Most Flashcards Are Bad. Here's How to Write Ones That Actually Work.

I've watched people spend an entire weekend building flashcard decks only to abandon them two weeks later. Not because spaced repetition failed them, but because their cards were so frustrating to review that the whole system felt broken.

The scheduling algorithm gets a lot of attention. So does the app's interface. But the single biggest factor in whether flashcards work for you is the quality of the cards themselves. A perfect scheduler can't rescue a bad prompt.

The card that teaches you nothing

A friend of mine in med school showed me a card she'd made during her second year:

Front: "Describe the renin-angiotensin-aldosterone system."

Back: A four-paragraph explanation covering the entire cascade from juxtaglomerular cells to sodium reabsorption, plus clinical relevance, plus drug targets.

This card was useless. Every time it came up, she'd read the front, feel a vague sense of dread, mumble something about renin and angiotensin, flip it, skim the wall of text on the back, and hit "Good" because she recognized enough words to convince herself she knew it.

She didn't know it. She recognized it. Those are different things, and the card's design made it impossible to tell the difference.

One fact per card

The single most useful rule I've found is: one atomic fact per card. Not one topic. One fact.

That RAAS card should have been six or seven cards:

• Where is renin released from? (Juxtaglomerular cells of the afferent arteriole.)
• What triggers renin release? (Low blood pressure, low sodium at the macula densa, sympathetic activation.)
• What does renin do? (Converts angiotensinogen to angiotensin I.)
• What converts angiotensin I to angiotensin II, and where? (ACE, primarily in the lungs.)
• What are the main effects of angiotensin II? (Vasoconstriction, aldosterone release, ADH release, thirst.)
• What does aldosterone do at the collecting duct? (Increases sodium reabsorption and potassium secretion.)

Each of these can be answered in one sentence. Each requires genuine recall, not recognition. And each one either passes or fails cleanly. You either know where renin comes from or you don't. There's no bluffing your way through with partial pattern matching.

This feels tedious to set up, and it is. More cards means more upfront work. But the review experience is completely different. Instead of a 45-second slog through a paragraph, you get a two-second retrieval that either fires or doesn't. That's the kind of practice that actually changes what you remember on exam day.

The vague-prompt trap

Another pattern I see constantly is cards where the question is so broad that almost any answer feels correct.

Front: "What is important about the liver?"

Where do you even start? The liver does about 500 things. Metabolism, detoxification, bile production, protein synthesis, glycogen storage. Any of those could be "the answer," which means none of them is really being tested.

Compare that with: "What is the main exocrine product of the liver?" Now there's one answer (bile), and you either retrieve it or you don't. The card has a clear success condition.

This applies outside medicine too. Language learners write cards like "Uses of the subjunctive in Spanish" with a back containing eight bullet points. Better to make eight separate cards, each with a specific sentence where the subjunctive is required and a brief reason why. "Espero que ____ (venir) manana" with the answer "vengas, subjunctive because esperar triggers it in the subordinate clause."

Cards that test recognition instead of recall

This one is sneaky. The card looks fine on paper, but the way it's structured lets your brain cheat.

Front: "The powerhouse of the cell is the ______." Back: "Mitochondria."

You're not recalling anything here. You're completing a culturally famous phrase. The cloze is so obvious that the answer surfaces automatically, like finishing song lyrics. You feel productive, but the card hasn't tested whether you actually understand mitochondrial function, electron transport, or ATP synthesis.

A better card: "Where does oxidative phosphorylation occur?" or "Why do cells with high energy demands (like cardiac muscle) have more mitochondria?"

The test should require you to reconstruct knowledge, not pattern-match a fill-in-the-blank.

When images actually help (and when they don't)

That same med school friend went through a phase of screenshotting entire textbook diagrams and pasting them on the back of cards. The front would say "Draw the cardiac conduction pathway" and the back would be a full-color anatomical illustration.

Problem: she was never drawing anything. She was looking at the image, nodding, and moving on. The card was a recognition exercise dressed up as recall.

Images work well when they're the question, not the answer. Show a histology slide and ask "What tissue type is this?" Show an ECG strip and ask "What rhythm is this?" Show a map and ask "What country is highlighted?" In those cases, the image forces genuine identification.

