A good flashcard asks for one fact, has one right answer, and is about something you already understand. Split lists into one card per item, add a short cue when two cards look alike, and rewrite any card you keep failing instead of grinding it. These rules come from Piotr Wozniak, who built the first spaced-repetition software, and they hold whether you write cards by hand or let AI draft them.
Why card design matters as much as the schedule
A spaced-repetition scheduler can only time reviews well if each review has a clear outcome: you remembered the card, or you did not. A card that asks for five things gets half-remembered, and a half-remembered card has no honest rating. It comes back too often, takes too long, and still slips. Most of the frustration people blame on the algorithm is really card design.
The rules below are a working subset of Wozniak's 20 rules of formulating knowledge, written for SuperMemo in 1999 and still the basis of most advice on the subject.
1. Understand it before you memorize it
Flashcards are for keeping what you have learned, not for learning it the first time. If you cannot explain a mechanism in your own words, cards about its details will feel arbitrary and fail often. Read the chapter or sit the lecture first, then make cards from what made sense.
2. One fact per card
Wozniak calls this the minimum information principle, and it matters more than any other rule. Each card should test one thing, so the answer is short and the rating is honest.
| Too much on one card | One fact per card |
|---|---|
| Describe the renin-angiotensin-aldosterone system. | Which enzyme converts angiotensin I to angiotensin II? (ACE) |
| What are the requirements for a valid contract? | Besides offer and acceptance, what must each party give for a contract to bind? (Consideration) |
The big question is still worth answering, just not on one card. Break it into the steps or elements, one per card, and the whole comes back when you need it.
3. Ask a question with one right answer
A card is ambiguous when two answers are both correct, or when you cannot tell from the front what kind of answer is wanted. “What about beta blockers and asthma?” could be answered several ways. “Why are non-selective beta blockers avoided in asthma?” has one answer: blocking beta-2 receptors can narrow the airways. Avoid yes-or-no questions too; you can guess them half the time.
4. Split every list
Lists are the most common way a card gets too big. Five elements of a tort on one card means five chances to forget, and one rating for all of them. Make one card per item instead, each asking for that item in context: “Which element of negligence requires that the breach caused the harm?”
If the order of a list matters, as in a sequence of steps, add cards that ask what comes after a given step. Keep a single “name all of them” card only if the exam really asks for the full list from memory, and expect it to be your hardest card.
5. Tell look-alike cards apart
Two similar cards interfere with each other: you learn one and start mixing up the other. Medicine is full of these, such as drugs with similar names, and so is law, with the same term defined differently in two areas. Add a short cue to the front that points at the right one, for example the subject or the context, and consider a card that asks for the difference directly.
6. Use images when the image is the question
Wozniak rates pictures highly, because we remember images better than words. They earn their place when you have to recognize something by sight: a structure on an anatomy diagram, a histology slide, an ECG. Put the image on the front and ask what it shows. An image that only decorates a text card adds nothing.
7. Rewrite cards you keep failing
A card you fail again and again is often called a leech. Pressing Again for the tenth time rarely fixes it, because the problem is usually the card. Look at why it fails: too big, too vague, or confused with a look-alike. Then split it, sharpen it, or delete it if you do not really need the fact.
What to check in AI-made cards
AI is good at the mechanical part of these rules. Given a page of notes, it will write short, one-fact cards faster than you can. What it cannot do is know your exam, so check three things before your first review:
- Is it right? Cards built from your own material are far safer than cards from a model's general knowledge, but a fact can still come out wrong. Fixing it now takes seconds.
- Is it clear? Rewrite any card whose question you cannot answer without seeing the source it came from.
- Do you need it? Delete trivia from your notes that no exam will ask. Every card you keep costs reviews for months.
How Cadence Deck applies these rules
Cadence Deck writes cards from the notes or PDF you give it, one fact or idea per card, in one of three styles: Recall for short factual cards, Concepts for why and how questions, and Balanced for a mix of the two. You see every card before it is saved and can edit or delete any of them, and each side of a card can hold one image. Reviews are then scheduled by FSRS, which brings a card back just before you would forget it.
Further reading
- Effective learning: twenty rules of formulating knowledge: Wozniak's original article, with many more examples.
- Improving Students' Learning With Effective Learning Techniques: the review that rates practice testing and spaced practice as the two most useful study techniques.