Position

This file says what is known about writing and keeping flashcards, what is not, and what this review adds. Each "not known" below is a statement about this search, not about the world.

Against the existing reviews

  • Nakata 2011 scores software features, not card wording; Webb 2020 pools activities, not card forms.
  • Lei 2022 counts word-card studies by who made the cards and finds ready-made ahead, without pooling and only for vocabulary.
  • Versteeg 2020, Xu 2024c and Öcel 2025 ask whether spacing, recall or one tool helps; none compares how cards are written or who writes them.
  • This review adds: card formulation, learner-made against ready-made decks, learners' use and abandonment, the mnemonic medium, generated cards, and adoption counts, in one pool.

Known

  • Learners drop cards once they can recall them, and dropping lowers retention (Kornell 2008; Karpicke 2009; Pyc 2011).
  • Learners split decks into small stacks and cram, against their own results (Kornell 2009; Hartwig 2012).
  • Learner-made cards beat premade cards under matched time in five of six experiments, for memory and, in most, application (Pan 2023b). Yet two vocabulary studies and a synthesis of 32 word-card studies lean the other way (against).
  • Generating beats reading in general, about 0.40, with large variation (Bertsch 2007, secondary; McCurdy 2020).
  • Related words studied together cause confusion errors, though not always lower scores (Nakata 2019; Ishii 2015).
  • Cards written for concepts or application help some learners more than detail cards (Lin 2018; Senzaki 2017).
  • Medical students who use Anki score higher, in observational data only (Deng 2015; Lu 2021).

Known with caveat

  • Pictures on cards help concrete words and may hinder abstract ones (Ide 2026); they help only when overconfidence is checked (Carpenter 2012).
  • The direction of a card changes what is learned: production or comprehension (Griffin 1996; Webb 2020).
  • Set size matters less than spacing (Nakata 2016).
  • Embedded prompts are followed by long demonstrated retention, without a control (Matuschak 2019).

Not known

  • Whether one fact per card beats a richer card: no study under any name.
  • Whether a cloze card beats a question-and-answer card in a spaced program; Kim 2023 compared the two activities, not the two card types.
  • Whether the twenty rules, taken together, raise retention over a learner's untrained cards.
  • Whether a ready-made deck written to those rules beats a learner's own deck.
  • Whether an essay with embedded prompts teaches more than the same essay with a separate deck, or with none.
  • Whether generated cards keep or lose the generation benefit that learner-made cards show.
  • How many learners abandon a deck, and when.

Where practice runs ahead of the evidence

  • The twenty rules (Wozniak 1999) are the field's de facto standard, repeated by the Anki manual and by every guide read; they cite no study.
  • Matuschak 2020 says so plainly: its advice rests on experience because research gives no exact guidance.
  • Shared medical decks are used by most respondents (Lu 2021) with no trial against learner-made decks in that setting.
  • Generated decks are deployed in courses (Khurana 2025) before any randomised test of their effect.
  • Mnemonic-medium essays are published as a new form, with only their authors' informal data (Matuschak 2019).

Where evidence is ignored by practice

  • Learners drop the cards they think they know, almost universally, and apps let them suspend or delete cards; the dropping studies show a cost (Kornell 2008).
  • Shared decks dominate in medicine (Lu 2021) while the largest controlled comparison on text material favours learner-made cards (Pan 2023b).
  • Detail cards dominate what learners and textbooks produce (Lin 2018) while conceptual cards help weaker learners more.
  • Learners believe massing works better, 72% after one session, while spacing won for 90% (Kornell 2009, secondary).

What this review adds

  • It separates writing a card (formulation), making a deck (who writes it), and keeping it (dropping, leeches, abandonment).
  • It names a baseline and a measure for each (baseline).
  • It proposes shapes and checks for decks shipped with study material, each graded (shapes and checks).
  • It does not claim a new result or the state of the art.