When your goal is to learn, ask for help that leaves you a next step to do. A small hint, a question or a pointer to the first error can preserve an attempt that a complete solution would remove.
This is a practice method, not a claim that withholding always teaches better. Explanations and worked examples can be useful, especially when the prerequisite knowledge is missing. A chatbot can also supply an incorrect hint or approve a mistaken answer.
Choose a safe practice task
Use a problem whose answer can be checked in course material, an answer key or another appropriate source. Follow your institution’s AI rules, and do not upload live assessments, restricted course content or private workplace material without permission.
These prompts are for low-risk learning exercises. For an actual medical symptom, urgent safety concern or consequential decision, stop the tutoring exercise and use the appropriate professional or official service. If someone is in immediate danger, contact local emergency services instead of waiting for AI. In the EU, 112 is the emergency number. For urgent medical help that is not an emergency, use your local urgent-care route; NHS 111 is an England-specific example. Asking the chatbot for a full answer does not make it a safe substitute.
Pattern 1: one hint, then an attempt
Here is a permitted practice problem and my attempt: [input]. Give one small hint without the solution. Wait for my next attempt. If my reasoning starts from a false premise, identify it. Do not pretend that an unsupported answer is established.
AI-generated teaching example, checked arithmetically:
Problem: solve 3x + 5 = 20. My attempt: subtract 5, giving 3x = 15. I am stuck.
Hint: what operation undoes multiplying x by 3?
Learner’s next step: divide both sides by 3, so x = 5.
Check: 3 × 5 + 5 = 20.
The hint leaves a meaningful operation for the learner. “The answer is 5” removes that step. A hint that says “subtract 3” would be wrong; the check is essential.
Pattern 2: ask a question, then check the answer
Using this permitted source, ask one question that tests my understanding. After I answer, compare it with the source, quote the relevant passage and explain any mismatch. Ask a follow-up only when it targets a specific gap.
Do not require the model to withhold correction through an arbitrary number of questions. That can let a false premise shape the rest of the exchange. A question should reveal what you can explain, not merely prolong the session.
Pattern 3: locate the first error
Compare my attempt with the supplied source or answer key. Identify the first unsupported step and give a clue rather than a complete replacement solution. If the key is missing or ambiguous, say what cannot be decided.
A forced “right or wrong” answer can conceal uncertainty. Check the model’s feedback against the source before practising it again. Research on sycophancy in language models gives a reason to challenge agreeable feedback; it does not establish how every current tutor behaves.
Know when to change approach
If successive hints leave you guessing, ask for the missing prerequisite or a worked example, or take the question to a teacher. Difficulty alone is not evidence of learning.
The Dunlosky and colleagues review supports practice testing and distributed practice across the settings it reviews. It does not validate these particular prompts or guarantee that a guided AI exchange produces retention. OpenAI’s prompting guidance similarly supports specific, iterative requests rather than a teaching-outcome guarantee.
Use the practice card to record which hint helped and how you checked it. Later, close the chat and try a different problem without help. That independent attempt is more informative than the model saying you understand.



