AI Tutor: How to Set Up ChatGPT or Claude as Your Personal Teacher

Why tutor mode doesn't switch on by itself

In the general guide on learning with AI, I already covered why a model defaults to solving a problem for you instead of walking you toward the answer: it was trained to be maximally useful right now, not useful for your long-term retention. Here's an expanded version of that same prompt: variants for different subjects, and what to do when the model still slips back into handing you finished answers.

The baseline template is one prompt for every subject, but it only holds up until the first non-standard case. From there come the questions that didn't fit in the general guide: how to run a tutor-mode conversation on a language, where there's no single correct answer, what to do when the model stubbornly plays examiner instead of helper, and how to tell when the prompt has simply stopped working.

Tutor mode for a specific subject

The same baseline prompt behaves differently in math, programming, and languages, because these subjects have a different structure to what counts as a "correct answer."

Math and the exact sciences. There's a single correct result here, so tutor mode works almost perfectly out of the box: the model can honestly say "incorrect" and point to the exact step where the logic went wrong. The one adjustment worth making to the baseline prompt: explicitly ask it to show intermediate steps separately rather than the full derivation at once, otherwise the model sometimes hands over a partial solution that isn't formally the answer but functionally is.

Programming. Here tutor mode runs into a temptation that doesn't exist in math: ask for pseudocode or a plain-language description of the algorithm as the first step, rather than code, and only after you've stated your own approach, ask the model to review it like a code review, rather than write the implementation from scratch. That keeps AI in the role of reviewing your thinking, not authoring the solution.

Languages. There's no single unambiguous correctness criterion here, and the baseline prompt needs its own adjustment: ask the model not just to fix mistakes, but to explain why a native speaker would phrase it differently, with an example from natural speech. Without that adjustment, the model usually just rewrites the sentence correctly without explaining what made your version sound unnatural, and the lesson doesn't stick.

What to do when the model still hands you a finished answer

Even with a working prompt, the model periodically slips back into "just answer it" mode, especially in long conversations, where a system instruction gradually loses weight against the accumulated context of the conversation.

Three working techniques for when this happens. First: directly remind it of the rule with one short phrase ("go back to tutor mode, don't give me the finished answer") instead of rewriting the whole prompt from scratch. Second: start a new conversation with the same system prompt, if reminders have stopped working, because a long context is sometimes easier to reset than to clean up. Third, the most reliable: lock the prompt in as a system message or a dedicated project, rather than pasting it in by hand at the start of every conversation, because manual entry is exactly what tends to get lost partway through a long session.

Another technique worth keeping in reserve if the first three don't work: explicitly ask the model, at the start of the conversation, to reread the system prompt and confirm it understood the constraint before answering the first question. It sounds excessive, but the extra second spent confirming costs less than ten minutes of a conversation where the model, on autopilot, just hands over finished answers, and you only notice once progress stops feeling real.

Worth a separate note on models that follow the "don't give a finished answer" instruction too literally and start asking leading questions even where you're asking for a direct fact ("which command stops a container in Docker" isn't grounds for a Socratic dialogue). A useful clause to add to the prompt itself: tutor mode applies to problems there's something to learn from, not to factual lookups that simply have a short correct answer.

A personal example: where I broke my own prompt

I wrote the first version of this prompt too strict: I told the model to never give a direct answer, even if I explicitly asked for one. Within a week I realized this was inconvenient in exactly the cases where I genuinely needed a quick fact in the middle of working on a real task, not a from-scratch teaching moment. The model honestly followed the instruction, and instead of answering "what's this HTTP status code called," it started asking what I already knew about response codes. I had to rewrite the rule so an explicit direct request would override tutor mode, not the other way around. The prompt has had a separate clause for exactly this exception ever since.

When tutor mode doesn't help

If tutor mode is set up correctly but progress still feels noticeably slower than usual, the likely cause isn't the prompt, it's that you're stuck at a specific difficulty level or practicing the wrong skill. That's a separate diagnostic problem, and the piece on dealing with learning plateaus covers how to tell a genuine plateau apart from a prompt that just needs retuning, and what to ask the model to get a specific answer instead of a generic "just keep working harder."

Tutor mode is a tool for a specific situation: when you have a topic and need a conversation partner who walks you toward the answer instead of just handing it over. It doesn't replace the other techniques from the general guide, spaced repetition, the Feynman technique, working with a study plan, it just determines the mode the model answers in for every question you ask within those techniques.