Methods · Generative practice & retrieval
B8
Delayed feedback
Attempt a whole set first, then review, so retrieval stays clean before corrections arrive.
When to use it
- You notice that instant corrections make you chase the AI’s next hint instead of thinking.
- You want exam-like conditions: a batch of items, then marking.
- Immediate feedback is tempting you to patch answers without finishing your own reasoning.
- You are practising a set where early answers should not leak into later ones.
What to ask AI
Type something like: “Give me five problems. I will submit all answers before you mark any. No hints until the batch is done.”
Ask it to mark first errors and patterns across the set, not rewrite every solution.
Keep the chat closed to solutions until your batch file is complete.
How it helps
Delayed feedback protects the retrieval attempt. You finish thinking before the answer key arrives.
Immediate feedback is powerful for fixing a single stuck step; it can also interrupt a stretch of independent work.
A batch-then-mark rhythm mimics tests and stops you from steering each item with mid-stream hints.
Use delay on purpose, not as neglect.
How to run it
- Agree on a set size and a no-feedback rule until submission.
- Attempt every item in writing without peeking.
- Submit the whole batch at once.
- Ask for patterns across errors first, then item-by-item notes as needed.
- Re-attempt the failed items later, optionally with immediate feedback on those only.
Example exchange
Please give me four medium chain-rule problems. I will submit all four before you mark any of them, no hints mid-set.
Here are four problems. I will stay silent on correctness until you send all four attempts.
Done. For 1 I got 6x cos(3x²). For 2 I dropped the inner derivative on (2x+1)⁵. For 3 and 4 I think I have the chain, but I am less sure about 4.
Pattern first: three of the misses share a missing inner-derivative factor. Item 2 also mixed up the outer function. Which one do you want to repair first under immediate feedback?
Copyable prompt
Give me [N] problems on [SKILL] at [LEVEL].
I will submit ALL answers before you mark any. No hints, partial solutions, or
corrections until the batch is complete. After I submit, summarise patterns
across errors before item-by-item notes.
The Tell
You will know the method has flipped when “delayed” still means you ask for a check after every line, or the AI breaks silence mid-set.
Then the delay never happened, and retrieval never stood alone.
Write a hard rule at the top of the chat and refuse mid-set markings until the batch is in.
Principle evidence
Strength of the underlying learning idea, not a claim about AI products.
Feedback timing research is nuanced: immediate feedback helps in many skill settings, while delaying feedback can protect uninterrupted practice and self-assessment. The point here is intentional delay to keep retrieval clean across a set.
AI delivery evidence
Whether an AI tutor delivers this method well is a separate question.
There is little specific research on AI tutors holding feedback to the end of a student-defined batch. The benefit depends on whether you and the model actually keep the rule.