Auditing your cards and deleting the bad ones

Six months into a deck, there's a small group of cards doing real damage. They come up constantly, they get graded Again, and they eat a share of every session out of all proportion to their number.

The learner assumes those words are just hard. Usually the words are fine and the cards are broken.

A card is a small piece of writing. Like any writing, some of it is badly made. A badly made card costs you every time it comes due, forever, until you change it or remove it.

Nobody schedules time to look at that, because reviewing feels like progress and editing feels like admin. So the bad cards stay.

Here's an audit you can run in one sitting. Set aside 20 minutes. You're looking for a specific, small number of cards, and you're going to do one of three things to each one.

Audit only the cards you've failed four or more times

Don't read your whole deck. Most of it is fine and reading it all is how the audit dies at card 60.

Sort by failures. In Verbamor, open Library and look at what keeps returning in Study with short intervals; in Anki, use the browser search prop:lapses>4.

You want cards you've failed at least four or five times. That's a small group. In a deck of 800 cards it's usually 20 to 50.

Anki formalises the failure-count idea. Its manual sets a default threshold of 8 lapses, at which point "Anki tags the note as a leech and suspends the card." A lapse is one time you forget a card you had learned.

Eight is late. By the eighth failure you've spent perhaps five minutes on one word across several months. Four failures is early enough to act and late enough to be a real signal.

One caution before you start. A card you've failed four times in the last week differs from a card you've failed four times over six months. The first is a new card you're still learning, and it's behaving normally.

Skip anything younger than a month. Age the sample first.

Write the list down, or keep it open in a second window. Now test each one.

Four tests, run in order, find why a card keeps failing

Run the four tests in order. The first one a card fails is the reason it keeps failing.

Test 1: does it ask for more than one thing? Read the card and count the separate pieces of information you must produce to get it right. If the answer is more than one, the card is overloaded.

Piotr Wozniak, who wrote the SuperMemo algorithms, states the rule directly: "Simple questions formulated for active recall in learning bring much better memory outcomes than complex questions even though one complex question may be equivalent to a large number of simpler questions."

His reason is mechanical. With a simple item "we can rely on a simple memory connection." Complex memories "may have their concepts activated in an incomplete fashion, or in a different sequence that depends on the context."

A card asking for a verb, its preposition and its irregular past tense fails in three independent ways, so it fails often.

Test 2: is the answer guessable from the prompt? Cover the answer. Look only at the front. If the sentence, the image or the length of the blank gives away the word, the card is testing your reading.

A card like "Necesito ___ dinero" with the answer "más" is fair, because the blank could hold half a dozen words. A card reading "El accidente fue ___" with the answer "accidental" hands you the answer in the first three words.

Images do this too. If the picture on a card for "cuchara" shows a bowl of soup with a spoon standing in it, the picture answers the question before you do.

Test 3: does something else in your deck compete with the card? Say the answer out loud, then ask what other card in your deck also wants that answer. Two cards that both take "to take" as their English side will fight.

Interference between similar items is a documented cause of failure, not a personal weakness. Anki's own manual gives the example of a learner confusing "disappoint" and "disappear". It recommends suspending one of the two cards until the other is well learned, then unsuspending it.

Test 4: do you know what the word means? Answer in plain English, right now, without flipping. If you can't, you've been memorising a shape rather than a word.

Wozniak's first two rules of formulation are "Do not learn if you do not understand" and "Learn before you memorize." A card can't teach you a meaning you never had.

Most bad cards fail test 1 or test 3. Those two account for the majority of what you will find.

Give every audited card one outcome: fix, split or delete

Every audited card gets exactly one of three outcomes. Deciding is the whole point of the audit, so don't leave a card in an unresolved state.

Fix when the word is worth knowing and one element is wrong. Rewrite the example sentence so it stops giving the answer away. Change the image so it points at this word and not a general scene.

Narrow the English side from "to take" to "to take, as in taking a bus". In Verbamor, edit the card in Library; the audio is regenerated for the new sentence, so the recording and the text stay in agreement.

Split when the card fails test 1. One card asking for three things becomes three cards asking for one thing each.

Here's a real overloaded card. Its front reads "to remember (French)" and its back "se souvenir de / je me souviens / souvenu". It asks for the verb, the reflexive construction with its preposition, and the past participle.

Split it into three: one for the verb in a sentence, one for a sentence that needs the "de", one for a sentence in the past. Each of those can be right or wrong on its own.

Splitting feels like it makes your workload worse. It makes it better within two weeks, because three simple cards each reach a long interval, while one triple card keeps resetting to the start. Wozniak's list makes the same recommendation for enumerations: "avoid lists, sets and enumerations (they can be tackled with cloze deletion)."

Delete when the word costs more than the seat it occupies is worth. Learners avoid this outcome, and it pays the fastest of the three. Anki's manual states the case plainly, advising deletion when "the material you're struggling with isn't important enough to justify the time spent."

Ask one question: will I hear or say this word in the next three months?

Delete a regional food term from one podcast episode, or a technical noun from a lesson about your job in a language you speak socially. Delete a word you saved because it looked pretty.

A deck is a claim about what you intend to use.

Watch the shape of your own decisions as you go. If you're deleting nothing, you're being sentimental about words you'll never say.

If you're deleting almost everything, you're using the audit to avoid hard words, which is a different problem, and one it doesn't solve. Most audits should produce some of all three outcomes.

Judge the audit by your review count four weeks later

Your review count drops the day after the audit. Your accuracy goes up. Both are real and neither is the interesting number.

The number to watch is your daily review count four weeks later. Bad cards generate reviews at a rate healthy cards don't, because each failure schedules another review soon.

Verbamor uses FSRS, an open-source algorithm that predicts when you're about to forget each card. It tracks three values per card.

Retrievability is "the probability that the person can successfully recall a particular piece of information at a given moment". Stability is "the time, in days, required for R to decrease from 100% to 90%". Difficulty is "the inherent complexity of the information."

A card you keep failing is a card whose stability won't grow. It stays near the front of your queue permanently. Twenty such cards can produce more monthly reviews than two hundred healthy ones.

Run the full card audit every three months. A quarterly pass gives each card enough review history to prove itself, and it keeps the editing habit from eating the studying habit.

Between audits, use one rule instead: when a card fails three times in a single week, edit it on the spot or delete it, then keep going. That single rule catches most new bad cards before they need an audit at all.

Expect the deck to feel too easy for about a week after the audit. That feeling is worth naming.

Koriat and Bjork (2005) showed that the sense of competence you get while the answer is available is unreliable. A freshly cleaned deck produces exactly that sensation.

Judge the audit at four weeks, on your review count and on whether the words show up when you speak.

When should you skip a card audit?

Skip the audit if your deck is under three months old. You don't have enough failure history to tell a bad card from a new one, and you'll delete words you were about to learn.

Skip the audit if you're in a bad stretch and behind on reviews. Auditing while a backlog sits there converts a review problem into an editing problem, and you'll finish neither. Clear the queue first, then audit.

Skip the audit if you build every card from a lesson recording and keep them narrow already. Cards mined from audio you actually heard tend to pass test 4 automatically, because you met the word in a real context with a real meaning attached.

The failure mode there is different: too many cards. That's a selection problem, and you fix it by keeping fewer words per lesson.

Do this much today. Sort by lapses and take the worst ten cards. Run the four tests on each, and give every one of them a fix, a split or a delete before you close the app.

Ten is enough to see whether your deck's problem is the words or the cards.