The hint ladder, and why it belongs in the content rather than the prompt

A hint ladder is an ordered set of hints written for one item, from a nudge that only orients the learner up to a worked answer, each rung recording the condition under which it is offered and the condition that ends it. Work on intelligent tutoring systems calls that last rung the bottom-out hint. The ladder is content: it lives beside the item, is authored by whoever owns the material, and is versioned with it.

The alternative — telling a model in a prompt to give a graded hint without revealing the answer — reads correctly and behaves differently every time. Two learners at the same point get different amounts of help, the same learner gets a different ladder on a second attempt, and nobody can say what the product does, because what it does is a distribution rather than a behaviour.

The rungs, from orienting nudge to worked answer

Three or four rungs is the working range. Ten is a ladder nobody climbs; one is an answer button with a longer name. The rungs differ in kind rather than in strength: orienting, conceptual, procedural, complete. Whether rung 0 is phrased as a question at all is a real choice, argued in ask or tell: when Socratic questioning helps.

What a rung has to carry to be usable

• Item id and rung index. The ladder is meaningless detached from the item, and the index makes ordering a property of the data rather than the prompt.
• The text itself, authored and reviewed. This is the string that ships, and what a subject lead signs off.
• The prerequisite concept the rung assumes. A hint leaning on something the class has not covered is worse than no hint, and this field lets the tutor skip a rung it cannot justify.
• An entry condition. Requests so far, elapsed time, or an answer matching a specific wrong pattern — the last of which is why a ladder pairs with a misconception library you author on purpose.
• An exit condition. What ends this rung: a correct next step, a further attempt, or moving on. Vague exit conditions break a ladder for the reason vague level descriptors break marking — two people implement them differently and neither is wrong.

Authored rungs are testable; improvised ones are only sampled

This is the whole argument. A ladder of 4 authored rungs across 60 items is 240 strings: a fixed set that can be diffed when it changes and shown to a teacher before term starts. Behaviour tests become ordinary assertions — this input in this state returns rung 2, and rung 3 is unreachable until the attempt count is met.

Where generation genuinely helps

Authoring 240 rungs by hand is why teams reach for the prompt. The productive compromise is to generate drafts and route them to an author: give the model the item, the worked solution, the prerequisite list and the known misconceptions, and have it propose 4 rungs in the house voice for a subject lead to edit. Generation also fits rewriting an approved rung at a different reading level, since the meaning is fixed first.

What generation should not do is decide at answer time how much to reveal. Two related behaviours sit outside this page: what happens when learners work the ladder for the answer, in students who press hint until the answer falls out, and the deliberate withdrawal of support, in fading, and when a tutor should stop helping.

Building the ladder as data — an authoring surface, a review step, versioning with the course material, selection logic the tutor executes — is work we scope under AI agents and automation, with the operational realities in what running AI agents in production takes. This page sits inside AI tutoring and learning support, part of our education and edtech work.

• tutoring
• instructional design
• learning support
• content model