AI Tutoring Is a Skills Problem, Not a Model Problem

Hand the most capable model on the market a student who is stuck on quadratics, and watch what it does. It solves the problem. Beautifully. Completely. With a clean derivation and a note about the discriminant. The student copies it down, understands nothing, and comes back tomorrow just as stuck. The model did not fail because it was dumb. It failed because it was helpful — and helpful, uninstructed, is the opposite of teaching.

That is the whole argument of this piece. AI tutoring is not gated by model intelligence. It is gated by behavior. A brilliant model with no teaching harness is a very expensive answer key. A modest model wrapped in a real pedagogy skill can run the loop that actually moves a learner — diagnose, question, scaffold, assess, adjust — and moving the learner is the entire job. The frontier you are waiting for already shipped. The thing missing is the instruction file.

Raw capability is not teaching ability

We keep conflating two things that have almost nothing to do with each other: knowing the answer and being able to teach it. The model knows the answer. It has always known the answer. Knowing more answers, faster, with fewer hallucinations, does not make it a teacher — it makes it a better encyclopedia, and nobody ever learned calculus by being handed one.

Teaching is a set of restraints as much as a set of capabilities. A good tutor withholds. A good tutor asks a question when it would be faster to just tell you. A good tutor watches for the misconception underneath the wrong answer instead of correcting the surface error. None of those behaviors emerge from scaling parameters. If anything they get harder to elicit, because a more fluent model is more tempted to fill the silence with a perfect explanation you didn't ask for and won't retain. Fluency is not pedagogy. Left ungoverned, fluency is the enemy of pedagogy.

Think about what the ungoverned smart model reliably does wrong. It over-explains — three paragraphs where one question would do. It hands over the answer at the first sign of struggle, because struggle reads as a problem to solve rather than the site where learning happens. And it almost never checks whether anything landed; it assumes comprehension and moves on. Those aren't intelligence failures. They are missing instructions.

Teaching is a loop, and the loop is a file

Real tutoring runs a loop, and the loop is old. Diagnose where the learner actually is. Ask a question that surfaces their thinking. Give them the smallest scaffold that lets them take the next step themselves. Check whether they can now do it unaided. Adjust and go again. This is the machinery behind the Socratic method, behind formative assessment, behind Vygotsky's zone of proximal development — the idea that you teach at the edge of what someone can do with help, not in the middle of what they already know or the fog of what they can't reach yet.

Here is the part that matters for anyone building on agents: that loop is expressible. It is a set of rules about when to do what. When the learner is confident and correct, get out of the way. When they're confident and wrong, ask — don't tell. When they're stuck and frustrated, scaffold; when they're stuck and fishing for the answer, hold the line. When a concept seems to click, test it before you believe it. Those conditionals are exactly the kind of thing a skill file encodes. A SKILL.md is not a personality. It is a policy for behavior under conditions — the tutor's version of a lesson plan that never gets skipped.

The tutor you actually want is not a bigger brain. It's a well-written procedure for when to question, when to scaffold, and when to assess — and a procedure is a file, not a frontier.

What the teaching skill actually encodes

Strip a good tutoring skill down and you find it does not add knowledge to the model. It adds governance. It converts a capable generalist into a disciplined pedagogue by specifying the moves — and, crucially, the restraints.

• Diagnosis before instruction. Open by finding out what the learner believes, not by explaining. The skill forces a question or a probe before any exposition — no lecturing into the void.
• Answer-withholding by default. The single most important rule, and the one raw models break instantly. A direct answer is the last resort, released only after the learner has done the reachable work. (This is the whole premise of a Socratic tutor skill.)
• Scaffolding, sized. Give the smallest hint that unblocks, then step back. The skill defines the ladder — from a nudging question, to a worked analogous example, to a partial solution — and climbs it only as far as needed.
• Assessment as a checkpoint, not an afterthought. Before declaring a concept learned, make the learner produce it unaided. The skill bakes in a "prove it back to me" step so comprehension is observed, never assumed.
• Adjustment on evidence. Route on what the learner just did. Confident-and-wrong, stuck-and-frustrated, and bored-because-it's-too-easy are three different states that demand three different next moves — and the skill branches accordingly.

That's the difference between a chatbot that happens to know things and a tutor. Not IQ. Instructions.

The behavior is portable; the model underneath is interchangeable

This is the part that should reframe how you think about the whole space. Because the pedagogy lives in a skill and not in the weights, it is portable. The same teaching behavior — the same diagnose-question-scaffold-assess loop — can be installed into a purpose-built study environment like DeepTutor, invoked inside Claude Code, dropped into Codex, or attached to any agent runtime that reads a skill file. Write the pedagogy once; run it anywhere the learner already is.

Swap the base model underneath and the teaching behavior survives, because the behavior was never in the model. That is why the interesting unit of progress in AI tutoring is not "which model" but "which skill." A curated bundle like education agent skills, or a study-focused kit such as tutor skills for StudyVault, is doing the real work — encoding the moves a good teacher makes and making them installable. The model is the engine. The skill is the driver's education.

Progress comes from the harness, not the parameter count

If the pedagogy is in the harness, then improving AI tutoring means improving harnesses — writing sharper conditionals, tighter answer-withholding rules, better diagnostic openers, cleaner assessment checkpoints. That work is legible, testable, and versionable in a way that "wait for the next model" never is. You can read a teaching skill. You can argue with it. You can A/B one scaffolding policy against another and keep the winner. You cannot do any of that with a parameter count.

None of this dismisses Bloom's finding that one-to-one tutoring pushes learners roughly two standard deviations above classroom instruction (Bloom, 1984). It reframes what closing that gap requires. The two-sigma result was never about the tutor's raw brilliance — plenty of two-sigma tutoring came from patient humans of ordinary intelligence running the loop well. The magic was the loop. And the loop is precisely the thing a skill can carry. This is why a whole category of these behaviors now exists as installable AI tutoring skills — because someone realized the teaching, not the thinking, was the missing file.

So stop waiting for the model that finally teaches. It isn't coming, because teaching was never a capability you scale into — it's a behavior you install. The smartest model on Earth, ungoverned, is a magnificent answer key. Wrap it in a skill that knows when to question, when to scaffold, and when to make you prove it back, and the same model becomes a tutor. The intelligence was always there. What you were missing was the instructions — and instructions are a file you can write today.

---

Part 2 of the AI Tutoring series. Previously: The AI Tutor That Refuses to Give the Answer. Next: Spaced Repetition Is the One Ed-Tech Idea That Survived. Browse the AI tutoring skills or more builder insights.