How to Make Good Anki Cards with AI (and When to Write Them Yourself)
Published
You paste a chapter into ChatGPT, ask for Anki cards, and a minute later you have 80 of them. It feels like a week of work, finished. Then the reviews start. What did the second group do better on? Which group? What causes forgetting? How many causes do you want? You press Again, then Delete, and start to wonder whether writing the cards yourself would have been faster.

That frustration now has a number attached. In a 2026 benchmark called Memory Machines, Ozzie Kirkby and Andy Matuschak gave AI models a source text and a reader's highlight, then scored the flashcards that came back. GPT-5.2, the strongest model in the test, still wrote unusable cards 36% of the time, and no model added since has done better. An Anki AI workflow pays off when the model writes drafts and you act as the editor. This guide gives you that workflow: copy-paste prompts for ChatGPT, Claude and Gemini, the route from NotebookLM to Anki, a six-question test for judging cards, an import file that works, and the cases where you should write the cards yourself.
Can AI Make Good Anki Cards?
Yes, as first drafts. A model can give you a usable share of cards from material you have already read, but a large share will need fixing or deleting, and the direct evidence on learning is thin.
Start with quality. The share of unusable cards in Memory Machines was 36% for GPT-5.2, 46% for Gemini 3 Pro, 50% for Claude Opus 4.5 and 71% for GPT-4o. The authors found that models were "rarely off-target." They usually spotted the right idea and then built the card badly, with wording that looks fine today and falls apart over months of review. Take the exact numbers loosely. The scores came from an AI judge that compared each new card with cards people had already rated, drawn from a set of about 1,500. That judge agreed with human raters only moderately well.
Newer models have not fixed this. The newest models in the report date from late 2025, and the authors have kept adding releases to the ranking table on the project's site. As of November 2026, GPT-5.2 still leads it. Claude Sonnet 5 wrote unusable cards 42% of the time, GPT-5.5 45%, Claude Opus 4.8 49% and Claude Fable 5 53%. The models you are most likely to use today, such as the GPT-6 family, Claude Opus 5.5 and Gemini 3.8 Flash, are not in the table yet, so nobody has measured them on this task. The pattern so far is that a newer model is no guarantee of better cards.
Accuracy is a separate problem. In a single-institution medical school study by Jay Khurana and colleagues, the team tuned its prompts and had people grade the output for invented facts. The best setup still produced one hallucination per 21 flashcards (4.8%). In a 100-card deck, that is about five cards teaching you something false.
The evidence on grades is the weakest part. In that same study, exam scores showed no significant difference between students who used the AI flashcards and students who did not. Students opted in, the users started from lower scores, and one flashcard group had only 14 people, so the result can't tell you much in either direction. As of November 2026, we could not find a randomized trial that compares AI-generated Anki cards with handwritten ones on long-term retention. Be suspicious of any page that quotes a precise retention gain for AI flashcards.
Use AI to save typing, and keep the judgment for yourself.
What Makes a Good Anki Card
A good Anki card asks for one thing, has one correct answer, and still makes sense when you meet it months later with no memory of where it came from.
Piotr Wozniak, the creator of SuperMemo, set down the classic version in his twenty rules of formulating knowledge. The best known is the Minimum Information Principle: keep each item as simple as you can. A small question gets retrieved the same way at every review. A card that asks for five things leaves you grading yourself on a blur. Andy Matuschak's guide to writing prompts adds a useful test: a good prompt is focused, precise and consistent, you can answer it nearly every time, and it still takes real effort to recall.
Those ideas boil down to something you can run on every draft. Call it the Six-Question Card Check, and delete or rewrite anything that fails.
Do you care about this? If you highlighted it for a reason, keep going. Trivia about the author or the journal rarely earns a place.
Do you understand the answer? Wozniak's first rule is "Do not learn if you do not understand." If you can't say why the answer is true, go back to the source first.
Does it test one idea? If the answer needs an "and," you probably have two cards.
Does it make sense alone? Read it as if you had never seen the chapter. "This study" and "the second group" fail.
Is there only one right answer? If you could give a different correct answer next month, the card will teach you nothing reliable.
Is it true? Open the source and check. Wozniak's rule 18, "Provide sources," makes this a ten-second job, which is why the prompt below puts a page number on every card.
