Cloze Cards for Language Learning
A word pair asks what a word means. A cloze card asks you to produce it inside a sentence, which is much closer to the thing you are actually training for. This guide covers what that buys you, what to hide, when cloze is the wrong card entirely, and how the two apps that support it differ.
What a cloze card is, and why it works
A cloze card is a sentence with a hole in it
- The term comes from 'closure': the mind's tendency to complete a pattern it recognises. Cloze tests have been used in language assessment since the 1950s, long before anyone called it sentence mining.
- The card asks a different question from a word pair. 'What does agotado mean' tests recognition. 'Estoy {{...}} despues del trabajo' asks you to retrieve the word in the shape the sentence needs.
- Retrieval is what builds memory, and the harder the retrieval, the more it builds, up to the point where you simply fail. A sentence with one gap sits close to that edge on purpose.
Context carries the half of a word a translation loses
- Collocation: which verb goes with which noun. No translation tells you that in Spanish you 'take' a decision rather than 'make' one, but a sentence does, silently.
- Grammar in use: the case, the aspect, the preposition, the article. These are the parts learners get wrong long after they know the vocabulary.
- Register: the same idea has a street version and an office version. A sentence you met in a film carries which one it was.
- A recall hook: a word learned inside a line you have heard has a scene attached to it. A word learned from a list has a position in a list.
Sentence mining is this idea with a workflow
- The rule that survived every variation of it: mine sentences where exactly one element is unknown. Everything else is comprehensible, so the new word arrives anchored.
- This is usually written as i+1, borrowed from Krashen's input hypothesis. Whatever you think of the theory, the practical version is unambiguous and easy to apply.
- It also explains the failure mode: a sentence with three unknown words is not a hard card, it is four cards pretending to be one, and you will fail all of them together.
None of this replaces the schedule that decides when you see the card. Cloze changes what a card asks; spaced repetition changes when it asks it. The guide to spaced repetition covers the other half.
Choosing what to hide
The sentence is the easy part. Which word you blank out decides whether the card is worth its place in your schedule.
One card, one thing. If the sentence teaches a verb, hide the verb, not the preposition next to it that happened to be interesting too. A card that tests two things fails for reasons you cannot diagnose.
A blank standing for half a sentence has no single correct answer, so you can never honestly mark it right. Nouns, verbs, adjectives and fixed expressions are the useful targets.
Hiding the dictionary form inside a sentence that requires an inflected one teaches you to produce the wrong shape. Hide the inflected form and you learn the word and its grammar in one pass.
The context has to narrow the answer down to roughly one word. 'Es muy {{...}}' could be a hundred adjectives; 'El cafe esta muy {{...}}, no puedo beberlo' is one.
If the hidden word appears again in the visible part, or if a synonym sits two words away, the card tests reading rather than recall. This is the single most common defect in generated cloze decks, and the reason to look at the sentences before exporting them.
You will see this card dozens of times. A line of dialogue is ideal; a subordinate clause pile from a legal text is a card you will start pressing 'again' on out of fatigue rather than ignorance.
When cloze is the wrong card
Cloze has become the default recommendation in language-learning communities, which means it is now applied to material it does not suit. Four cases where a plain pair is the better card.
Numbers, days, colours, kitchen objects, body parts. The mapping is one to one, there is no collocation to learn, and a sentence is pure overhead. Use word pairs and spend the saved time on more of them.
Cloze needs a sentence you can read. At A1 the sentence is the difficulty, not the gap in it. Start with pairs, add cloze once reading a simple sentence is no longer the hard part.
In a new writing system the bottleneck is recognising the characters, not producing a word in context. Cloze cards there measure your decoding speed and hide what you meant to test.
Before a trip or an exam with a published word list, breadth beats depth. Two hundred pairs reviewed once each will serve you better than fifty beautiful sentences.
Anki and Mochi do it differently
Both support cloze, and they disagree about what cloze is: a note type in one, plain syntax in the other. Knowing which is which saves a broken import.