Images on the back work when they genuinely clarify something text can't. A diagram showing the spatial relationship between heart valves is useful because spatial relationships are hard to describe in words. But a screenshot of a paragraph from a textbook is just a lazier version of typing out the answer.

The "I'll fix it later" problem

Bad cards compound. You add them during a busy week, intending to clean them up. You never do. They enter the review cycle and start accumulating scheduling history. Six months later you have 200 low-quality cards that the algorithm faithfully serves to you at optimized intervals, and every review session is a mix of good cards and garbage.

The fix is annoyingly simple: spend an extra thirty seconds on each card when you create it. Ask yourself two questions before you save: "Can this be answered in one sentence?" and "Will I know if I got it wrong?" If the answer to either is no, rewrite it before it enters the system.

The thirty seconds up front saves minutes per review for the lifetime of the card. It's the most leveraged time you can spend in the entire flashcard workflow.

What about cloze deletions?

Cloze cards (fill-in-the-blank) are fine when used carefully. They work best for specific facts embedded in context:

"The half-life of amoxicillin is approximately {{c1::1 hour}}."

The surrounding sentence provides enough context to make the question meaningful without giving the answer away. You need to recall the specific number, but the sentence reminds you what you're being asked about.

Where cloze cards go wrong is when people delete too much or too many things from one sentence:

"{{c1::Amoxicillin}} is a {{c2::beta-lactam}} antibiotic with a half-life of {{c3::1 hour}} that works by {{c4::inhibiting cell wall synthesis}}."

Now you've got four cards generated from one sentence, and some of them are terrible. The c1 deletion asks "what drug?" with almost no useful context. The c4 deletion is answerable by elimination if you've seen the other three. Each card is weaker than a dedicated question-answer pair would have been.

One to two cloze deletions per sentence, max. If a sentence contains four things worth remembering, write four separate cards.

Studying with other people's decks

Shared decks are a double-edged situation. On one hand, someone has already done the work of creating thousands of cards. On the other hand, you inherit their quality decisions, their phrasing, and their blind spots.

The biggest issue with shared decks is that you skip the encoding step. Writing a card forces you to process the material and decide what matters. Downloading someone else's deck bypasses that entirely. You end up reviewing prompts you don't fully understand, written in someone else's mental model.

This doesn't mean shared decks are worthless. Medical decks like AnKing are popular for good reason: they cover a standardized curriculum systematically. But treat them as a starting point. Suspend cards that don't make sense to you. Edit prompts that are vague. Add your own cards for material the deck doesn't cover well.

The worst approach is downloading a 30,000-card deck, unsuspending everything, and treating it as gospel. You'll drown in cards you don't understand, your retention will tank, and you'll blame the method when the real problem is that you're reviewing someone else's knowledge map instead of building your own.

Where AI fits into this

This is obviously self-serving since Imprimo uses AI to generate cards, but I think the nuance matters.

AI is good at decomposition. Feed it a page of notes and it can break the content into atomic facts reliably. It's good at generating the boring structural work: turning a paragraph into six question-answer pairs that each test one thing. This is the part of card creation that takes the most time and adds the least value.

AI is less good at knowing what matters to you. It can't tell that you already understand the RAAS cascade but keep forgetting the specific receptor subtypes. It doesn't know which exam is next week. It generates cards based on the content it sees, not the gaps in your memory.

The best workflow I've found is to let AI create the first draft and then spend a few minutes pruning. Delete cards covering things you already know cold. Edit cards where the phrasing doesn't match how you think about the topic. Add cards for connections the AI missed because it only saw one page of notes, not the full course context.

This is faster than writing everything from scratch, and the output quality is higher than what most people produce under time pressure during a busy semester. The key is that you stay the editor, not a passive consumer.

The minimum viable flashcard

If all of this feels like a lot of rules, here's the short version. A good flashcard has three properties:

It asks one thing. Not two, not a paragraph summary, one retrievable fact or connection.

It has a clear answer. You should be able to grade yourself in under three seconds. If you need to reread the back and think about whether your response "sort of" matched, the card is too vague.

It tests recall, not recognition. The answer should come from your memory, not from context clues in the question.

That's it. Get those three right and the scheduling algorithm, whichever one you're using, can do its job. Get them wrong and no amount of algorithmic sophistication will save you.

Your cards are the curriculum. The algorithm is just the delivery mechanism. It's worth spending your time on the part that actually determines what you learn.