The Grammar of a Hint

The zone is often quoted and rarely drawn honestly, so the figure above draws it: three bands, with the middle one — the only one a tutor can work in — between two moving edges. The edges matter more than the band. They move as the learner grows, which means yesterday’s hint is today’s insult: help pitched at a student who has outgrown it is not scaffolding but friction. A good tutor re-measures constantly. A good piece of software has no excuse for not doing the same, since the evidence of what the student can now do alone arrives with every answered question.

David Wood, Jerome Bruner and Gail Ross, studying adults teaching three- to five-year-olds to build wooden pyramids, gave the principle its sharpest form. The tutor’s moves should be contingent: the more the learner struggles, the more specific the help; the moment they succeed, control hands straight back 1. They called the tutor’s job scaffolding — not because the scaffold does the building, but because it holds just long enough for the builder to reach the next course, and is then taken down. The word has been stretched to mean “any support at all” in the decades since; the original was stricter, and the strictness is the content. Scaffolding that never comes down is a cage with good intentions.

What a hint is made of

So what is the unit of help? Not an explanation — the research on full explanations is bracing. A meta-analytic review of instructional explanations found them reliably weaker than their feel: an explanation that satisfies the explainer often fails to attach to the explainer’s model of the student, and learners frequently prefer and remember their own partial ideas over the cleaner sentence offered from outside 16. The hint, by contrast, is a small, aimed thing. Reading the tutoring studies, its grammar has roughly four moods.

Reorient — turn attention to the feature of the problem that matters, without naming any step: “what do you know about the angle at the centre?” Narrow — rule out a class of approaches: “integration is going to fight you here; try a substitution.” Step — give one move of the solution, the next move only, and let the student execute and continue. Path — lay out the whole route, the skeleton of the solution, without the working. Each mood reveals more; each costs the student more of the problem. Allan Collins and colleagues called the general family fading and contingency rules decades ago in their studies of human tutors 17; Kurt VanLehn’s synthesis of the intelligent-tutoring literature found the same moves in the machine tutors that approach human effectiveness 14.

The ladder in the hero plate arranges these moods by price. Its vertical axis — the probability that the student still finishes the step unaided — is modelled, not measured, but its slope is the honest shape: a reorientation barely dents it, the full path flattens it. This is why the order of the ladder is not negotiable. A tutor that opens with the path has skipped every rung where learning was still for sale.

The assistance dilemma

Put a number on the trade and you get the field’s central tension, which Koedinger and Aleven named the assistance dilemma: when to tell, when to tell conditionally, and when to wait 11. The two curves in the figure below are its whole story. With more assistance, performance during practice improves — students get further, feel better, cover more material. With more assistance, what survives to a delayed test declines — because practice under heavy scaffolding is not rehearsal of the skill but observation of it. The tutor who optimises the visible curve is spending the hidden one.

The dilemma is genuine — there is no setting that wins both curves — but it is not unmanageable. The resolution the literature converges on is sequencing: assistance first where knowledge is absent, then its deliberate withdrawal. That is the deep content of the next section.

Worked examples, and the discipline of fading

The strongest single result in this literature is also the least intuitive: for a novice, studying a worked solution can beat solving the equivalent problem. Sweller and Cooper ran the comparison on algebra: students who studied worked examples between problems learned faster and transferred better than students who solved every problem themselves 4. The mechanism is cognitive load. A novice wrestling with an unsolved problem spends most of a limited working memory on search — trying moves, backing out, trying again — and little on the structure of the solution itself. The worked example hands over the structure so it can be studied. John Sweller’s broader theory, cognitive load theory, grew from this result and remains one of the discipline’s productive engines.

But the effect has a famous shadow, the expertise reversal effect: the same worked examples that help novices hurt more experienced learners, who gain nothing from studying what they can produce and are slowed by the redundancy. Alexander Renkl’s studies of learners studying examples found the difference that matters inside this: those who explained the steps to themselves — why this move, why here — learned; those who read passively did not, and the self-explaining students were not the ones who spent longest 56. Michelene Chi’s original demonstration had made the same point from the other side: the quality of self-explanation predicted how much students took from the very same text 5.

Hence the discipline the figure above draws: fading. Begin with worked examples; move to completion problems, where most steps are given and the learner supplies one; withdraw more with each stage until the learner faces full problems — and withdraw on evidence, not on a fixed clock. Renkl and Atkinson formalised the schedule as a transition structured by developing ability 7. The point is not “examples are good.” It is that the ratio of given to asked is a dial, the dial should move, and the direction of its motion is set by what the learner can now do alone — the zone, again, in arithmetic form.

The debate that surrounded this literature — Kirschner, Sweller and Clark’s argument that minimally guided instruction fails novices, and Hmelo-Silver and colleagues’ reply that the successful forms of guided discovery are scaffolded, not bare — is best read as a fight over where on the fading schedule a classroom usually stands 89. Nobody in that argument believed a novice should face a blank page. Nobody believed an expert should be handed a full script. The fight is about the dial’s starting position, which is to say: it is the assistance dilemma wearing a different coat.

Where the model thumbs the scale

The honest limits here are the limits of any help.

First, help can be gamed — mostly by accident. A student under time pressure will click through hints until the answer appears, and a system that measures progress by completion will reward exactly that. The tutoring literature calls the pathology help abuse, and its studies are sobering: left unstructured, learners seek help too rarely when they need it and too cheaply when they don’t. The remedy is not to ration help but to price it — hints that are earned by an attempt, counted, and that hand control back at every rung. Corbett and Anderson found that even the locus of the feedback decision matters: students allowed to choose when to request feedback learned differently from those who received it automatically, and the difference itself depended on their competence 13. Help-seeking is not a failure mode; it is a skill with its own development, and a tutor that ignores it trains the wrong thing.

Second, the ladder must be climbed from the bottom every time. The rung a given student needs on a given step is an empirical fact, not a setting. The contingency principle says the help tracks the attempt, and the attempt is new information every time.

Third, the whole enterprise is bounded by the zone. No hint rescues a student from material they lack the prerequisites for; below the band, the honest move is to walk back down the prerequisite chain — a different act entirely, and one the machine is better at than the human, because it remembers what was never mastered.

What we read from it

EuraStudy’s tutor withholds by design — an earlier dispatch argued that handing over the answer is not teaching 18. The hint grammar is the constructive half of that argument. When a student asks for help inside a problem, the tutor does not explain; it climbs. The first rung reorients. If the student’s next attempt moves, the ladder stops — control handed back, per Wood, Bruner and Ross 1. Only on a second request does it narrow; only after that does it spend a step; and the full path is the last rung, not the first, priced visibly so that taking it is a decision rather than a reflex.

The reading from a half-century of this work is compact. Help has a geography — there is a band where help is possible, and it moves; measure before you assist. Help has a grammar — reorient, narrow, step, path; the moods differ in what they cost, and the costs are the point. Help has a schedule — examples before problems, and the scaffold fades on evidence, never on a timer. And the whole of it rests on the oldest line in the notebook: the next move should belong to the student. The tutor’s sentence is finished only when theirs begins.