Writing Good Flashcards: 20 Rules with Examples

Most flashcards don't die from the algorithm. They die from themselves: too long, too vague, with five answers at once. After three weeks you click "Again" without knowing which part was missing, and eventually you simply stop opening the deck.

The good news: writing good cards is a craft, and that craft is well documented. The most important source is a text by Piotr Wozniak, the inventor of SuperMemo, from 1999: "Twenty rules of formulating knowledge." Add two essays everyone seriously learning with flashcards should read: Andy Matuschak's "How to write good prompts" (December 2020) and Michael Nielsen's "Augmenting Long-term Memory" (July 2018).

A candid sentence on evidence, because you rarely read one elsewhere: the 20 rules are accumulated experience, not individually tested laboratory findings. What is well supported is that retrieval practice works at all. A meta-analysis by Yang et al. (2021) covering 222 studies with 48,478 learners found a mean advantage of g = 0.499 for tests and quizzes over their respective control conditions in classroom settings. That figure does not tell you which card format is best, however.

The 20 rules in one sentence each

Here are all 20 rules in Wozniak's order, in plain English, with what each means in practice for you.

1. Don't learn what you don't understand. A card about something you haven't understood only trains memorizing word chains.
2. Learn first, then memorize. Read the chapter, understand the connection, then write cards.
3. Build on foundations. The simple concepts first, the details later.
4. Minimum information principle. Ask as little as possible per card. More on this below.
5. Cloze deletions are simple and effective. Cloze cards are quick to make and they work.
6. Use images. Where a picture says more than a sentence, use the picture.
7. Use mnemonics. Memory aids belong on the card, not only in your head.
8. Image occlusions work as well as text cloze. Hidden image areas function like cloze.
9. Avoid unordered sets. "List all …" is a trap.
10. Avoid enumerations. Long sequences are hard to review.
11. Fight interference. Similar cards disrupt each other.
12. Optimize wording. Shorter, clearer, unambiguous.
13. Connect to other memories. Link new material to what you already know.
14. Personalize and give examples. Your own example sticks better than the textbook one.
15. Use emotional states. What surprises or annoys you is easier to remember.
16. Context cues simplify wording. A short prefix like "Pharma:" saves you half a sentence.
17. Redundancy does not contradict the minimum information principle. Asking the same knowledge from two directions is allowed.
18. Cite sources. So you can look things up later.
19. Date your information. For knowledge that can become outdated.
20. Prioritize. Not everything deserves a card.

If you only take four away, take 1, 4, 9 and 20. The rest is refinement.

Minimum information: the rule that decides everything

The minimum information principle is often misunderstood. Wozniak does not mean "make as few cards as possible." He means the opposite: each individual repetition should demand as little as possible at once. You break complex answers apart.

Why? If one card asks three facts and you forget one, you grade the whole card as wrong. The scheduler brings it back soon, even though you could confidently recall two thirds. And you don't even know which third was missing. Michael Nielsen describes exactly this in his essay: he breaks a question about creating symbolic links in Unix into smaller sub-questions and writes that this makes it clearer which component is missing when you get something wrong. He calls them atomic questions.

The first card is the before: one paragraph as the answer, where you forget something different every time. The three cards after it are the after. Each has exactly one correct answer. (The examples in this article deliberately use content from the article itself, so you can try them directly.)

Two important caveats from the essays. First: "atomic" does not mean "trivial." Nielsen stresses that an atomic question may presuppose demanding concepts. It is atomic because it is clearly delimited relative to what you already know. Second, Matuschak says that "focused" depends on prior knowledge. A familiar combination may be one unit for an advanced learner, while a beginner needs several cards for it. There is no universal rule such as "at most one word per answer."

And Nielsen has one more trick: you may additionally keep the original big question as an integrative question. Parts first, then the composition.

Matuschak's five properties of good cards

Andy Matuschak thinks of cards as small tasks you perform repeatedly, which should change your understanding. For him, good retrieval prompts are:

• focused: one thing per card,
• precise: no question that several answers fit,
• consistent: the same correct answer every time,
• tractable: you can realistically answer them,
• effortful: you really have to pull something from memory.

Many people overlook the last point. Hints on the card are allowed, but they must not give away the answer. A card where you can derive the answer from the question feels good and accomplishes almost nothing.

