Ask a chatbot a question you are stuck on, and by default it gives you a full, correct, immediate answer. That is what “helpful” means to most AI products out of the box. It is also close to the opposite of what a good human tutor does when you are stuck on something you are supposed to be learning — a good tutor asks you a question back, gives you a small nudge, and makes you do the next step yourself. The direct-answer default is efficient and teaches you very little; the guided-struggle approach is slower and is where the actual learning happens.
This article is a small set of prompt patterns that push AI from “answer machine” toward “guide,” for the specific situations where struggling productively is the point.
The concept: withholding is the tutoring skill
A human tutor’s most important skill is not knowing the answer — it is knowing when not to give it. A student who is handed the answer the moment they hesitate never builds the specific mental muscle of working through the hesitation themselves. A student who gets one well-placed hint, tries again, and eventually gets there on their own remembers the solution far better, because they built it rather than received it.
AI defaults to the “hand over the answer” mode because that is usually what users want for quick tasks — looking up a fact, drafting an email, fixing code fast. Learning is a different mode, and it has to be requested explicitly, because the model has no way to know you want to be made to struggle unless you say so. The default also leans toward agreement rather than correction: assistants trained on human preference judgments show measurable sycophancy, including revising a correct answer when a user pushes back and tailoring feedback to what the user appears to want (Sharma and colleagues, Anthropic, 2023). A tool tuned to please is not a tool that withholds by default.
The reason withholding is worth the friction is that producing an answer and reading one are not the same operation. Experimental work found that taking a memory test improved later retention more than an additional study pass did (Roediger and Karpicke, 2006), and practice testing came out well across a wide range of conditions in a broad review of study techniques (Dunlosky and colleagues, 2013). Every hint you do not need is a retrieval you got to keep.
The common misconception
The mistake is assuming a chatbot naturally behaves like a tutor because it can explain things well. Explaining well and tutoring well are different skills. An AI tool that explains a concept beautifully the moment you ask is doing the same thing a textbook does — giving you the finished answer to read. A tutor’s job is to make you produce the answer, with just enough support that you succeed rather than flail. Getting AI into that second mode takes a specific instruction, not just a good question.
Pattern 1: the hint ladder
Instead of asking for the answer, ask for hints that escalate only when you ask for the next one:
I am trying to solve this myself: [describe the problem].
Do not give me the answer or a full explanation.
Give me one small hint — the smallest nudge that might unstick me.
Wait for me to try again. Only give me a slightly bigger hint if I ask for one.
Only give me the full explanation if I ask for it directly after at least two hints.
This keeps you in the driver’s seat. Most of the time, the smallest hint is enough to get unstuck, and you keep the satisfaction — and the memory — of having produced the answer yourself.
Pattern 2: ask me a question back
For a concept you are trying to understand rather than a specific problem to solve, ask the model to interrogate your thinking instead of explaining:
I think I understand [concept], but I am not confident. Instead of explaining it to me, ask me questions that would reveal whether I actually understand it.
Ask one question at a time. Based on my answer, ask a follow-up that targets whatever seems shaky.
Do not tell me if I'm right or wrong until after at least three questions — just keep asking.
This flips the interaction: instead of you evaluating whether the model’s explanation makes sense, the model evaluates whether your understanding holds up under questioning — much closer to what an oral exam or a good study partner does.
Pattern 3: confirm, don’t correct
When you already have an attempted answer and just want to check it, ask for confirmation with a clue rather than a full correction:
Here is my answer: [your attempt].
Tell me only whether I am right or wrong, and if wrong, give me one clue about where the error is — not the corrected answer.
I will try again before you tell me more.
This preserves a second attempt instead of converting every check into a full explanation, which is the same discipline behind deliberate practice with AI — attempt, minimal feedback, retry, and only then the full answer.
Where hint mode is the wrong choice
Do not use hint-withholding prompts for questions where a direct, correct answer matters immediately — a medical symptom, a safety question, an urgent factual check, or anything where getting stuck and guessing wrong has a real cost before you get to try again. Hint mode is for learning contexts where struggling and occasionally being wrong is safe and productive. Ask directly and get the full answer when the stakes call for it.
Hint mode is also the wrong tool when you are simply trying to get unblocked on a task you need finished, not a skill you are trying to build — the skill-preservation framing applies here: decide first whether this is something you are learning or just something you need done, and pick the mode accordingly.
What one well-aimed question surfaces
Consider an illustrative case. Someone learning basic statistics gets stuck on when to use a paired versus independent-samples test. Asking directly gets a clear, complete explanation that feels satisfying and is mostly forgotten by the next session. Using the question-back pattern instead, the model asks: “In your specific scenario, are you measuring the same subjects twice, or different subjects in two separate groups?” That single question, answered honestly, reveals the actual point of confusion — the person had been measuring the same group at two time points and mistakenly treating it as an independent-samples case. The correction that follows lands harder because the learner found the specific gap themselves, prompted by one well-aimed question, rather than reading past it in a general explanation.
Signs the pattern isn’t working
Hint mode is not free of friction, and some friction is a signal worth reading rather than pushing through. If you have gone through three or four escalating hints and still cannot make progress, that usually means the gap is a missing prerequisite, not a shortage of hints — stop the ladder and ask directly: “What do I need to already know to make sense of this hint?” A tutor’s withholding is meant to make you produce the next small step, not to make you guess at something you have no foundation for yet. Escalate to a full explanation, or to a human teacher, rather than grinding through hints that are not landing.
Combine this with the broader tutoring loop
These three patterns are a lighter-weight complement to the four-prompt tutoring loop — that loop is the right tool for building understanding of a new topic end to end (explanation, examples, quiz, revision); these hint-withholding patterns are what to reach for specifically in the moment you are stuck and tempted to just ask for the answer.
Practice card
Pick one thing you are currently stuck on — a practice problem, a concept from a course you are taking, a work question you are trying to actually understand rather than just resolve. Use the hint-ladder pattern first. Notice how much smaller a hint you actually needed compared to the full explanation you would have defaulted to asking for. The tutor-mode struggle-prompt card has all three patterns ready to copy, plus a log for which one worked best on which kind of problem.
