You read the chapter twice, highlighted half of it, and it felt familiar. Then the exam asked you to produce something from nothing and the familiarity turned out to be worthless.
That's the core problem with how most students use AI too. They ask for a summary, nod along, and feel prepared. The better use of AI for studying is the opposite: make it ask you questions, make you answer from memory, and tell you exactly where you went wrong. This post gives you the prompts that do that, the limits you need to know about, and a section on where not to use AI at all.
Why quizzing beats summaries
Retrieval practice, which means pulling an answer out of your head instead of looking at it, is one of the best-supported study techniques there is. In a widely cited 2006 experiment, Roediger and Karpicke found that students who practiced recalling a passage remembered more a week later than students who reread it, even though rereading made the rereaders feel more confident. A 2013 review by Dunlosky and colleagues in Psychological Science in the Public Interest rated practice testing and spacing your study over time as high-utility, and rereading and highlighting as low-utility.
So the job for AI is not "explain this to me again." It's "test me, then show me what I missed." Everything below is built around that.
The setup prompt (paste this first)
Chatbots default to being helpful, which means explaining. You have to override that in the first message. I wrote this version to be short enough that you'll actually re-paste it when the model drifts.
You are my exam coach for [COURSE, e.g. Intro to Microeconomics].
Rules for this session:
1. Ask me ONE question at a time, then stop and wait.
2. Do not give hints or the answer until I've attempted it.
3. After my answer, say whether it's correct, partially correct or wrong, and
name the specific idea I missed or got right.
4. Use only the material I paste below. If a question needs something outside
it, tell me instead of guessing.
5. Mix question types: definitions, "why" questions, and short applied problems.
6. After 10 questions, list the topics I got wrong or half-right.
MATERIAL:
[paste your notes, slides text or textbook section]
Rule 4 matters most. Without it the model fills gaps from its general knowledge, and that's how you end up studying something your professor never taught or, worse, something wrong. If you're new to why that matters, the giving context lesson covers it.
Prompts that make you retrieve
Question generator from your notes
Use this when you've finished a chunk of reading and want to check what stuck.
From the notes below, write 8 questions of increasing difficulty. No answers yet.
Include at least 2 that ask "why" or "what would happen if", not just
definitions. After I reply to all 8, mark each one and explain misses.
Explain-back (the Feynman check)
This one exposes fake understanding quickly. You explain the concept in your own words and the model plays a confused first-year, then grades you.
I'm going to explain [CONCEPT] as if to a classmate who missed the lecture.
Act as that classmate: ask me clarifying questions when something is vague or
skips a step. When I say "done", give me: (a) anything I got wrong, (b) the one
step I skipped that matters most, (c) a 1-5 score for how well someone new
could follow it.
If you can't get through it without looking at your notes, that's your answer. Close the notes and try again tomorrow.
The "where would I get this wrong" prompt
Given this topic, list the 5 mistakes students most commonly make on exams
about it, and write one short question for each that would catch the mistake.
Treat the list as hypotheses. The model is predicting common errors, not reading your professor's mind. Past papers and your own graded work are better evidence, and you can paste those in too.
Interleaved mixed quiz
Studying one topic until it's perfect feels productive. Exams mix topics. Near the end of your prep:
Here are the topics for my exam: [LIST]. Give me 12 questions shuffled across
all of them, one at a time, no labels showing which topic each belongs to.
A worked example (illustrative)
The output below is illustrative, written by me to show the shape of a good exchange. It isn't a captured transcript from a specific model.
Coach: Question 3. A city caps rents below the market rate. Name two effects
you'd expect over several years, and say why.
You: Landlords make more money because demand is high.
Coach: Wrong direction. A cap below market price reduces what landlords can
charge, so profit per unit falls, which tends to shrink supply over time
(less new building, more conversion to other uses). You also skipped
the demand side: more people want units at the lower price than there
are units, so you'd expect shortages and non-price rationing such as
waiting lists. Marked: wrong. Topic to review: price ceilings and
shortage.
Notice what makes it useful. It names the topic, it explains the mechanism rather than just stating a verdict, and it ends with something you can write down.
How do I space my reviews with a chatbot?
