10 prompts that turn any AI into a patient tutor
Left to itself, a chat assistant lectures. It gives you a clean, complete answer, you nod, and a week later you remember none of it. These ten prompts flip the arrangement: the model explains only as much as it needs to, then asks you something, waits, and corrects you. Four were tested in September 2026. Each has a copy button, a note on when to use it, and the one edit to make first.

- Prompts10
- Read time12 min
- StatusTested September 2026
How to use this pack
Fill in the square brackets, paste the prompt, and then do the hard part: answer the question it asks you before you scroll. Every prompt here ends by handing the work back to you, and the learning happens in the ten seconds you spend trying. Two habits help. Start a fresh chat per topic so the model's picture of what you know stays accurate. And when you get something wrong, say so plainly; the tutor prompts are written to adjust, but only if you tell the truth about what confused you.
Prompts 1, 2, 6 and 8 were run through Gemini 3.6 Flash, the model on Gemini's free tier, via the API in September 2026; the rest are written to the same pattern but were not run. Some vendor links on this site are affiliate links; see our disclosure.
Why make the model ask you questions
The three big assistants now each ship a study feature: OpenAI launched Study Mode in July 2025, Google followed with Guided Learning in Gemini that August, and Claude's Learning mode arrived in Claude for Education in April 2025 before spreading. All three do the same thing. Switch them on and the model stops answering and starts asking. That the companies converged on this design is the strongest hint about how learning works: the research on retrieval practice keeps finding that pulling an answer out of your own head, even wrongly, beats reading a good explanation of it. A lecture feels like progress because it is fluent. A question feels worse because you might fail. The failure is the point.
There is a second reason. A model that explains cannot tell what you already know, so it explains everything at the same depth. A model that asks first can skip the parts you have and slow down on the parts you do not. Every prompt below is written so the first move is a small explanation and the second move is yours. If your assistant has a study toggle, these prompts still work alongside it and give you finer control over how hard it pushes. Note that ChatGPT's Study Mode shortcut moved in spring 2026, which produced a wave of "Study Mode is gone" posts; it was still there, under a different entry, and was back in OpenAI's own marketing by August.
Understanding
1. Explain it like I am new, then check
Explain [topic] to me as if I am smart but completely new to it. Use one everyday analogy, then the real explanation in plain words, under 250 words total. Avoid jargon; if a technical term is unavoidable, define it in the same sentence. Then ask me one question that checks whether I understood, and wait for my answer before continuing. Do not answer the question yourself.
Use it for the first ten minutes with any new subject. Tested with compound interest: the model gave a snowball analogy, a clean definition, and a check question with a specific number in it ("on what total will Year 2 interest be calculated?"), then stopped. Edit before sending: the word "new". If you know a little, say what; "I understand simple interest but not compounding" produces a much better explanation than a blank slate.
2. The Feynman check
I am going to explain [topic] in my own words, as if to a 12-year-old. Your job is to act as the Feynman check. Read my explanation and then: 1) point out every place I used a word I did not explain or hand-waved past a step, 2) identify the single weakest link in my understanding, 3) ask me one question that would expose whether I really understand that weak link. Do not explain the topic yourself yet. Be direct, not encouraging. My explanation: [write it here, badly if necessary]
Use it for the night before an exam or a presentation. Tested with a "wanted poster" explanation of vaccines: it flagged three undefined words, named the mechanism of antibodies as the weak link, and asked what an antibody physically does to a germ. That was the right question. Edit before sending: nothing in the prompt; the edit is to write your explanation without looking anything up. If you check first, the prompt tests your reading, not your understanding.
3. Worked example, then a fresh problem
Teach me how to solve this kind of problem: [type of problem, e.g. balancing chemical equations, calculating net present value, conjugating irregular verbs in the past tense]. First, solve one example step by step, and at each step say why you did it, not just what you did. Keep it to the minimum number of steps. Then give me one new problem of the same type but with different numbers or words, and stop. Do not solve it. When I send my answer, tell me if it is right, and if it is wrong, tell me which step went wrong and nothing more, so I can retry.
