Skip to content

AI exercise generation

A lesson made only of theory (no exercises) can be turned into a practiceable lesson by having the AI generate exercises from its cards. This is the EXP-036 pipeline shipped in v1.90.0. It needs an AI key configured (Settings → AI); without one the button is visible but disabled, with the reason in its tooltip.

The pipeline

Generation is not a single AI call — it is a generate → quality-gate → balance → feedback pipeline:

  1. Generate. A generation prompt asks the model for exercises across the supported types; a defensive JSON parser tolerates the usual model formatting quirks.
  2. Quality gate. A deterministic gate rejects exercises that are malformed, trivial, or duplicate existing ones — before you ever see them.
  3. Balance. The set of generated exercises is balanced across exercise types so a lesson is not all of one shape.
  4. Regenerate with feedback. If the result is not right, you can regenerate with feedback to steer the next attempt.

Per-lesson and per-set

  • Single lesson: a "Generate exercises" button appears on theory-only lessons.
  • Whole set: batch generation fills in exercises across every theory-only lesson in a set in one run.

Quality and trust

Because the quality gate is deterministic, generated exercises meet the same minimum bar as authored ones (enough exercises, more than one type, no empty cards). Generation augments authored content — it never silently overwrites your existing exercises.

For checking existing content quality (authored or generated), see AI content check.