The AI Feynman Technique Coach skill for Claude

The AI Feynman Technique Coach is a free skill from the library we keep ourselves. A skill is an instruction file, SKILL.md, that gives an AI assistant a fixed way of working for one task. No software to install, no subscription, no integration with anything: a text file of 1,196 words that tells Claude how to walk you through the feynman technique, which questions to ask, which words it may not use and at which moment it has to stop and wait.

The problem the skill solves is the illusion of understanding. You read an explanation, you nod, you think you have it. And then someone asks you about it and you hear yourself using words you cannot unpack. That happens to everyone preparing for a pitch, an exam, a client conversation or an explanation to family. Feynman's method puts a hard test against that: if you can explain it to a twelve year old without jargon, you understand it. Fail that test and you know immediately where your knowledge is thin.

Written for anyone who has to hand something over: business owners who want to make their own service understandable, advisers explaining a complex recommendation to a board, students preparing for an exam, and marketers who have to describe a technical product in human language. The frontmatter of the file names exactly those moments, with the phrases people actually use: explain this, help me understand this, I want to really get this, make this simple, I need to be able to pitch this, explain this for a layperson. And the reverse question too: check whether I really understand this, test my understanding, where are the gaps in my knowledge.

One thing is worth knowing up front, because it decides whether you will like this skill or not. It does not do the work for you. In step 3 it marks the weak spots in your explanation and asks you to fill them, with an explicit ban on answering on your behalf. Anyone after a finished text is in the wrong place. Anyone who wants the subject genuinely in their head is exactly right. It belongs to the AI and automation service and sits alongside the rest in the skill library (in Dutch), with no account and no sales email afterwards.

When an explanation does not land, people usually look for the cause in the listener. Almost always it sits in the explanation itself. The rules the skill imposes in step 2 read like a list of exactly those mistakes, and that is no accident: they were written to make each of them impossible.

Jargon papers over the gap. A technical term is an abbreviation of an idea. Use it while your listener does not have the idea, and you have said nothing. Worse: technical terms feel like knowledge to the speaker, while they often mark the exact point where the thinking stopped. So the file bans jargon, technical terms and buzzwords in the first explanation, and adds that a term you genuinely cannot avoid is explained straight away with a comparison.

The sentences are too long. There is a hard limit in the file: fifteen words per sentence at most in the explanation for a twelve year old. That is not a style preference but a thinking test. A long sentence can glue three half ideas together without anyone noticing that the middle one is missing. Cut it into three and the missing link stands out immediately.

The explanation is about the theory instead of about the world. What it asks for is concrete examples from daily life, and there is even a rule about conjunctions: avoid "because" followed by a technical reason, use "because" with human logic. That sounds like a detail, but it forces you off the layer of definitions and onto the layer of causes someone recognises.

The length is unlimited. Given an explanation without a bound, everyone fills the uncertainty with more text. So there is a ceiling of 250 words on the first explanation. Whatever does not fit is probably something you do not yet understand well enough to leave out.

The tone is model tone. Rule 3 in the file literally bans the stock sentences language models warm up their answers with, from "it is important to understand that" to "let us take a look at". It also says the skill gives no empty compliments: not "great question" first, but straight to the substance. That makes the answer shorter and the conversation more honest.

In diff form, using the mortgage example from the file: what goes out and what comes back in its place.

A skill is only as good as its instructions, so here they are, in the order they appear in the file. At the top sits the frontmatter with the name ai-feynman-technique-coach, version number 1.0.0, the MIT licence and a long description that decides when Claude picks the skill up. That description is written in a strikingly practical way: it lists the phrases people use when they ask for an explanation, right down to asking whether something can be put in plain English. Say something like that to Claude while the skill is loaded and you get it automatically, without a command.

This is the top of the file, so you can read it before you download anything:

---
name: ai-feynman-technique-coach
version: 1.0.0
license: MIT
description: "Applies the Feynman Technique to any subject you
  genuinely want to understand. Forces you to explain a complex
  concept so simply that a child gets it, finds the gaps in your
  knowledge and fills them with sharp analogies. ALWAYS use this
  skill when someone says 'explain this', 'help me understand
  this', 'I want to really get this', 'make this simple', ..."
---

Then comes the goal in two lines: take the user past surface knowledge, build real understanding by forcing simplicity, finding gaps and using analogies that stick. After that the core: four steps that according to the file have to be followed strictly. Step 1 frames with four questions and produces a working definition of one sentence. Step 2 writes the explanation for a twelve year old, with five hard writing rules and a fixed format of four headings. Step 3 finds the gaps with six Socratic questions and a three line marking per weak spot. Step 4 rewrites everything with at least three analogies, a metaphor and two lengths.

Then the output structure is written out in full, including the order in which the blocks come back: working definition, explanation for a twelve year old, knowledge gaps found, then the wait line, and only after that the final explanation with the elevator version, the in-depth version and the memorable metaphor. That fixed format is the reason two sessions with this skill look alike and can be put side by side to see whether your understanding has grown.

Under the steps sit five rules about tone and honesty. Be honestly hard when the explanation is vague. No compliments without content. Avoid the stock sentences of language models. Point at your source and say when you are not certain, with a reference to Feynman's well known line about fooling yourself most easily. And fifth: two iterations at most, because if it has not landed after two rounds the problem runs deeper than a better wording.