Use it for anything procedural: math, accounting, grammar, code. The "which step went wrong and nothing more" line stops the model handing you the full solution the moment you slip. Edit before sending: paste one problem from your own textbook or course as the type; the model's invented examples are sometimes easier than the real ones.
4. The misconception hunter
I am learning [topic]. List the five most common misconceptions that people at my level ([level, e.g. high school, first-year undergraduate, self-taught beginner]) hold about this topic, with the correct idea beside each. Then, without telling me which is which, give me five statements about the topic, some true and some containing one of those misconceptions, and ask me to sort them. Wait for my answer before revealing the key.
Use it for subjects where you feel confident, which is exactly when you are carrying a wrong idea. Tested with photosynthesis: the five misconceptions were the classic ones (oxygen comes from CO2, the dark reactions happen at night, plants do not respire), each paired with the correction, and then it set the sorting task and stopped. Edit before sending: the level. "High school" and "medical student" produce completely different lists.
Remembering
5. Spaced-repetition flashcards
Turn the material below into 20 flashcards for spaced repetition. Rules: one fact or idea per card, never two. The front is a question or a cloze deletion with one blank, not a heading. The back is under 25 words. Mix the types: about half recall ("What is X?"), a quarter "why" or "how" questions, a quarter applied ("Given Y, what happens to X?"). No card should be answerable from the wording of the front alone. Output as a table with two columns, Front and Back, so I can paste it into Anki or a spreadsheet. Then list any three cards you think are too easy and offer harder replacements.
Material:
[paste notes, a chapter summary, or a transcript]Use it for anything you need to still know in three months. The "no card answerable from its own wording" rule is what separates real cards from trivia. Edit before sending: the count. Twenty cards from a chapter is right; twenty from a paragraph produces padding.
6. A study plan from a syllabus
Here is the syllabus or topic list for [course or exam], and the date I need to be ready: [date]. I can study [hours] a week, in sessions of about [minutes]. My current level: [honest description, including what I have already covered]. Build a week-by-week plan to the date. Rules: order topics so each builds on the last; give every session one topic and one concrete task (read, practice problems, explain aloud, flashcard review); put a review session every week that only revisits earlier weeks; leave the last [X] days for practice tests only; flag the three topics most likely to be examined heavily and give them extra sessions. Output as a table: week, session, topic, task, time. Then ask me two questions that would change the plan if my answers were different. Syllabus: [paste]
Use it for any exam with a date. The two closing questions are the model's chance to catch what you left out, usually "which topics scare you" and "how did the last exam go". Edit before sending: the hours. Give the number you will actually do, not the number you wish you would; the plan is only useful if it fits.
Reading and speaking
7. Summarize a paper for a beginner
Below is a research paper (or its abstract and conclusion). Explain it to me as a beginner in five parts, each under 80 words: 1) the question they asked, in plain words; 2) what they did to answer it, without method jargon; 3) what they found, with the one number that matters; 4) what it does not show, including the limits the authors admit; 5) why anyone outside the field should care. Then give me one question I could ask an expert about this paper that would show I understood it. Do not add claims that are not in the text. Paper: [paste]
Use it for reading outside your field, and for checking whether a headline matches the study behind it. Edit before sending: paste the paper's own text, not a news article about it; part 4 is the valuable part, and only the paper contains its own caveats.
8. Language-practice conversation partner
Be my [language] conversation partner. My level: [beginner / about A2 / intermediate]. Scenario: [e.g. ordering coffee and a pastry at a cafe, checking into a hotel, explaining a problem to a landlord]. Speak only in [language], one or two short sentences at a time, and always end with something that needs a reply from me. When I make a mistake, put a corrected version of my sentence in brackets after your reply, and nothing more; do not explain the grammar unless I ask. Keep the vocabulary to what a [level] learner would know, and introduce at most one new word per turn. Start the scenario now.
Use it for daily ten-minute practice. Tested in Spanish: the opening was two short sentences and a question, in the target language only, nothing else, which is exactly the restraint most assistants lack. Edit before sending: add "if I write in English, reply in [language] anyway and carry on" if you tend to bail out. Use the voice mode on ChatGPT or Gemini for the same prompt spoken aloud.
