Asking a model to write a brand-new quiz question every time a student taps Practice sounds fine until you multiply it by a whole school doing daily revision. The cost adds up quickly, the wait between questions grows, and — worst of all — much of that generated content is near-identical to questions the model already produced last week. We wanted the variety of fresh AI questions without paying to regenerate the same ideas over and over.
Generate once, reuse widely
The solution is a shared question bank, organised by grade, subject, and topic. The first time a particular slice of the curriculum needs questions, the model generates a batch and they are cached. Every later request for that grade and topic is served from the bank. Across a school, the marginal cost of a practice session trends toward zero, and questions appear instantly.
The hard part: not repeating questions to a student
A shared bank only helps if it does not feel repetitive. So the bank is large and growing, and each student has a record of what they have already seen. When a student starts a session, we draw questions they have not encountered yet, mixing topics and difficulty so the set feels fresh even though nothing was generated on the spot. When a student is close to exhausting a topic, that is the signal to top the bank up — not to regenerate from scratch every time.
Quality stays in the loop
Because questions are stored rather than thrown away, they can be reviewed. Weak or off-curriculum questions can be retired so they are never served again, and the bank improves over time instead of producing the same mistakes repeatedly. A cache turns out to be a quality tool as much as a cost one.
The result is the experience students want — quick, varied practice that earns points, badges, and a daily streak — built on infrastructure a school can actually afford to run.