Have the model write questions, not summaries

A summary is the most obvious thing to ask a model for, and for eight months it was the only thing I asked for. Learners read the summary, said it was clear, and remembered almost none of it a week later.

The fix was not a better summary. It was to stop generating one.

What the summary was doing

A summary moves the work of organising material from the learner to the model. That is exactly the work that produces retention, so handing it over removes the thing you were trying to cause.

• Reading a summary feels like understanding.
• Producing one is understanding.
• The model is very good at the half that does not help.

What replaced it

Now the model reads the source and returns questions — no answers, ordered easiest first. The learner answers from memory, and only then sees the source again.

Given the passage below, write 5 questions that
cannot be answered without having read it.
Do not include answers. Order them easiest first.

Two details that mattered

Questions that could be answered from general knowledge were worse than useless — learners answered them correctly without reading and concluded they were done. The "cannot be answered without having read it" clause removed most of those.

Ordering mattered too. Starting hard made people quit; starting easy made them continue into the hard ones.

The model should ask the questions. The learner should do the remembering.

Where it fails

It does not work on material the learner has no foothold in at all. With nothing to retrieve, retrieval practice is just failure, and a short explanation first beat any question I could generate.