Good flashcards test one fact, in one direction, using material you have already met in context. Almost every complaint that flashcards do not work traces back to cards breaking one of those conditions: a card asking for four meanings at once, one that only ever runs in a single direction, or one copied from a word list you have never seen in use. Card design, not review software, decides whether a deck teaches you anything.
The minimum information principle#
Piotr Woźniak, who designed the SM-2 algorithm behind most spaced-repetition software, published a set of rules for formulating knowledge in 1999. Its core rule is the minimum information principle: formulate material as simply as possible, so each item tests a single, atomic piece of knowledge.
The reason is scheduling. A spaced-repetition system tracks how well you know each card and stretches its interval accordingly. A card containing four facts is scheduled as a unit, so the three facts you know cold are dragged back into review by the one you keep failing. Split it into four cards and each finds its own interval, three of which will quickly exceed a month.
Atomic also means unambiguous. If a card can be answered correctly in three ways, you cannot grade yourself honestly, and dishonest grading corrupts every interval the scheduler computes afterward.
| Weak card | Problem | Better version |
|---|---|---|
| tomar = to take, to drink, to catch | Three facts on one card | One card per sense, each with a sentence |
| Front: "the subjunctive" | Not a question, no single answer | Cloze one subjunctive verb in a real sentence |
| Front: llevar. Back: to carry/wear/take/spend | Ambiguous grading | Front: sentence with llevar; back: this sense only |
| Front: "German separable verbs" | Tests a whole topic | One card per verb, in a sentence showing separation |
| Front: word. Back: word + 40-word note | Note is never actually read | Word plus one short example, note trimmed |
Build cards from sentences you met#
The strongest card is a sentence you already understood with one unknown element inside it. This is the sentence-mining principle, and it rests on encoding specificity: Tulving and Thomson (1973) showed that retrieval is best when the cues at recall resemble the cues at encoding. A word encoded inside a sentence from a novel you were enjoying has the situation and the collocation available as cues. A word from a bilingual list has one cue, and it is the wrong one.
Sentence cards also solve a problem word cards cannot: they show which sense, which register, and which grammatical frame the word takes. Knowing a verb exists is not the same as knowing it takes a preposition or is only used in speech. A sentence carries that for free.
The practical consequence is that card creation should happen where reading happens. If saving a word requires switching apps, retyping the sentence, and finding a translation, you will do it for a week. That is the workflow LingoBlend builds around: read the text you chose, tap any blended word for its meaning and grammar, and it enters your dictionary and review schedule directly.
Cloze deletion#
A cloze card presents a sentence with one element removed and asks you to supply it. Ayer ____ (llegar) tarde a la reunión. It is the highest-value format for language learning because it tests production inside context, which is closer to what you actually want to do than recognizing a translation is.
Cloze cards are also naturally atomic: the deletion defines exactly what is tested, so nothing about a correct answer is ambiguous. One sentence can yield two cards from different deletions, rarely more.
Use cloze for verbs, prepositions, collocations, and endings, where the sentence supplies the constraint. Use a plain word card for a concrete noun with no interesting grammar. LingoBlend's fill-in-the-blank game applies the cloze format automatically to your saved words, using Tatoeba example sentences or AI-generated ones for Pro users.
Add an image and a mnemonic#
Two additions consistently improve recall, each costing seconds.
The first is an image. Allan Paivio's dual coding theory (1971) proposes that verbal and visual information are stored in separate but connected systems, and material encoded in both is easier to retrieve than material encoded in one. For concrete nouns and many adjectives, a picture is a second, independent route to the word, and it breaks the translation dependency, since an image does not route through your first language on the way to the target. We cover the mechanism in our post on dual coding for vocabulary.
The second is a personal mnemonic. Atkinson and Raugh (1975) tested the keyword method with Russian vocabulary and found it substantially improved recall: you link the foreign word to a similar-sounding word in your own language, then build a vivid image connecting that keyword to the meaning. The image needs to be yours and it needs to be strange; a mnemonic someone else wrote is much weaker, for the same reason a premade deck is. The full technique is in our guide to the keyword method.
Add both to the words that keep failing rather than to every card. The cost is real and most words do not need the help.
Test both directions#
A card that shows the target word and asks for the meaning trains recognition. That is useful for reading and is not the same skill as production. Recognition improves faster, which is why learners who only drill it are surprised when they cannot say anything.
Recall cards run the other way: you see the meaning, or the image, and produce the target word. They are harder, take longer to mature, and are the ones that make speech possible. Run both, weighted toward recognition early and shifting toward recall as your vocabulary grows.
| Direction | Front | Tests | Best for |
|---|---|---|---|
| Recognition | Target word | Passive understanding | Reading, listening, early stages |
| Recall | Meaning in your language | Active production | Speaking, writing, intermediate on |
| Image only | Picture, no text | Direct concept to word link | Concrete nouns, breaking translation habits |
| Cloze | Sentence with a gap | Production in context | Verbs, prepositions, collocations |
LingoBlend's flashcard game implements the first three as Recognition, Recall, and Immersive modes, so one saved word can be tested in any direction without separate decks.
Why premade decks underperform#
Downloading a 5,000-card frequency deck feels efficient and usually is not. Three findings explain why.
The generation effect, described by Slamecka and Graf in 1978, holds that information you produce yourself is remembered better than information you read. Making a card is itself processing: choosing the sentence, deciding what to delete, picking the sense that matters. Encoding specificity, again, means a card someone else wrote carries their context, not yours. And the self-reference effect, described by Rogers, Kuiper, and Kirker in 1977, found that material processed in relation to yourself is recalled better than material processed semantically alone.
Premade decks are still worth using in two situations: as a starting vocabulary base in the first weeks, before you have met enough material to mine, and for closed sets like numbers and months where context adds nothing. Import them, then let your own cards take over. Our guide to importing Anki decks and word lists covers the mechanics.
A short checklist#
Before a card enters your deck, check that it tests one thing, that the answer is unambiguous, that the sentence came from something you read or watched, and that you understood the sentence when you met it. Cards failing the last test are the ones you will still be failing in six months. Where cards come from matters as much as how they are built, which is why mining real material, whether TV dialogue or a book, is the habit worth building first. The feature list shows how the two halves connect.