It closes with three parts you do not see on many skills. A worked example, the mortgage conversation you see in FIG.01 above. A source list with three titles by Feynman himself and two books for further reading. And a short list of the situations where you should not use the skill.

The SKILL.md names its sources, and that is deliberate: you can read them and decide for yourself whether you agree with the approach. The method is named after Richard Feynman, who lived from 1918 to 1988, physicist, Nobel laureate and known by the nickname The Great Explainer. There are three titles by the man himself: Surely You Are Joking, Mr. Feynman from 1985, What Do You Care What Other People Think from 1988, and The Feynman Lectures on Physics, published between 1963 and 1965. For further depth the file names two books about learning: Make It Stick by Peter Brown, Henry Roediger and Mark McDaniel, and A Mind For Numbers by Barbara Oakley.

Worth knowing: the four steps as they stand here are not a method Feynman himself wrote down in this form. They were derived later by others from the way he worked and taught, and that is exactly how the skill presents them: as a technique named after him that rests on his starting point. That starting point is his, and it is sharp: if you cannot explain something to a twelve year old, you do not really understand it yourself.

Why does that starting point work so well? Because it is the one test you can run on yourself without an expert in the room. You cannot hear yourself say you understand something and check it. You can hear yourself stumble the moment you take the technical term out. Three extra tests are built on top of that. The writing rules of step 2, which strip the jargon. The six questions of step 3, which point at the gap as soon as your explanation thins out. And the follow-up test of step 4, which asks whether your analogy survives ten questions or falls apart at the third.

That last point is underrated. A good analogy does more than illustrate: it carries the structure of the idea across, so your listener can think onwards by themselves. A bad analogy only pretends, and breaks at the first follow-up question, after which your audience not only fails to grasp the idea but stops trusting your explanation. So the file demands three from different domains, from sport, cooking, driving, nature, children or money: what falls away in one analogy becomes visible in the next.

And then the honesty rule the file adds itself: if the explanation is uncertain, that has to be said. Rule 4 refers to Feynman's well known line that you are the easiest person to fool. For an AI skill that is more than a nice sentiment. A language model effortlessly produces a fluent explanation of something it is not sure about, and that fluency reads as reliability. The instruction to name what is uncertain is therefore not modesty but a correction mechanism.

The zip is 3.6 KB and contains one folder, ai-feynman-technique-coach, with the SKILL.md inside it. Installing means putting that file in the right place, and the right place differs per environment. SKILL.md has been an open standard since December 2025, so the same skill also works in Codex, Cursor and Gemini CLI. You are not downloading a Claude file but a working instruction that any modern AI assistant can read.

Once it is in place, using it is simple: name the subject and say who you have to explain it to. Allow fifteen to thirty minutes, because that is the time the file itself names for a full cycle. Most of that is you typing in step 3, and that is not a design flaw but the point.

It is at its strongest when you shortly have to explain something out loud to someone who knows nothing about it. An investor, a board, a client, a family member, an examiner. In all of those cases recognising the material is not enough: you have to produce it, in your own words, under follow-up questions. That is exactly what the four steps train.

It is also useful as a self-test on something you think you have mastered. Pick a subject from your own field and let the skill break your explanation open. The six questions of step 3 are uncomfortable precisely because you no longer ask them yourself: about your own field you have stopped looking naively.

And there is a third use that is less obvious: getting your own offer sharp. Cannot explain your own service in 250 words without jargon? Then your client does not understand it either. The jargon ban of step 2 works there the same way it does in physics. Write the description out first and the padding shows up on its own, usually as adjectives that nobody would miss if you cut them. What is left once you remove the fine words is what you actually sell.

There are limits too, and they are simply in the file. The first: pure fact lookup. For a date or a definition, searching is faster than a four step cycle. The second: fast task execution. The method costs fifteen to thirty minutes, so if you need something finished within ten, this is the wrong instrument. The third: subjects you are already expert in and only want a second opinion on. You get an explanation at beginner level back, and that is an answer to a question you did not ask.

We add one honest limit the file does not name. It tests whether your explanation is internally consistent and whether you see the gaps yourself, but a fact check it is not. If your explanation contains an error that is carried through consistently, it will happily build three analogies for it. If you are about to explain something that carries weight, such as medical, legal or financial material, check the facts with a real source or a colleague in the field before you pass it on.

Looking for something else, then the library probably has a better fit. If the problem is not understanding a subject but keeping the practice up, the AI Atomic Habits Coach is the one to read next. To strip a subject down to its building blocks rather than explain it, the First Principles Thinker skill is the logical choice. For reaching insight through questions instead of through explanation, the Socratic Method Coach skill does that systematically.

Download the skill, pick the subject you have to talk about next week and see whether you can get it round in 250 words without technical terms. It costs half an hour and nothing at all, and you know straight away where your explanation wobbles.

The same test works on working with AI itself: ask someone to explain what they asked the AI and why, and within a minute you can hear whether they understand it or are only clicking along. Should your team practise that way instead of watching a demo, that is what the AI training does. And if you are looking for more instruments of this kind, the rest are in the library of 100 free skills (page in Dutch).