The first card repeats the answer in the question, so you only read it off. The second requires you to recall the name of the principle yourself.

Rules 9 and 10: lists and enumerations

"List all …" is the classic bad card. Wozniak says: avoid unordered sets, and if unavoidable, order them. For sequences you cannot avoid, he recommends cloze deletions with heavily overlapping gaps, meaning several cards in which a different part of the sequence is missing while the rest stays visible as context.

Matuschak suggests two approaches: either you ask about individual missing elements, or you ask targeted questions about the components. Which path fits depends on how precisely you must know the list. For an exam where you must enumerate every element, you need the overlapping gaps. If you only need to know what a single element does, the individual question is enough.

Cloze or basic?

The Anki manual describes the card types that checkst also understands when you import a deck:

• Basic: front, back, one direction of recall.
• Basic (and reversed card): two cards from one note, once forward, once backward.
• Basic (optional reversed card): the reverse direction only if you fill in a specific field.
• Cloze: you hide text passages, in Anki with the syntax {{c1::answer}}. An extra field shows additional information on the back.

The reverse variant is not a luxury. For vocabulary, the manual distinguishes recognizing a foreign word from producing it. These are two different skills, and that is exactly what Wozniak's rule 17 is for: redundancy is allowed.

A practical rule of thumb: cloze for sentences where exactly one term is the core (definitions, statutes, mnemonic sentences). Basic for genuine questions with why or how. What you will not find is proof that one type is fundamentally superior to the other. The meta-analysis by Rowland (2014) found larger benefits when learners had to recall freely during practice rather than only recognize. Yang et al., by contrast, found no significant difference between these categories in classroom settings. No fixed ranking of card types can be derived from this.

The first is a classic fill-in-the-blank, the second a why-question. Both are atomic, but they train different things.

Images and image occlusions

Anki has by now built in rules 6 and 8: since version 23.10, image occlusion is native. You hide areas of an image and quiz them. There are two modes: with "Hide All, Guess One" all areas are hidden and one is revealed. With "Hide One, Guess One" only the quizzed area is hidden. The manual names anatomy and geography as typical use cases.

An honest note: according to the brief, there is no clean comparison showing that image occlusions are better than equivalent text cards. Image occlusions make sense when the knowledge is genuinely spatial, such as where a muscle attaches.

Rule 11: interference

The two image occlusion modes above are a perfect example: "Hide All, Guess One" and "Hide One, Guess One" look almost identical, and if you have two separate cards for them, you will confuse them sooner or later. Wozniak says: recognize the confusions and sharpen the cards. In practice, if you keep swapping two cards, write a card that asks the difference directly, and make the two originals more distinguishable.

Rule 14: your own examples, your own cards

Here there is actually controlled data. Pan et al. compared self-created with premade digital flashcards in six online experiments with 284 students in total, at equal time budgets, with a final test after 48 hours. The self-made cards were better: d = 0.45 for definition questions and d = 0.29 for application questions. Compared with heavily incomplete templates, the advantage for definition questions was even larger.

The limits belong here too: these were short-term studies, not long-term use over a semester. "Quality" there mainly meant correct and complete. And the study does not test any of the 20 rules individually. Still, it is a good argument against adopting premade decks blindly and against saving AI-generated cards unchecked.

Rules 18 to 20: sources, dates, priorities

These three sound bureaucratic, but they save you trouble. A source on the back ("lecture handout slide so-and-so") helps when you doubt something three months later. A date helps with knowledge that becomes outdated, such as guidelines or legal status. And prioritizing means: not every sentence in the lecture handout deserves a card. Whoever cards everything soon has a review mountain that stops being fun.

Revise cards instead of rewriting them

Matuschak recommends revising cards while you review them. This is the most underrated advice of all. Most bad cards are not recognized while writing them, but on the third repetition, when you get annoyed by them. That is exactly the moment to shorten, split, or sharpen.

And checkst?

checkst schedules your reviews with FSRS, understands basic and cloze cards, and imports Anki decks as .apkg. If you have the AI suggest cards from a lecture handout, you choose basic, cloze, or mixed and see every card before it is saved. That is exactly where these rules help: throw out any card that asks more than one thing, and rewrite one or two yourself. According to Pan et al., that is worth the effort.