Performing
9. The interview drill
Run a mock interview with me for [role] at [type of company]. Here is the job description: [paste]. Here is my background in three lines: [paste]. Ask me one question at a time, starting with an easy one and getting harder, mixing behavioral questions ("tell me about a time") with technical or situational ones drawn from the job description. After each answer, give me a score out of 10 and one specific way to improve it, in under 60 words, then ask the next question. Do not give me model answers unless I ask. After eight questions, stop and summarize my three strongest and three weakest patterns.Use it for the week before an interview; run it three times with the difficulty rising. The score is not the point; the "one specific way to improve" is. Edit before sending: paste the real job description. Generic drills produce generic questions, and the real interviewer will not ask those.
10. What would an expert ask?
I have been learning [topic] for [time] and I think I understand it. Act as a senior expert in this field who has been asked to find out, politely, whether I actually do. Ask me the five questions you would ask a junior colleague to test real understanding rather than memorized facts: questions about edge cases, about why things are done one way and not another, and about what breaks when an assumption changes. Ask them one at a time. After each of my answers, say whether an expert would accept it, and if not, what an expert would have said, in two sentences. Do not soften the verdicts.
Use it for the end of a course, before you put a skill on your CV, or before you ship something built on it. Edit before sending: the field. "Statistics" gets you textbook questions; "A/B testing for a small e-commerce site" gets you the questions someone would actually ask.
What the tests showed
We ran prompts 1, 2, 4 and 8 through Gemini 3.6 Flash, the model on Gemini's free tier, in September 2026. In every case the model did the thing these prompts are for: it stopped. It explained, asked, and waited, rather than answering its own question, which is the failure the "wait for my answer" line is there to prevent. The Feynman check was the most useful single output, because it named the weak link ("you treat antibodies as a magic word") rather than listing everything that could be improved. The misconception list was accurate for the level asked. The one thing to watch is politeness: even with "be direct, not encouraging" in the prompt, the model tends to open with praise. Add "no praise" if it bothers you.
Which assistant to run them in
All three flagship assistants run these well on their free tiers. Independent comparisons through 2026 give Claude's Learning mode the strictest hand at withholding answers, which suits prompts 2, 4 and 10. ChatGPT's Study Mode works on the free plan and its voice mode is the natural home for prompt 8. Gemini's Guided Learning is strongest for diagrams and step-by-step material, and the free tier's Deep Research is useful for gathering the paper for prompt 7. Students should check the free tiers and deals guide first: Google is giving eligible college students a year of AI Pro, and OpenAI four months of Plus, as of September 2026.
Quick answers
Do I need ChatGPT Study Mode or Gemini Guided Learning to use these prompts?
No. Study Mode, Guided Learning and Claude's Learning mode are built-in versions of the same idea: the model asks before it tells. These prompts do the same job in any chat, including free tiers and assistants without a study toggle, and they give you more control over how hard the tutor pushes.
Why should the AI ask me questions instead of explaining?
Because retrieving an answer from your own memory is what makes it stick, and reading a clear explanation only feels like learning. Every prompt in this pack ends with the model handing the work back to you: a check question, a sorted list, a fresh problem. If you answer wrong, it tells you and tries again.
Which AI is the best tutor?
For holding the line and not giving the answer away, Claude's Learning mode is the strictest in independent comparisons. ChatGPT's Study Mode works on the free plan and is the easiest to start with. Gemini's Guided Learning is strongest for visual and step-by-step material. All three run these prompts well; pick the one whose free tier you already use.
Sources
- Glasp — AI study modes compared: ChatGPT Study Mode vs Gemini Guided Learning vs Claude Learning Mode, 2026
- Multi Chats — Study Mode vs Guided Learning vs Gemini in Classroom, August 2026
- FindSkill — Where did ChatGPT's Study Mode go, August 2026
- iatroX — ChatGPT Study Mode vs Gemini Guided Learning vs Claude Learning Mode for medical students, September 2026