Here is what failures look like. The drafts below were written for this guide to show common patterns (they are not output copied from a model), and the facts come from our guide to active recall.
| Problem | Typical draft | Fix |
|---|---|---|
| Too much at once | What did Roediger and Karpicke find in 2006? | Split it into two cards |
| Lacks context | What did the second group do better on? | Name the study and the group |
| Many valid answers | What improves memory? | Ask about one named finding |
| Unordered list | Name Dunlosky's ten techniques. | Delete, or ask for the top two |
| Gives itself away | Retrieval practice means {{c1::retrieving}}. | Hide a term the sentence doesn't echo |
| Trivia | Which journal published the 2006 study? | Delete unless you must cite it |
Take the first row. The draft asks for a whole paper, so any answer is half right. The fixed card reads: In Roediger and Karpicke's 2006 experiment, which group scored higher one week later: restudy or recall? A second card covers the five-minute test, where the result flips. The vague What improves memory? becomes Which technique did Dunlosky's 2013 review rate as high utility alongside practice testing?
How to Make Anki Cards with AI in Five Steps
- Read, then pickFinish the chapter, then mark about ten passages.You
- Draft with AIThe model writes 1 to 3 cards per highlight.AI
- Cull the draftsKeep, fix or delete every card before importing.You
- Import into AnkiSave a .txt file and check the preview.You
- Fix as you reviewEdit or delete any card that trips you up.You
1. Read First, Then Pick What Deserves a Card
Finish the chapter before you open a chatbot. Wozniak's second rule is "Learn before you memorize," because scattered facts have nothing to attach to until you have the overall picture. A reading method such as SQ3R helps if you tend to skim.
Then highlight the passages you want to keep, and start with about ten. This is also the setup the Memory Machines team tested: the model got the full source plus the reader's highlight. The highlight tells the model what matters to you, which it has no other way of knowing.
2. Draft the Cards with This Anki ChatGPT Prompt
Paste this into ChatGPT, Claude or Gemini. Attach your source file or paste the text, then fill in the three brackets at the top and the deck name further down.
Help me make Anki flashcards from a source I have already read. SOURCE: [attached, or paste it here] MY HIGHLIGHTS: [paste the passages I want to remember] WHAT THIS IS FOR: [for example, first-year pharmacology exam] Write cards only for my highlights. Write 1 to 3 cards per highlight, and fewer if that is all a highlight deserves. Rules for every card: 1. Test one idea. If the answer needs "and", split the card. 2. The question must make sense alone. Never write "this study", "the author" or "the passage". Name the thing. 3. Ask a question with one correct answer that will still be the same answer in six months. 4. Keep answers as short as you can, ideally under 15 words. 5. When a highlight explains a reason, ask why or how instead of asking for a definition. 6. No yes/no questions. No lists of more than 3 items. Turn a longer list into several cards. 7. Use only facts stated in the source. If a highlight is unclear, skip it. 8. Use a Cloze card only for a fixed term, name or number that belongs inside its sentence. Everything else is Basic. Output one code block in exactly this format, with no blank lines: #separator:Pipe #html:true #notetype column:1 #deck:[DECK NAME] #tags:ai-draft Basic|Question goes here?|Short answer<br><small>Source: p. 12</small> Cloze|A sentence with {{c1::the hidden term}} inside it.|Source: p. 12 Format rules: one card per line. Never use the | character inside a card. Never start a line with #. Never start a field with a double quote. Write the characters < > and & in card text as < > and &. After the code block, list any highlights you skipped and why.
The rules map to the failures in the table above. The source restriction in rule 7 lowers the odds of invented facts, but it won't remove them, as the 4.8% figure shows.
3. Cull the Drafts Before You Import Them
Go through the output line by line and run the Six-Question Card Check. Every card gets one of three verdicts: Keep, Fix or Delete. Do this before importing, while the cards are still lines of text and deleting one costs nothing. Once a bad card is in Anki, it charges you a few seconds every time it comes up, for months.
Set a limit so the job gets done. Try one 25-minute block for ten highlights, with a Pomodoro timer or the Yellow Tomato Chrome extension, which blocks the tabs you would otherwise wander off to.
Expect to lose a lot. At the benchmark's rates, more than one card in three from even the best model needed real work or the trash. If you deleted nothing, read them again.
For a card with a good idea and bad construction, ask for options and pick one yourself:
This Anki card has a problem: [say what is wrong in one sentence]. Card: [question] / [answer] Source passage: [paste it] Write three different rewrites. Each must test one idea and have one correct answer. Use only facts from the passage. Keep the format Basic|Question|Answer.
