Review

This file reads the literature on how flashcards are written, made, used and kept, strand by strand. Whether retrieval and spacing work is settled in the core review and is not repeated. "(secondary)" marks a number read in an abstract or a summary. Grey scores are in sources, and a source under 10 of 20 is "example only".

0. Existing reviews

Nakata 2011 scored nine flashcard programs on 17 criteria from paired-associate research; most do little to raise retrieval effort, and no shared design guideline exists (secondary). The author's handbook chapter was not opened (Nakata 2019b).
Webb 2020 pooled 100 effects from 22 vocabulary studies of flashcards, lists, writing and fill-in-the-blanks: delayed meaning-recall gains averaged 39.4%, and direction of learning moderated gains (secondary).
Versteeg 2020 (scoping, 120 articles) finds spaced learning in health professions poorly defined (secondary).
Xu 2024c (systematic) finds flashcards popular and correlated with grades; Öcel 2025 (25 studies) finds Quizlet studies positive (secondary).
Nguyen 2021 and Teymouri 2024 are narrative reviews of Anki and digital flashcards; Kornell 2007, Bjork 2013 and Kang 2016 review study decisions and spacing (secondary).
Lei 2022 synthesised 32 word-card studies by effect-size plots, without pooling: more showed larger effects for ready-made than self-made cards, and for paper than digital (secondary); no review found compares card formulations.

1. Seminal work

Wozniak 1999 gives twenty rules: understand first, minimum information, cloze, imagery, avoid sets and enumerations, combat interference, optimise wording, context cues, sources and dates.
It cites no study; its numbers are the author's records, such as one item claimed to save 25-fold.
Slamecka 1978 named the generation effect, memory for what the learner produces over what is read (described through Bertsch 2007).
Watkins 1975 explained proactive inhibition as cue overload: a cue shared by many items retrieves each less well (title only; Nairne 2002 develops the principle).

2. Card formulation (Q1)

Practitioner rules converge: one idea per prompt, one correct answer, no yes or no prompts, no lists, context named (Matuschak 2020; Nielsen 2018; Bjornstad 2020).
Matuschak 2020 adds that prompts must be effortful and tractable, and says research gives no exact guidance.
The Anki manual says to understand before memorising and to split cards that hold too much (Anki 2026a; Anki 2026c).
Direction: first language then second was more versatile when both production and comprehension were needed (Griffin 1996, secondary).
Pictures: Swahili words were not learned better from pictures than from translations until overconfidence was removed (Carpenter 2012, secondary).
Pictures on flashcards helped concrete words, and simple pictures hindered abstract ones (Ide 2026, secondary).
Card elements: word-formation notes went with vocabulary growth, in a model explaining 15.8% of variance (Reynolds 2020, secondary).
Cloze form: fill-in-the-blank sentences and flashcards gave no overall difference at two weeks, 150 learners (Kim 2023, secondary).
No comparison of cloze and question-and-answer cards inside a spaced program was found; a Q/A answer reports the same gap (LLSE 2017b).
Cue strength: a vague hint leaves room for retrieval and an obvious one turns recall into recognition, by a reading of the literature (PNSE 2015).

3. Who makes the cards (Q2)

Pan 2023b, six experiments, 48-hour test: learner-made beat premade cards, d = 0.45 for memory and 0.29 for application (secondary).
Generating beats reading by about 0.40 over 86 studies (Bertsch 2007, secondary); the size depends on how constrained generation is (McCurdy 2020, secondary).
Text generation beat reading, g = 0.41, not explained by time on task (Schindler 2023, secondary); learners choose cues that suit their own memory better than observers do (Tullis 2022, secondary).
Against: premade word cards may be better (Wilkinson 2020, secondary), ready-made cards gave more recall than self-made ones (Sage 2019, secondary), and the word-card synthesis leans the same way (Lei 2022, secondary).
Deng 2015 (full text, 72 of 111 students): Anki cards, which the authors class as user-made, predicted licensing scores, about 1 point per 1,700 cards; commercial premade cards did not (p = 0.763).
Deng 2015 is a retrospective survey, and its authors ask for a randomised trial before a causal claim.
Lu 2021 (full text, 201 respondents): at least 80% used premade decks and 57.4% made cards; users scored higher, d = 0.38.
Teacher-made cards: one third of nursing students used them, and users scored higher, d = 0.42 (Ingebrigtsen 2025, secondary).
More teacher-made cards reviewed went with higher physiology grades (Santos-Ferreira 2024, secondary); a class that crowdsourced 16,000 questions outscored the class before (Bow 2013, secondary).
Premade psychiatry cards were used by 56% and did not predict exam scores (Sun 2021b, secondary).
Practitioners favour own decks for complex subjects, shared decks as a supplement (Anki 2026a; Nielsen 2018).

