ChatGPT as a Tutor: The Prompt That Works

The default failure is subtle enough that most people never notice it.

You ask ChatGPT to explain something. It produces a clear, well-structured, genuinely correct explanation. You read it, and you understand it — that part is real, the feeling of comprehension is not a lie. Then you try to use the idea four days later and it's gone. Not fuzzy. Gone.

This isn't a flaw in the model. It's the model doing exactly what it was built to do. A general assistant is optimized to be maximally helpful right now, and the most helpful possible response to "explain X" is a complete, well-organized explanation of X. Unfortunately, reading a good explanation is one of the least durable ways to learn anything. The learning happens when you struggle to produce the answer yourself — and a maximally helpful assistant removes exactly that struggle, every single time, on purpose.

Every AI tutor that has actually beaten a classroom in a controlled study works by undoing this. When Harvard's Gregory Kestin built the tutor that beat his own active-learning physics class, the design was mostly a list of prohibitions: keep responses to a few sentences, give away one step at a time, never provide the full solution in a single message.

You can impose most of that yourself. Here's how, and then an honest account of what's left over.

First: turn on Study Mode

Before any prompt, use the built-in tool. OpenAI's Study Mode was designed with educators and learning scientists across roughly 40 institutions specifically to make ChatGPT ask rather than tell. It's on every plan, free included. Google's Gemini Guided Learning does the same thing on LearnLM, with better diagrams.

Turning that on gets you most of the way. The prompt below is for when you want tighter control — a specific difficulty level, a specific format of questioning, or a session that ends with something you can actually review.

The prompt

Paste this at the start of a conversation. Replace the two bracketed lines and nothing else.

You are my tutor for: [TOPIC]
My goal: [what I want to be able to DO when we're done]

Run the session by these rules, and don't break them even if I ask you to.

DIAGNOSE FIRST
Before teaching anything, ask me 3-4 short questions to find out
what I already know about this. Start easy, escalate until I miss
one. Then tell me plainly where you think I'm starting from and
what we should skip.

TEACH IN SMALL PIECES
- Maximum ~150 words per turn. Short is the point.
- One idea at a time. End every turn with a question I must answer.
- Never give me a complete solution or a full explanation up front.
- If I'm stuck, give the smallest possible hint, not the answer.
  Make me try at least twice before you reveal anything.

MAKE ME PRODUCE, NOT RECOGNIZE
- No multiple choice. Make me explain, derive, or write it out.
- Regularly ask me to explain a concept back in my own words, and
  tell me specifically what my explanation got wrong or left out.
- Bring back things I struggled with earlier in the session, later
  in the session, without warning me.

WHEN I'M WRONG
Don't just correct me. Name the specific misconception I appear to
have, tell me why it's a tempting mistake, then make me redo it.

IF I ASK FOR THE ANSWER
Refuse once and give a hint instead. If I ask again, give it — but
then immediately ask me a variant question to check I actually
got it.

AT THE END
When I say "wrap up", output a LEARNING LOG in this exact format:
- Covered: (bullet list)
- Solid: (things I got right without help)
- Shaky: (things I needed hints on)
- Wrong: (specific misconceptions I showed)
- Next session should start with: (your recommendation)

Begin with the diagnostic questions now. Ask them one at a time.

Two details do most of the work here. "End every turn with a question I must answer" is what stops the drift back into lecturing. And "one at a time" on the diagnostic matters more than it looks — ask for four questions at once and you'll get four questions and answer them all at your own comfortable level, which defeats the point of escalating until you miss one.

How to actually run a session

• Say your goal as a capability, not a subject. "Learn statistics" gets you a Wikipedia article with extra steps. "Be able to look at an A/B test result and say whether I believe it" gets you a course.
• Answer badly on purpose, once. Early on, give a deliberately confused answer and see whether it diagnoses the confusion or just restates the right version. If it restates, the rules are slipping — say "you're explaining again, go back to asking" and it will correct.
• Type your answers out properly. The temptation is to think the answer and skim the response. The retrieval effect requires actual retrieval; thinking "yeah I know this" is precisely the illusion that produced the problem.
• Always run the wrap-up. The learning log is the only artifact that survives the session, and it's what makes the next one possible.

The four things no prompt can fix

Here's where the honesty comes in, because the pitch usually stops at the prompt.

1. It doesn't remember what you got wrong. This is the big one. The prompt makes it revisit your mistakes within a session; nothing carries across sessions. Tuesday's tutor has no idea you fumbled conditional probability on Sunday, so it can't do the one thing tutoring is actually for — spending time exactly where you're weak.

Partial fix: keep the learning logs in one document and paste the last one in when you start. This works, and it's genuinely tedious, and you'll stop doing it around week three.

2. It has no curriculum. It will teach whatever you point at, in whatever order you point at it. It won't tell you that you're trying to learn transformers without understanding matrix multiplication, because it doesn't know what comes before what — it only knows what you typed. Self-directed learners systematically pick the wrong order, and this is invisible to you by definition: you can't see the prerequisite you don't know you're missing.

Partial fix: ask it up front for a 10-lesson outline and make it justify the sequence. This is much better than nothing. It's still generated fresh each time, so ask twice and you'll get two different curricula.

3. It drifts. Long conversations wash out early instructions. Twenty turns in, it's writing four-paragraph explanations again. Watch for the turn where it stops asking you questions — that's the tell, and it usually happens right when the material gets hard, which is the worst possible time.

Partial fix: re-paste the rules block every 15-20 turns, or start a fresh chat per session and lead with the learning log.

4. It grades generously. Ask it to evaluate your explanation and it will find something to praise. It's poor at telling you that you've mostly got it but the load-bearing part is wrong, which is the most common state of partial understanding and the most important one to catch. Its assessment is conversational, not scored against a standard.

Partial fix: ask "what would a skeptical expert say is missing from my explanation?" — adversarial framing gets noticeably harsher and more useful feedback than "how did I do?"

When a prompt stops being enough

For a well-defined topic over a couple of weeks, the prompt above plus Study Mode is genuinely good, and it's free. Use it. Most people never get this far and it's a real upgrade over reading explanations.

The ceiling shows up on the long stuff. Somewhere around month two of learning something substantial, you notice you're doing all the work the tutor should be doing: maintaining the curriculum, remembering your own weak spots, deciding what's next, re-imposing the rules, and grading yourself. That's not a prompting problem you can prompt your way out of. Persistent memory of your errors, a sequenced syllabus, and honest assessment are things a product has to be built around — they're state and structure, not instructions.

That's the line where a purpose-built AI tutor starts to matter: one that runs the diagnostic once and keeps the result, generates a real syllabus for your goal, grades your practice against a standard, and reorders what's next based on what you actually got wrong — without you hand-carrying a text file between chat windows.

If you want the evidence for why those four properties are the ones that matter, we went through the tutoring research here. And if you'd rather compare the actual options first, the rundown of AI tutors for adults includes the free ones and says plainly where each is the better choice.