Contingent hinting

Get unstuck without getting rescued.

When to use this: While you practise
Learning impact: Highest-leverage move
Learning evidence: Strong research tradition
Where you are with the topic: Building understanding
How often to use it: Worth repeating often

When to use it

You have twenty minutes tonight and need real practice, not more highlighting.
Homework is done, but you still want one honest attempt before bed.

What to ask AI

Type something like: “I'm working on problems and I want contingent hints, not solutions. Rules:”.

Tell it to ask one question at a time and wait for your answer before continuing.

Tell it your current level and ask it to stay one step harder than comfortable.

Keep the hard part on your side: you retrieve, attempt, explain, or decide before you ask for help.

How it helps

They get back a clear, well-structured, entirely correct solution.

You read it, follow every line, and think: yes, obviously. Then they attempt the next one and are stuck in exactly the same place.

This is the most common way AI makes learning worse, and it is invisible from the inside, because the experience of reading a good solution is genuinely pleasant and feels like progress.

Something did happen: you now understand that solution.

What didn't happen is the thing that would have let you do the next one.

The mechanism you skipped is the struggle.

Being stuck, and working out what to try, is not a regrettable preliminary to learning, for a large class of material, it is the learning.

How to run it

Attempt first, in writing. Non-negotiable. A hint before an attempt is just a smaller solution.
Say what you've tried and where you stopped. This is what makes the hint contingent instead of generic.
Ask for rung 1. Almost half the time, "is my approach reasonable?" is enough, you were on the right track and stalled on confidence, not method.
Attempt again between every rung. The rung is useless if you don't use it before asking for the next.
Stop climbing the moment you're moving. Don't take the remaining hints because they're available.
Log which rung you needed. Over a fortnight, the pattern tells you something: consistently needing rung 4 on one topic and rung 1 on another is a precise map of where you're actually weak.

Example exchange

You

I'm stuck on this related-rates problem. Here's my attempt so far, please don't give the answer. Smallest hint only.

You've set up related rates, which is the right type. What quantity is changing that you haven't named yet?

You

Oh, the radius is changing too, not just the height.

Good catch. Write the next equation yourself, then come back if you stall again.

Copyable prompt

I'm working on problems and I want contingent hints, not solutions. Rules:
never give me a complete solution; give the smallest hint that would unstick me
and then stop; if I haven't shown you an attempt, ask for one before helping.
If I ask for the answer, remind me of these rules and give me a hint instead.

The Tell

Here is how you know this method has flipped: you finish a problem set having been unstuck every time and couldn't start any of them again tomorrow. You were helped through, not taught.

In that moment the AI (or the schedule, or the story) did the thinking, and you only recognised a finished product.

Tighten the prompt for contingent hinting so you attempt, decide, or retrieve before anything is handed to you.

Principle evidence

Strength of the underlying learning idea, not a claim about AI products.

The underlying learning idea is rated strong. Contingent tutoring, help calibrated to you’s last move, in the smallest useful dose, is a classic finding from human tutoring research (Wood, Bruner & Ross, 1976, on scaffolding; later ITS/tutoring process literature). It sits with scaffolding and the zone where you can almost succeed: support that enables success then withdraws. Related: assistance dilemma / help-abuse in intelligent tutoring, more help is not always better. AI contingent hinting inherits that principle strongly; whether current chat models deliver calibrated contingency reliably is still speculative.

AI delivery evidence

Whether an AI tutor delivers this method well is a separate question.

The learning principle may be strong, but that does not prove an AI session delivers it well. Treat AI delivery as speculative unless a study of tutoring with this method is named, and none is claimed here.