What makes a tutor app different from a plain chatbot?

A chatbot answers questions; a tutor withholds the answer and guides the learner to it. The difference lives almost entirely in the system prompt — instructing the model to ask one question at a time, give escalating hints rather than solutions, and adapt difficulty to the learner. Same model, completely different learning outcome.

What is hint laddering?

Instead of revealing the answer when a learner is stuck, the tutor gives the smallest possible nudge first, then a bigger one, then a worked partial step — escalating only as needed. This keeps the learner doing the cognitive work, which is what actually produces learning. You implement it by telling the model to track how many hints it has given and never skip straight to the solution.

How does the tutor detect misconceptions?

Ask the model to classify a wrong answer — is it a careless slip, a missing prerequisite, or a specific known misconception? Tagging the error type lets you respond appropriately and, over time, surface patterns. A learner who keeps confusing two concepts needs targeted review, not just "try again".

Do I need to fine-tune a model for this?

No. A capable general model with a well-crafted system prompt handles tutoring across most school and early-university subjects. Fine-tuning helps only for highly specialised curricula or a very specific pedagogical style, and it is far more work than iterating on the prompt. Start with prompting and add retrieval for course-specific material before considering fine-tuning.

How do I keep the tutor from just giving the answer?

State it explicitly and repeatedly in the system prompt — "Never state the final answer. Respond with a question or a hint. If the learner asks directly for the answer, give a hint instead and explain why." Models default to being helpful by answering, so you must actively constrain that instinct, and you should test it by trying to extract the answer yourself.

How to Build an AI Tutor App

Building a Socratic AI tutor

The best human tutors rarely hand over answers — they ask the question that lets you find it yourself. A good AI tutor app reproduces that Socratic loop: it poses one problem at a time, judges your attempt, and offers escalating hints rather than the solution. Remarkably, almost all of this behaviour comes from the system prompt, not from any special model. This guide covers the prompt, the question-and-hint loop, and progress tracking — and the builder below generates the tutor system prompt for any subject and level.

How the tutor loop works

There are three moving parts. The system prompt defines the tutor: teach by asking, one question at a time, never reveal the full answer, adapt to the learner’s level and goal. The turn loop sends each learner answer plus the running history back to the model and asks it to do three things — judge the answer, respond with a graded hint if wrong or move on if right, and pose the next step. The hint ladder is the heart of it: the model gives the smallest nudge first and escalates only when the learner is still stuck, so the learner keeps doing the thinking.

Layer misconception tagging on top: ask the model to classify each wrong answer as a slip, a missing prerequisite, or a known misconception, and store that tag. Over many turns this gives you a map of what the learner has and has not mastered, which you can render as a progress view and use to revisit weak concepts.

Tips and pitfalls

The single hardest part is stopping the model from being too helpful — its default is to answer, so you must explicitly and repeatedly forbid revealing the solution, and then test by trying to extract the answer yourself. Always inject the learner’s level so difficulty calibrates. For course-specific material, add retrieval over the syllabus rather than fine-tuning. And render the generated questions and hints, not a canned answer key, so the learner experiences a real conversation. Use the builder below to produce a tutor prompt you can paste into any model API and start from there.