anki-flashcard-skill

Anki flashcard skill

A set of instructions that teaches an AI assistant to turn study material into Anki cloze deletion flashcards. The rules are informed by cognitive science research (e.g., desirable difficulties, elaborative encoding, retrieval, spacing, interleaving, generalization, interference, concrete examples) with the aim of building cards that help the learner efficiently encode new material, retain it longer and be able to transfer and generalize the learning better.

It works with any capable AI system and outputs a .txt file that imports into Anki with almost no configuration.

The whole skill is a single file: SKILL.md.

Quick start

1. Download SKILL.md.
2. Give it to your AI assistant, then add your study material:
   • Claude: add it as a skill so it activates automatically, or paste it into a Project's instructions.
   • ChatGPT, Gemini, Copilot, others: paste it as custom or project instructions, or attach it alongside your material in the chat.
3. Ask for flashcards, e.g. "Make Anki cards from this."
4. The assistant first shows 4–5 sample cards and asks about coverage (high-yield or exhaustive), card language, and deck name. Check that the samples make sense, then ask for the full set.
5. You receive a tab-separated .txt file. In Anki (desktop, version 2.1.55 or newer): Drag and drop the file in Anki, or use File → Import, select the file, confirm that Field 1 maps to Text and Field 2 to Back Extra, and import. The header lines in the file set the separator, note type, HTML handling, and tag column for you.

What to feed it

One lecture, one article, or stop at the end of a learning conversation you have had and ask it to make flashcards of the stuff you have learned in the conversation you just had. Large files or textbooks should not be used. The default is high-yield coverage, with a few highly relevant cards that assume that you have read and understood the material; ask for exhaustive coverage if you need more detail.

What the skill enforces

Every card must pass a quality checklist before the assistant outputs it. The core rules:

• One idea per card (minimum information principle), with cloze answers of 1–3 words. Attributions such as "Cohen and Squire (1980)" are kept as a single retrieval unit.
• Self-contained cards. Each card is understandable on its own, months later, out of order. No pronouns.
• Retrieval, not guessing. Bare clozes by default. Hints are only used where context cannot constrain the answer, because strong hints make retrieval feel fluent while measurably impairing later recall. Cards are checked for being too open (several answers fit) and too closed (answerable without having studied).
• Brief elaboration in the Extra field. A one-sentence "why" for causal facts, a mnemonic for arbitrary ones, criterial features when the card carries concrete examples, and a short source link last.
• Discriminative contrast cards for confusable concepts, and concrete-example cards (typical, typical, atypical) for central abstract concepts.
• Lists dissolved into cued single cards or scaffolded with one cloze index per item.

Key sources behind the rules include Roediger and Karpicke (2006) on test-enhanced learning, Vaughn and Kornell (2019) and Vaughn et al. (2022) on hint strength, and Micallef and Newton (2022) on example-based cards. Full references are given inside the skill file where each rule is stated.

Before you paste

• Do not paste copyrighted textbook text, or other people's personal data, into external AI services. Use your own notes, open materials, or material you have the right to share.
• Read the cards. The assistant can misread the source or get a source link wrong. Any link it cannot verify should be flagged with [URL NEEDED].
• Generating cards is not the same as learning. The learning happens when you repeatedly retrieve the answers in review.

Versions

Versions are dated (for example 2026-08-20) and stated in the file's frontmatter. Each release is tagged on GitHub and archived on Zenodo with its own DOI. The concept DOI above always resolves to the latest version. See CHANGELOG.md for what changed.

How to cite

If you use or redistribute the skill, please cite it. Using APA 7:

Ingebrigtsen, M. (2026). Anki flashcard skill (Version 2026-08-20) [AI skill file]. Zenodo. https://doi.org/10.5281/zenodo.22033018

BibTeX:

@misc{ingebrigtsen2026anki,
  author    = {Ingebrigtsen, Magnus},
  title     = {Anki flashcard skill},
  year      = {2026},
  version   = {2026-08-20},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22033018},
  url       = {https://doi.org/10.5281/zenodo.22033018},
  note      = {AI skill file}
}

GitHub's "Cite this repository" button (from CITATION.cff) gives the same reference in other formats.

Contributing

Issues and pull requests are welcome, particularly reports of cards that slipped through the checklist, and evidence that would change a rule. Adaptations are welcome under the license terms; please keep the attribution section in SKILL.md intact.

License

This work is licensed under a Creative Commons Attribution 4.0 International License. You may share and adapt it for any purpose, including commercially, as long as you give appropriate credit.

Author

Magnus Ingebrigtsen, PhD candidate and university lecturer in psychology, UiT – The Arctic University of Norway.

For studenter (norsk)

Denne fila er en instruks til en KI-assistent om hvordan den skal lage Anki-kort. Du trenger ikke lese hele fila selv.

1. Last ned SKILL.md.
2. Åpne en chatbot og lim inn eller last opp fila.
3. Last opp notater, artikler eller lignende, og skriv: «Lag Anki-kort av dette.»
4. Chatbotten vil da først vise deg noen eksempelkort og spørre om språk, navn på kortstokken og hvilket omfang du vil ha (eks. alle detaljer eller viktigste konsepter). Se over at kortene gir mening før du ber den generere resten.
5. Du får en .txt-fil. I nyere versjoner av Anki kan du enten bare dra-og-slippe fila i Anki, eller velge Fil → Importer, velg fila, sjekk at felt 1 går til Text og felt 2 til Back Extra, og importer.

Noen råd: Ikke lim inn opphavsrettsbeskyttet materiale eller andres personopplysninger i eksterne KI-tjenester. Les gjennom kortene – som du sikkert vet kan slike modeller ta feil. Ikke kast inn enorme mengder med tekst. Prøv deg fram med mindre artikler og tekster.