Learning anything faster with AI: from \"explain like I'm 12\" to practice quizzes

The single most underrated use of ChatGPT is as a tutor. People discover it for emails and spreadsheets first, miss this entirely for a while, and then once they try it for learning, never look back.

Tutoring is a job AI is structurally well-suited to. It is patient, never bored, available at 2 a.m., happy to explain the same thing for the fifth time without sounding tired, and willing to be quizzed for an hour without complaint. What it needs from you is structure — a four-prompt loop that turns it from a chat into a private lesson.

This article is that loop, and how to use it.

Why the obvious approach doesn't work

The instinctive way to use AI for learning is to type "explain compound interest" and read whatever comes back. That works for the first thirty seconds — but it does not actually teach you anything. You read a paragraph, nod, and forget it.

What is missing is the parts that make humans learn anything: working examples, active recall, and feedback. The four-prompt loop is built around those.

The four-prompt loop

Prompt 1 — Get the explanation. Prompt 2 — Get worked examples. Prompt 3 — Get quizzed. Prompt 4 — Identify what to revisit.

Do all four in a single conversation. Each one builds on the previous.

Prompt 1: Get the explanation, three ways

Pick a topic. Anything you have wanted to understand. Start with:

I want to learn [topic]. I have [some background detail — what you know already, what your goal is, why you care].

>

Explain it three different ways, in roughly 150 words each:

>

1. To someone with no background at all (curious 12-year-old level). 2. To a working professional in a different field — accurate but accessible. 3. To me, given what I told you about my background, with the most useful framing for someone with my goal.

>

Use one short concrete example in each. Tell me if any of them is oversimplified to the point of being misleading.

Three things to notice. You are not asking for a definition; you are asking for three explanations. You are telling the model your background so version 3 is calibrated. And you are inviting it to flag when an analogy breaks down — which surprisingly often it will, if you ask.

By the third version, something usually clicks that did not in the first. If it does not, ask: "Try version 3 again, but with a different analogy and a different example."

Prompt 2: Get worked examples

Understanding the concept and being able to apply it are different things. The bridge is worked examples — problems with the solution explained step by step. A reliable prompt:

Now give me three worked examples that show this in practice. For each one:

>

- Describe the situation in one or two sentences. - Walk through the solution step by step. - Highlight the single most important move — the part where someone would get stuck without understanding the concept.

>

Start with an easy example. Make the second medium. Make the third one that has a genuine subtlety in it.

This is where AI is materially better than most textbooks: it can generate examples calibrated to where you are. If something is too easy or too hard, say so. "Make the third example harder — show me one where the obvious approach gives the wrong answer."

After three worked examples, you should have a feel for what doing the thing looks like, not just what it is.

Prompt 3: Get quizzed

Now active recall. This is the most important prompt and the one most people skip. Active recall — being asked, retrieving an answer from memory, getting feedback — is the most reliable learning method we know of. Reading is much weaker. AI makes this effortless:

Now quiz me. Ask me ten questions about what we have covered, mixing easy and hard. Rules:

>

- One question at a time. Wait for my answer before moving on. - After I answer, tell me whether I'm right, what I missed, and (if I was wrong) which specific concept I should revisit. - Do not give me the answer if I write "skip." - At question 6 or so, throw in one question that requires combining two things, not just recalling one. - At the end, summarize which two or three concepts I should review most.

The "one at a time" instruction matters. Without it the model will ask all ten at once, which is useless because you read them all together and skip the retrieval step. Forced one-by-one, you actually have to think.

After ten questions, the summary at the end is where the real learning happens. You will see, plainly, what you still do not understand.

Prompt 4: Identify what to revisit

This is the cleanup pass. After the quiz, follow up with:

Based on what I got wrong or said "skip" to, give me a short revision plan:

>

- The two concepts I should revisit, and why I missed them. - A short worked example for each that targets exactly where I tripped. - A re-quiz of three questions, only on those two concepts.

>

Do not re-cover the things I clearly understood.

This is impossible to do alone — you cannot reliably notice what you do not understand. The model can, because it has a record of your wrong answers. The revision plan is calibrated to your gaps, not generic ones.

Run prompt 3 and 4 again a day later, asking the model to recover what you covered yesterday. You will be surprised by how much sticks compared with reading alone.

Where this works best

The loop is general. People have used it to learn:

• A new technical skill (Python, statistics, accounting, machine learning, a specific framework or tool)
• A foreign language vocabulary or grammar point
• A regulatory regime (GDPR, the EU AI Act, a country's tax basics, a specific contract type)
• A new role they just started in (what does "Series A" mean, what is a sales motion, what is OKRs)
• A medical or scientific topic they want to understand at a serious-amateur level
• A new product or codebase, by pasting in documentation and quizzing yourself

The pattern is the same. Explanation, examples, quiz, revision.

A few small variations

Spaced repetition. Have the model space your review across days. "Quiz me on yesterday's material first, today's material second."

Teach-back. After the quiz, ask the model to be a curious student and have you explain the topic. Pretending to teach is one of the most reliable ways to find your own gaps.

Compare and contrast. Once you understand a topic, ask: "What is the most common mistake people make confusing X with Y?" The error analysis often teaches you more about the concept than the original explanation did.

The first-principles version. For topics with a long history or a lot of jargon, try: "Explain this from first principles — starting from things any educated adult already understands — and build up to the current state of the art without using jargon. Introduce each piece of vocabulary only when we need it."

A subtlety worth knowing

The model is not infallible on subject knowledge. On well-known topics it is reliable; on niche topics, fast-moving research areas, or precise factual details (specific dates, specific quotes), it will sometimes invent. The fix is the same as everywhere else: when something matters precisely, ask the model to use search, paste in the source material, or check the answer against another source.

But for the learning process — explaining, structuring, quizzing, calibrating — it is excellent. Even when it is slightly off on a specific fact, the structure of the lesson it generates is sound, and you can correct the fact and continue.

The whole loop, in one line

Explanation → examples → quiz → revision. Run it tomorrow on something you have been meaning to understand. Twenty minutes later you will be further along than three months of half-finished textbooks. And you will discover that the slowest part of learning was never the topic — it was the lack of a patient teacher.