Spacing means revisiting the same material over days instead of cramming it once. A common rule of thumb is to review after one day, three days, a week, then two weeks, and tighten the gaps for anything you keep missing. That's a heuristic, not a law. What matters is that you come back after you've partly forgotten.
If you start on 8 October, that schedule lands on 9, 11 and 15 October, then 22 October. Count back from your exam date and fit as many reviews in as you can.
The catch: a chatbot won't remember last week's session. Don't rely on it. Do this instead.
- At the end of each session, send: "List the topics I got wrong or half-right, one line each."
- Paste that list into your own notes file with the date.
- At the start of the next session, paste it back: "These are my weak topics. Start with 5 questions on them, then mix in new material."
That's a spaced-review system that costs you two minutes and doesn't depend on any app's memory feature.
Which prompt for which situation?
| Situation | Use | Skip |
|---|---|---|
| Just finished a lecture or chapter | Question generator from notes | Asking for a summary |
| Concept feels "sort of" understood | Explain-back | Asking for another explanation |
| 1-2 weeks before the exam | Weak-topic list + interleaved quiz | Making new flashcards for everything |
| Night before | Short mixed quiz on weak topics | Learning new material |
| Quantitative subject (maths, physics, stats) | Ask for similar problems, attempt on paper, then compare | Asking it to solve and "walk you through" first |
For maths and physics specifically, treat the model's worked solutions with suspicion. Chatbots can make arithmetic and algebra mistakes that look confident. Check final answers against your textbook's answer key or a calculator.
What goes wrong, and how to catch it
Wrong answer keys. The model can mark you wrong when you're right, or right when you're wrong. When a verdict surprises you, check the original notes or textbook before arguing with it. If the model is quoting something, ask "which line of my notes supports that?" A model that can't point to one is guessing. More on this failure in the hallucinations lesson.
Drift back to lecturing. After a few turns the model starts giving hints or answers unprompted. Re-paste rule 1 and 2 from the setup prompt.
Easy questions. Models default to mid-difficulty recall. If you aren't getting stuck, ask for "harder, exam-style questions that combine two concepts" or paste a real past paper as a style guide.
Missing context. If you give it two paragraphs of notes it will ask you things the notes can't answer. More material in, better questions out. For long sources, see working with long documents.
Built-in tutor modes. ChatGPT has a study mode that guides you with questions instead of answering outright; OpenAI documents it in its help center, and menu labels and availability have changed more than once, so check the live page. Anthropic introduced a learning mode as part of Claude for Education, and I couldn't confirm on 2026-10-08 that it is a general setting for every Claude user. Either way, a plain-text rule like the setup prompt above works in any chatbot, and you won't depend on a feature that may move.
When not to use AI
This section is more important than the prompts.
- Graded work with a no-AI rule. Generating, rewriting or "polishing" an essay or problem set that you submit as your own is the quickest way into an academic integrity hearing. Detectors are unreliable, but instructors also spot a sudden change in voice, and some courses ask you to submit your chat history.
- Anything your syllabus doesn't clearly allow. Rules vary by course and instructor even inside one university. Read the syllabus. If it's silent, email the instructor and keep the reply.
- Citations and references. Models invent plausible-looking sources. Never paste a reference into a paper without opening it yourself.
- The first pass at understanding. If you ask AI to solve the problem before you've struggled with it, you lose the struggle that builds memory. Attempt first, even badly.
- Sensitive data. Don't paste anything that includes classmates' names, grades, or your institution's non-public material.
A simple test: could you do this same thing tomorrow, on paper, in the exam room? If your AI session is training you to do that, good. If it's doing the part the exam will test, you're borrowing knowledge you'll have to return.
Putting it into a week
You don't need a system. You need a loop:
- After each lecture, run the question generator on your notes (10 minutes).
- Once a week, run explain-back on the two concepts that feel shakiest.
- Save the missed-topic list every time.
- In the final fortnight, run interleaved quizzes and paste the weak list at the start.
If you want a reusable library of prompts for other tasks, the prompt library has copy-paste versions, and how to use AI for research covers the longer-form side, such as literature reviews, where the rules about checking sources apply even more strongly. If you teach rather than study, the teachers' guide is the other side of this.
