The AI writes the steps, never the answer

Every AI homework tool has to answer one question, and most of them answer it by accident: when a kid is stuck, does the AI do the work or does it make the work doable? We answered it on purpose, and then wrote the answer into the system so it can’t drift.

The rule in Vantee is one line: Vantee plans, asks and checks; the kid does the work. No essay, no project, no poster, no answer comes out of the model. What comes out is a plan, a question, and a check that the work happened.

Why “helpful” is the failure mode

A model that drafts the paragraph is, moment to moment, the most helpful thing in the room. The kid is unstuck, the assignment is done, the parent sees a finished page. Three things happened that nobody wanted:

1. The kid is exactly where they started. The hard part — the part where you don’t know what to write and you write anyway — got skipped.
2. The parent can’t trust “done” any more. Once the tool can produce the artifact, a finished artifact stops being evidence of anything.
3. The kid learns that being stuck is a thing you route around with a prompt. That one compounds.

The second point is the one that kills a product. A learning app’s whole value is that a parent believes its signal. The fastest way to destroy that signal is to let the AI be able to generate the thing it is supposedly measuring.

What “writes the steps” looks like

A goal arrives as a sentence — from a parent, a teacher, or the kid. “Get ready for the math competition in March.” “Read more.” Vantee turns it into a plan of small, dated, concrete pieces, and each piece names something the kid will do or make. A step says write the opening paragraph of your report. It never contains an opening paragraph.

That distinction is enforced in two places rather than trusted:

• The prompt that drafts a plan is explicit that a step describes work, not content — and a test asserts the prompt still says so, because a prompt is a file someone can edit on a Tuesday.
• A plain-language check runs over every generated step title. It exists for readability (our readers are 6-to-18-year-olds, many reading in a second language), but it also catches the drift where “steps” start swelling into content.

The deliverable is the kid’s. Finishing it is the kid’s own step, and nothing grades it with a model.

The part people find surprising: the AI doesn’t decide what you practice

This is where most of the engineering went. It would be easy — and it’s what we tried first — to hand a model the results of a quiz and ask what the kid should work on next. It reads well in a demo. It is also unaccountable: ask twice, get two answers, and no way to explain to a parent why their kid spent a week on fractions.

So the loop that decides what a kid practices is ordinary code. A check runs, and plain functions over its results pick the topics in focus, the topics already solid, and whether the goal is met. Same inputs, same decision, every time, and every decision traceable to an answer the kid actually gave.

The model’s job is narrower and it’s a job code is bad at: writing a question, a hint, a lesson, a sentence of encouragement in the right voice. Language, not judgment.

There’s a cost to this split and it’s worth naming. A deterministic loop can’t notice something clever that isn’t in its rules. We took that trade because the failure mode on the other side is a kid told they’re weak at something by a system that can’t say why.

Three rules that follow from it

A check steers, it never labels. Scores exist in the system; a kid mostly doesn’t see them. The words a kid gets are “practising now”, “learning now”, “got it”. We let a real number through only when it came from an objective check with a real answer key — never from a model’s impression of their work. A number a model guessed, shown to a 10-year-old as their number, is the single worst thing this kind of product can do.

Every judgment is reversible. The next piece of evidence overrides the last one, by design: an adult’s correction is scoped to one cycle rather than permanent, and a step marked done can be reopened. Nothing a kid does gets filed away under a lasting conclusion about them.

Credit goes to follow-through, not to hitting the target. A completed plan pays out because it was seen through. A check pays the same whatever the score. We want the kid who finishes a hard plan at 60% to feel better about it than the kid who abandoned an easy one at 95%.

What this costs us

Honest ledger. Saying no to “the AI does it” means saying no to the demo where a model writes a report in eight seconds, and that demo sells. It means a kid can still be stuck at the end of a session, because we won’t resolve being stuck by producing the answer. And it means our plans are only as good as the small set of rules behind them, which is a ceiling we’ll keep raising by hand.

What we get is a “done” that means something. A kid who got to the end of a piece of work knows they got to the end of it, and so does everyone else.

---

Vantee is a family AI assistant that turns goals into daily challenges kids run themselves. Have a look.