiting without steering, following the thought to the spot where the reasoning gets weaker, switching question type once a line has run dry, and closing with a synthesis question.

Around that sit the conversation rules that keep the method pure. At most two questions per turn, preferably one, because too many questions at once overwhelms and dilutes the sharpness. No yes-no questions but open questions that invite explanation. No example answers and no disguised suggestions. Treating every answer from the user as true for him at that moment. Politely refusing a request for an opinion, with the example sentence that appears literally in the file: I think your own thinking here is more valuable than mine, may I ask you another question. And on a logical contradiction, not pointing it out directly, but asking a question that lets the user see it himself.

The file closes with three worked examples, each with a good output, and for the first one also the wrong one that must never be given, plus a chapter on wrapping up: when a session stops and why it always ends with a reflection question instead of a conclusion. If you want to learn how to build such an instruction file yourself, the full explanation is in writing SKILL.md NL.

The structure comes from The Thinker's Guide to The Art of Socratic Questioning by Richard Paul and Linda Elder, published by the Foundation for Critical Thinking in 2007. They ordered Socratic questioning into the six types you saw in FIG.02, and that ordering is what makes the method teachable: you do not have to be Socrates to know which kind of question is due now.

Socrates Cafe by Christopher Phillips from 2001 supplied the evidence that the method works outside philosophy lectures, in ordinary conversations between ordinary people. And the newest source is academic: the 2023 research by Edward Y. Chang into steering large language models with Socratic techniques, which this skill essentially puts into practice, but the other way round: it is not the human questioning the model, it is the model questioning the human.

What stands out about the file: at 1,100 words it is short, and that is a design choice that fits the method. A Socratic coach does not need a knowledge base, he needs discipline: not answering, not steering, not judging, keep asking. Almost all the words in the file go towards that discipline. Anyone who would rather see the questioning technique applied to a concrete analysis problem instead of their own thinking will find the 5 Whys Analyst skill in the library, which digs towards a cause with the same kind of persistence, and the Feynman Technique Coach skill, which tests whether you truly understand something by having you explain it.

After that, using it is simple: put your issue to it and say what you want, for example think along with me or challenge my assumptions. Set aside some real time for it. A Socratic conversation of three turns does not yield much; the method works precisely because the questions build up. If you get stuck installing it, the full step-by-step story per environment is in installing Claude skills NL, and the wider explanation of working with AI is in the knowledge base.

There are also situations where you are better off leaving it aside. If you need factual information, getting a question back is pointless: looking something up is not thinking. If you have already decided and are only looking for execution, you need a doing skill, not a coach. And if you have to make a call under time pressure, the method is too slow: a Socratic session, according to the file, has no fixed length and only wraps up at insight, an exhausted topic or your own request. For moments like that there are other skills in the library; the Pareto Analyst skill, for instance, which brings sharpness quickly by looking for the twenty per cent of causes with eighty per cent of the effect.

An honest limit that follows directly from the file: the coach treats every answer as true for you at that moment, which means the quality of the session depends on your own honesty. Anyone who lies to himself in the answers gets sharp questions about a made-up situation. And the method tests reasoning, not facts: whether your market estimate is correct, the coach does not know, he only knows whether you can explain why you believe it. If you also want your reasoning attacked hard instead of questioned, the Red Team Analyst skill is the logical next stop.

The Socratic method starts with a question you honestly put to yourself. For your online visibility that question is surprisingly often: do I actually know where I stand, or do I only think so? The free SEO scan answers that with data instead of a feeling, in a few seconds. And if you want to keep talking about what AI can do for your business beyond that, from separate skills to full automation, we simply do that in a conversation.