Anki for language learning

Anki does one thing extremely well: it decides which card to show you today, based on your history with that specific card, so that you review each item at roughly the moment you are about to forget it. Everything else about it — the interface, the deck options screen, the add-on ecosystem — ranges from adequate to actively hostile.

That trade is worth making. The scheduling is the product, and it is the reason spaced repetition works at all.

Most language learners who try Anki stop within a month. Almost none of them stop because of the algorithm.

Why decks die

Three failure modes, in order of frequency.

The cards are word pairs. hablar → to speak. This card is fast to make and teaches a translation, which is not the thing you need. It does not teach that hablar con alguien takes con, that the stress falls on the second syllable, or what the word looks like inside a sentence. You end up able to translate a list and unable to say anything — the understand-but-cannot-speak problem with extra steps.

The deck was downloaded. Someone else’s 5,000-card deck contains 5,000 items you have no relationship with. Card creation is itself a learning event — deciding what the item is, where its meaning stops, and how to cue it. Skipping it removes half the value and all of the context.

Twenty new cards a day. This looks ambitious in week one. By week six the daily review count is 150 and climbing, because reviews accumulate behind new cards. Miss four days, come back to 400 due, and the deck becomes a source of guilt rather than a tool.

Card design

Sentences, one unknown each

The unit is a sentence containing exactly one thing you do not know. Everything else in it should already be solid.

Front:

I’ve never been there.

Back:

Je n’y suis jamais allé. /ʒə ni sɥi ʒa.mɛ za.le/ 🔊

That card teaches the word, the structure it lives in, the pronunciation, and the liaison — four things, one card, one review.

Compare the word-pair version: jamais → never. One thing, and not the useful one.

Production direction, mostly

A card that shows the target language and asks for the meaning tests recognition. A card that shows the meaning and asks for the target language tests production. The second is harder and closer to what you actually want to do with the language.

A reasonable mix for one item:

Card type	Front	Tests
Production	English meaning	Can you say it
Recognition	Target sentence	Can you understand it
Audio	Sound only, no text	Can you hear it

Three cards per sentence is heavy. Two — production and audio — is a good default, and audio-only cards are dramatically underused. Many learners can read a language far beyond what they can hear, and an audio-only card is the only card type that catches this.

Audio on everything

If a card has no audio, you are learning the word from spelling, and you will learn a wrong pronunciation you then have to unlearn. This is severe in English, French, and any language with opaque orthography, and it is the single most common reason a learner’s vocabulary is unusable in speech.

Phonetic transcription where it earns its place

Not on every card. On the cards where spelling lies.

rouge /ʁuʒ/ Aussprache /ˈaʊ̯sˌʃpʁaːxə/ though /ðəʊ/

For a learner who knows the handful of IPA symbols relevant to their own problem sounds, this is a precise instruction. For a learner who does not, it is decoration — so introduce four or five symbols first, tied to their actual difficulties.

Cloze for grammar

Grammar does not belong on a translation card. It belongs in a cloze deletion:

Ich {{c1::bin}} nach Berlin gefahren.

This trains the choice you actually have to make in speech, in the context where you make it. Full-sentence translation cards for grammar produce recall of the sentence, not the rule.

Settings worth changing

The defaults are mostly fine. Four are not.

New cards/day: 5–10. The most important number in the entire application. This governs your review load six months from now.

Maximum reviews/day: 100–200. A cap prevents a bad week from producing a 400-card day that makes you quit.

Use FSRS. Anki’s newer scheduler. It fits the intervals to your actual performance and typically produces meaningfully fewer reviews for the same retention. Turn it on: Deck options → FSRS.

Bury related cards. Stops the production and audio card for the same sentence appearing in one session, where the first gives away the second.

Leave everything else alone. The deck options screen is where enthusiasm goes to die.

Reviewing properly

Answer out loud before you flip. If you flip while thinking “yes, I know this”, you got exposure and skipped retrieval, which is the entire mechanism. Say the sentence. Out loud. Then flip.

Only two buttons matter. Again if you did not produce it, Good if you did. Hard and Easy invite fine-grained self-assessment you cannot perform accurately, and overusing Hard creates cards that never mature.

Reviews before new cards. If you have ten minutes, do reviews and add nothing. Reviews are the compounding part; new cards are the part that can wait.

Delete cards you hate. Every deck accumulates a few items you have failed fifteen times. They are usually badly made — ambiguous cue, two unknowns, a word you have no context for. Delete or rewrite them. A card you dread poisons the whole session.

When you fall behind

You will. The recovery matters more than the prevention.

Do not try to clear a 500-card backlog in one session. Instead:

1. Set new cards to 0 for two weeks
2. Cap reviews at your normal daily number
3. Let it drain

Or, if the backlog is months old, delete it and start again with the cards that still matter. A 300-card deck you open daily is worth more than a 4,000-card deck you avoid.

The real friction

Everything above assumes the cards exist. They mostly do not, because making a good card is genuinely slow: find the sentence, confirm the translation, get audio, get the transcription, format it. Two to four minutes per card, honestly, which is why decks are abandoned in week three.

This is the specific bottleneck Teachee removes for lesson material. A sentence captured during a lesson already carries its transcript, translation, phonetic transcription and audio, and exports directly to an Anki deck — so the cards are built out of language you personally reached for and failed to produce, which is the highest-value material available and the hardest to assemble by hand.

Platform notes

Platform	Cost	Notes
Windows / Mac / Linux	Free	Where you make cards
Android (AnkiDroid)	Free	Full featured
iOS (AnkiMobile)	~$25 one-off	Funds the project; worth it
AnkiWeb	Free	Sync, plus browser review

Make cards on desktop, review on the phone. Reviewing on a phone means it happens in queues and waiting rooms, which is the difference between fifteen minutes a day and zero.