Ask for a task they can start, not finish

Kapur (2014) found that students who wrestled with a problem before any teaching outperformed students taught first, but only when the task let them generate partial solutions and the lesson afterwards connected their attempts to the canonical method. The design constraint is precise: startable with current knowledge, unfinishable without the new idea. That is fiddly to write by hand and quick to generate with a clear prompt. (Manning's productive-failure-desirable-difficulty-designer skill exists for exactly this.) For the classroom sequencing itself, see "Let them struggle first, then teach".

In practice

Concept I will teach next lesson: [concept]
What students can already do: [prior knowledge]

Design a 20-minute problem for them to attempt
BEFORE any teaching:
- Solvable in part with what they already know.
- Not fully solvable without the new concept.
- List 3 flawed approaches they will likely try,
  and what each one gets right.
- Give me a consolidation script comparing their
  approaches to the canonical method.

• Give the AI the target concept and what students already know, then ask for a problem partially solvable with that prior knowledge but not fully.
• Ask for the two or three flawed approaches students will likely invent, so you can name and compare them during the consolidation.
• Keep chatbots closed during the struggle phase; if students can ask for the answer, the generative work never happens.
• Budget as much time for the compare-and-contrast afterwards as for the struggle; the learning lands in the consolidation, not the failure itself.