Anki: a separate note type
- The marker is numbered: {{c1::word}}. The template renders {{cloze:Text}}, which is what turns it into a blank.
- A CSV carries none of that, so a cloze import under the wrong note type shows the raw braces on the card. A native .apkg carries the note type with it.
- Because cloze cards are sentences, and sentences contain commas, cloze text exports must be tab-separated or the sentence tears in half at the first comma.
Mochi: syntax inside the card
- No numbering: {{casa}} is enough, and several groups in one card are handled by Mochi's own rules rather than by c-numbers.
- It lives in the same Markdown as the rest of the card, so formatting, media and cloze coexist without a template change.
- A native .mochi package carries the audio inside it, and importing one needs no Mochi Pro subscription.
Either way, the sentence should speak
- The recording is of the sentence with the markers stripped, so you hear it complete, including the word you were asked to produce.
- Nothing needs installing on either side: no text-to-speech add-on in Anki, no subscription in Mochi.
- Generate the same set twice, once per platform, rather than converting one export into the other syntax.
The details of each live in their own guides: Anki cloze cards covers the note type, the c-numbering and the import traps, while the Mochi markdown and cloze guide covers the syntax and what Mochi's Markdown understands.
If the sentences should come from something you watched rather than from a topic, mine them first: turning subtitles into a deck walks that route.
Review load, mixing card types, and where sentences come from
Cloze decks fail for budgeting reasons more often than for pedagogical ones. These are the decisions that keep one alive past the first month.
A cloze card takes longer to answer than a word pair, roughly two to three times as long once you account for rereading the sentence. That means a deck of 200 cloze cards is not equivalent to 200 pairs in your daily budget; it is closer to 500. Decide the mix by minutes per day, not by card count, and cut the deck rather than the schedule when the minutes run out.
There is no reason a language collection should be all one type. The pattern that holds up: pairs for the long tail of concrete vocabulary, cloze for anything with grammar or collocation attached, and a small number of production cards for the words you actually want to say. Tag by type at import time so you can suspend a whole class of cards on a week when you have twenty minutes instead of an hour.
Anki lets one note hold c1, c2 and c3, producing three cards. It looks efficient and behaves badly: the schedules diverge, so you meet the same sentence on three different days with a different gap each time, and after the first pass the context is no longer new. The second and third cards end up testing your memory of the sentence rather than of the word. One marker per sentence keeps every card independent, which is why NextLang generates exactly one.
In order of value: a line from something you watched, a sentence from something you read, a sentence written for you at your level, a sentence from a textbook. The first two carry a memory hook the others cannot fake. Practically, that means mining subtitles and articles you have actually consumed, and using generation for the gaps where you have no source yet.
A cloze sentence with the sentence audio on the answer side reinforces pronunciation and prosody at no extra review cost. Move the audio to the question side and the same note becomes a listening card, where you hear the sentence and produce the missing word. That is the cheapest upgrade available to a cloze deck, and it requires a template edit rather than new cards.
Premium access includes:
Frequently asked questions
For anything with grammar or collocation attached, yes, at a higher cost per card. For concrete vocabulary with a one-to-one translation, no. The useful question is not which is better but which each word needs, and most collections want both.
Cloze deletion is the card format: a sentence with something hidden. Sentence mining is the habit of harvesting those sentences from real material you consume. You can do one without the other, and they are strongest together.
One. Two markers turn one note into two cards sharing a sentence, and the second card is mostly a re-read. If a sentence deserves two gaps, it is usually two sentences worth of material.
No. Anki uses a numbered marker inside a separate Cloze note type; Mochi uses double braces as plain syntax inside the same Markdown card. Converting a file between them by hand is where decks break, so generate each natively instead.
They can be generated, with one condition: look at them before they leave. A generated sentence is a first draft that is right most of the time, and the failure mode is a sentence that answers itself, which takes a second to spot and a moment to fix.
Keep reading
Related guides that build on what you just read.