4. How learners use cards (Q3)

Dropping cards had small but consistently negative effects on learning, in four experiments (Kornell 2008, secondary).
Learners remove an item once they can recall it, which leads to poor retention (Karpicke 2009, secondary).
Dropping after one correct recall saved trials and lowered final scores (Pyc 2011, secondary).
One large stack beat four small ones, which mass the reviews; spacing won for 90% of learners, and 72% believed the opposite (Kornell 2009, secondary).
Learners self-test to diagnose, not to learn (Kornell 2009b, secondary).
Of 374 students, most understood practising to a higher criterion but not longer lags (Wissman 2012, secondary).
Of 901 students, digital flashcard use only partly followed the evidence (Zung 2022, secondary).
About 70% of 415 students used flashcards, and use on all three exams went with higher scores (Golding 2012, secondary).
With a partner, learners dropped fewer cards; alone, their global judgments ran about 20% above their test (Imundo 2025, secondary).
Required Anki use went with better grades, and with reluctance and little enjoyment (Seibert Hanson 2020, secondary).
Anki marks a card a leech after 8 lapses and suspends it; its manual's first remedy is to rewrite the card (Anki 2026c).

5. Granularity, interference, understanding (Q4)

No study of the minimum information principle was found under that or a related name (protocol, F7 and F8).
With spacing equal, studying in small or large sets differed little; spacing mattered more (Nakata 2016, secondary).
Five or seven retrievals beat one or three, but per minute one retrieval gained most (Nakata 2017, secondary).
Semantically clustered words were harder to learn (Tinkham 1993; Tinkham 1997, secondary).
Related words raised confusion errors without lowering scores (Nakata 2019, secondary); words whose referents look alike interfered, semantic relation alone did not (Ishii 2015, secondary).
Nation 2000 advises against teaching lexical sets together (opening read in a quotation, LLSE 2018b).
Conceptual cards beat detailed cards for less able learners only, and flashcards overall did not beat free study (Lin 2018, secondary).
Cards written for comprehension and application raised exam scores in three studies of 73, 62 and 434 students (Senzaki 2017, secondary).
Nielsen 2019 turns one proof into dozens of cards over a few hours, by his own account.

6. Text with embedded prompts (Q5)

Quantum Country embeds 112 questions in an essay of about four hours' reading (Matuschak 2019).
Six months after release, 195 users had shown a month of retention on at least 80% of the cards.
After about 95 minutes of review, demonstrated retention rose from about 2 to about 54 days per card.
Informal test: regular review moved accuracy from 89% to 96% (16 users), a two-week delay from 91% to 87% (25 users); no essay-without-prompts control.
"Mnemonic medium" returned no record in ERIC or arXiv; the nearest formal work, adjunct questions, is held by the core.

7. Generated cards (Q7)

Model-made Anki cards held 1 error per 21 cards (4.8%); users' exam scores did not differ, and use was self-selected (Khurana 2025, preprint, secondary).
Generated contextual cards in Anki against usual teaching gave d = 2.48 in 60 learners, cards and scheduler confounded (Karrar 2026, secondary).
Experts judged generated MOOC cards accurate (Bachiri 2025, secondary); generated keyword cues were tested against hand-made ones (Lee 2023, abstract only).
Students' stated and observed preferences for generated mnemonics disagreed, over 2,684 preferences from 45 students (Balepur 2024, secondary).
Six students wanted generated cards editable and transparent (Li 2026, secondary).
No randomised comparison of generated and written cards on delayed recall was found.

8. Practice and adoption (Q6)

Repository stars: Anki 32,379, AnkiDroid 12,116, the FSRS add-on 4,119, Orbit 1,838 (GitHub 2026h, read 2026-10-29).
In 2020, 68.3% of 560 students from 102 US medical schools used Anki (Halperin 2024, self-selected, secondary).
Daily Anki users were 56% of 165 at one school (Wothe 2023, secondary), and about 70% of 60 used it at another (Harris 2022, secondary).
Anki users scored higher in one more single-school cohort (Gilbert 2023); in another, heavier users scored higher on average, not significantly (Levy 2023, 45 students) (both secondary).
Of 57 Ottawa respondents, 98.2% used unofficial resources, student-made Anki decks among the most common (Makus 2023, secondary).
Quizlet's own page refused the request; an uncited figure gives 60 million monthly users in 2021 (Wikipedia 2026e, example only).
Guides add cost rules: memorise what would cost over 5 or 10 minutes to look up (Branwen 2019; Nielsen 2018).

Q/A sites

Language Learning, Psychology and Neuroscience, Mathematics Educators and Academia Stack Exchange (read 2026-10-29): 19 threads, 2 under 10 of 20; opinion, no data.

Hiding the word trains production, hiding the meaning trains comprehension; study both as goals require (LLSE 2016a).
Split a word with several meanings, or give it a short sentence (LLSE 2018a); put grammar into cloze sentences (LLSE 2016h).
Flashcards give a shallow, context-free meaning; combine them with reading (LLSE 2016b).
Leeches come from confusable pairs and misunderstanding; separate the pair in time (LLSE 2016c; LLSE 2016e).
An old card costs little per review, yet 20,000 cards at 10 seconds is about 55.5 hours over a few years (LLSE 2016d); daily use and backlogs cost most (LLSE 2017a).
Premade parts that help recognise a word are fine; parts that help use it should be personal (LLSE 2016f).
Mathematics is compressed by problems and mental models more than by prompts (MESE 2014; MESE 2024).
Universities leave memorisation to students (AcSE 2016; AcSE 2024); sharing a set is fair, questions first (AcSE 2023, example only).
Examples only: a new card appears before it is learned, so learn it first (LLSE 2016g); a reversed sibling only shortens intervals (PNSE 2016).