What card-making actually teaches
Three things, and they're worth naming precisely so you can decide whether you're losing them.
- Selection. Deciding what's worth testing forces you to rank the material, which requires knowing what matters. This is the biggest of the three.
- Reformulation. Turning a statement into a question requires understanding it — you can copy a sentence without comprehension, but you can't reliably turn it into a question with one answer.
- Discrimination. Noticing that two items are confusable, and carding them together, comes from processing the material as a whole.
A generated deck supplies none of these unless you engage with the output. Which points at the actual answer: don't accept the deck, edit it. Culling and rewriting a generated deck exercises selection and reformulation directly, at maybe a fifth of the time cost of writing from scratch.
The honest comparison
| Handmade | Generated and accepted | Generated and culled | |
|---|---|---|---|
| Time per lecture | 45–90 minutes | 2 minutes | 10–15 minutes |
| Selection practice | Full | None | Most of it — you decide what stays |
| Reformulation practice | Full | None | Partial — you rewrite the weak ones |
| Card quality | Variable; depends on your skill | Variable; often list-shaped and duplicated | Good, because both passes filter |
| Coverage | Whatever you had energy for | Everything, including trivia | Everything important |
| Probability you actually have a deck by week 8 | Low, honestly | High | High |
That last row is the one that decides it in practice. The comparison isn't handmade versus generated; it's generated versus nothing, because the handmade deck for five modules across twelve weeks is a commitment most students don't sustain.
The cull is the technique
- 1
Generate, then delete before you read carefully
First pass is deletion only, at four seconds a card: anything with more than one answer, anything referencing the document, anything duplicated, anything you wouldn't care about forgetting. Expect to remove half to two thirds.
- 2
Split every list-shaped survivor
Any answer with commas in it becomes several cloze cards. This is where a mediocre deck becomes a good one, and it's genuine reformulation work.
- 3
Rewrite the phrasing of anything you'd have asked differently
If a question is technically fine but not how your course frames it, rewrite it. That judgement is exactly the selection skill the handmade route trains.
- 4
Add the three or four questions no generator writes
The application and failure questions — 'what happens if this step breaks', 'given this scenario, what follows'. Generators are weakest here and these are the highest-yield cards in the deck.
- 5
Verify anything numerical against the source
Specific values are where generation errors concentrate, and a wrong card learned confidently is worse than no card. This is why page-level citations matter — see AI flashcard quality.
When to write from scratch anyway
- The twenty hardest cards in a module. For material you genuinely struggle with, the making is worth the time and the generated version won't capture why it's hard for you.
- Anything where your course's framing is unusual. A generator works from the text; if your lecturer's treatment differs from the standard one, you'll notice and it won't.
- Your own error log. Cards made from your mistakes are the highest-value cards you'll ever have, and nothing can generate them — see how to learn from mistakes.
- Anything visual you should be able to draw. Redraw-from-memory cards have to be built by you.
Pre-made decks are a different question
Downloading a large shared deck for your subject is a third option and it has a distinct problem: it's someone else's course. A 20,000-card deck for a subject contains your syllabus plus a great deal that isn't, and the review time spent on material you'll never be examined on is real.
For standardised exams with a fixed syllabus — licensure tests especially — well-maintained shared decks are excellent and the argument reverses, because the syllabus genuinely is the same for everyone. For a university module taught by one lecturer with their own emphasis, a deck generated from your own slides tracks the exam far better. Thaxas vs Anki covers where each approach fits.
The evidence, such as it is
The generation effect — that self-produced material is better remembered than material you read — is well established, and it's the empirical basis for the handmade argument. It's a real effect and it's smaller than the effect of actually doing spaced retrieval at all.
So the ordering that follows from the evidence is: reviewing a generated deck beats not reviewing a handmade one, reviewing a culled deck beats reviewing an accepted one, and reviewing a handmade deck is best if you'll genuinely build and sustain it. Be honest about which of those you are, rather than choosing the theoretically optimal one and ending up in the first row of the table with no deck at all.