Study & Learning
Writing AI Flashcards That Actually Work
By Jim Vernon, Editor, AI Intelligence International · Published 3 March 2026 · Reviewed against our editorial standards · About the author
Generating a thousand flashcards takes about four minutes. Drilling a thousand bad flashcards takes about forty hours and teaches you very little, which is why card quality matters far more than card volume.
This article sets out the specific properties of a card that produces durable memory, and the prompt structure that gets a model to produce them reliably instead of producing paragraphs with question marks attached.
Key takeaways
- One card, one fact: A card that asks you to list four causes of something is really four cards fused together.
- The answer must not be recoverable from the question: Generated cards frequently leak.
- Cloze deletions and when to use them: Fill-in-the-blank cards work well for definitions, sequences and formulas where the surrounding structure is itself worth knowing.
- Prompting for card sets: Give the model the source text rather than the topic name.
One card, one fact
A card that asks you to list four causes of something is really four cards fused together. You will get three right, one wrong, and mark the whole card as failed, which sends the three you knew back into the review queue for no reason.
Atomicity also makes failure informative. When a card fails, you want to know exactly which fact is missing, not that some part of a cluster is shaky.
The instruction that works: ask explicitly for atomic cards, one testable fact each, and state that compound answers should be split into separate cards rather than joined with commas.
The answer must not be recoverable from the question
Generated cards frequently leak. A question phrased around the same distinctive vocabulary as its answer can be solved by pattern recognition, and pattern recognition disappears the moment the exam rephrases anything.
Test each card by asking whether someone who had never studied the topic could guess it from context. If they could, the card is measuring English, not knowledge.
Ask the model to avoid reusing rare words across the two sides, and to prefer questions that require producing a term rather than recognising one.
Cloze deletions and when to use them
Fill-in-the-blank cards work well for definitions, sequences and formulas where the surrounding structure is itself worth knowing. They work badly for concepts, because they let you rely on sentence rhythm.
A good rule: use cloze when the blank is the only thing you would have to produce in a real setting. Use plain question-and-answer when the whole idea needs reconstructing.
Never blank more than one item per sentence. Multi-blank cards degrade into the compound problem described above.
Prompting for card sets
Give the model the source text rather than the topic name. Cards built from a topic name reflect the model's general knowledge, which may not match your syllabus; cards built from your notes reflect what you will be examined on.
Specify the count, the format and the exclusions. Something like: twenty atomic cards from the text below, question and short answer separated by a pipe, no compound answers, no questions answerable without reading the source, no trivia about names or dates unless the text emphasises them.
Ask for a difficulty spread. A set that is entirely easy feels good and moves nothing; a set that is entirely hard causes abandonment. Roughly a third recall, a third application, a third edge cases works well.
Pruning the set
Expect to discard a quarter to a third of generated cards. This is normal and much faster than writing the survivors by hand. The pruning pass is where you also catch factual errors, so it is not wasted time.
Delete anything you already know cold. A card you get right every time is costing review minutes and providing nothing.
Merge near-duplicates immediately. Generated sets tend to circle the same idea from three angles, and three cards on one fact distorts your sense of coverage.
Review discipline
Grade honestly. The temptation to mark a nearly-right answer as correct destroys the scheduling algorithm, and the algorithm is the only reason spaced repetition beats re-reading.
Keep sessions short and daily rather than long and weekly. Twenty minutes a day sustains a card base of several hundred comfortably; two hours on Sunday does not, because the intervals collapse.
Retire the deck when the exam passes. Carrying dead cards forward is the main reason people abandon spaced repetition entirely.
Worked example: a pharmacology chapter
Source: eight pages covering drug classes, mechanisms and common side effects. First pass generates sixty cards from the raw text with the prompt structure above.
Pruning removes eleven duplicates, six cards where the drug name appeared in both question and answer, and four with mechanisms that contradict the chapter. Forty-one survive.
The surviving set splits naturally: fifteen name-to-class cards, fourteen class-to-mechanism, twelve side-effect associations. Drilled at expanding intervals, this reaches stable recall in about nine days at fifteen minutes a day.
The eleven-minute pruning pass was the highest-value part of the exercise, because every error caught there would otherwise have been memorised as fact.
Frequently asked questions
How many flashcards should one chapter produce?
After pruning, thirty to fifty for a dense technical chapter and ten to twenty for a discursive one. If a chapter yields two hundred, most of them are testing detail that will not be assessed.
Should the AI write the answers too?
Yes, but verify them against the source. Writing answers by hand is the slowest part of card creation and the part where generation helps most, provided you keep the verification step.
Are image cards worth making?
For anatomy, diagrams, maps and anything spatial, yes — text descriptions of visual material are poor substitutes. Generate the question text with AI and attach the image yourself.
What is the biggest mistake with generated decks?
Keeping everything. An unpruned deck buries the cards that matter under duplicates and trivia, and the review time it consumes is the time you needed for practice papers.