Don't hand the verdict to the model. When Kirkby and Matuschak used Claude Sonnet 4.5 as a judge, only 56% of the cards it approved were really usable. That rose to 78% once it could compare against human-rated cards for the same highlight, which you won't have. A second AI opinion can catch clumsy wording. The final call stays with you.
4. Import the File into Anki
Copy only what is inside the code block, without the backtick lines around it if your chatbot shows them. Paste it into a plain text file and save it as UTF-8 with a .txt ending (in TextEdit, choose Format, then Make Plain Text first). In Anki, click File, then Import, and choose the file. Look over the preview, then import.
Your file should look like this:
#separator:Pipe
#html:true
#notetype column:1
#deck:Learning science
#tags:ai-draft
Basic|In Roediger and Karpicke's 2006 experiment, which group scored higher on a test one week later: restudy or recall?|Recall<br><small>Source: Psychological Science, 2006</small>
Basic|Why can rereading feel more effective than self-testing right after studying?|Familiar words feel known even when you cannot produce them<br><small>Source: active recall guide</small>
Cloze|The measurable benefit of retrieval practice on later tests is called the {{c1::testing effect}}.|Source: active recall guide
Cloze|Dunlosky's 2013 review rated {{c1::practice testing}} and {{c2::spacing}} as high-utility techniques.|Source: Psychological Science in the Public Interest, 2013The lines starting with # are file headers, which Anki's manual lists in full. They set the separator, allow HTML, say that the first column holds the note type, and pick the deck and tag. The ai-draft tag lets you find this batch later. Each cloze number becomes its own card, so {{c1::...}} and {{c2::...}} in one sentence make two cards, as the manual on cloze deletions explains.
This exact file was run through the import code in Anki 26.09.4 (the open-source anki package). It produced 4 notes and 5 cards, all in the right deck with the right tag, and the same test showed what breaks. None of these problems stopped the import, so read the preview before you confirm:
| What went wrong | What Anki did | How to avoid it |
|---|---|---|
| Tabs turned into spaces | Whole card went into Front | Use the pipe separator |
A ``` line copied at the top | Headers ignored, junk notes | Copy only the block's contents |
A | inside a card | Card split, extra text dropped | Reword the card |
A card line starting with # | Line skipped | Start with the note type |
| A field starting with a quote | Quotes removed | Move the quote later |
| A misspelled note type | Row skipped | Copy the name from Anki |
For vocabulary, write Basic (and reversed card) in the first column and Anki makes a card in each direction. For formulas, Anki uses \( and \) around inline math, and the manual warns that a }} inside math can end a cloze early, so add a space between those braces.
5. Fix Cards the First Time They Annoy You
The import is one more draft. In Anki's Browse window, search for tag:ai-draft to see the batch. By default, Anki introduces 20 new cards a day, so a batch of 40 arrives over two days.
Review as usual, and when a card makes you hesitate for the wrong reason, press Edit and fix it right there. Matuschak recommends the same habit: revise any card that keeps failing or that you no longer care about. Deleting is allowed too. If you haven't already, enable FSRS in Anki's deck options. It is a scheduler that models how likely you are to forget each card, and it aims for 90% recall by default.
Our spaced repetition guide covers why these gaps work. If you study from paper and have a fixed exam date, the spaced repetition calculator gives you review dates without any app.
How to Use NotebookLM to Make Anki Cards
Google renamed NotebookLM to Gemini Notebook in July 2026, so that is the name you will see on screen. The flashcard feature carried over. There is no button that sends cards to Anki, but Google's help page confirms you can download a deck as a CSV file, and Anki can read that.
- Open a notebook and add your sources. You need edit access to make flashcards.
- In the Studio panel, click the pencil icon on Flashcards. Choose a difficulty and paste the prompt below into the prompt box.
- Go through the deck in the viewer and delete cards that fail the Six-Question Card Check. Google's help page lists deleting a card but not editing one, so wording fixes wait until the cards are in Anki.
- Click Download to save the CSV. Google notes this may be unavailable if a Play Books ebook is one of your sources.
- In Anki, click File, then Import, and choose the CSV. Check that the first column maps to Front and the second to Back, pick a deck, and add the tag
ai-draft.
Make flashcards only about these topics: [paste your list or highlights]. Test one idea per card. Every question must make sense without the notebook and have one correct answer. When the sources explain a reason, ask why or how. Do not ask about the authors, the documents or their structure. Keep answers under 15 words.Look at Anki's import preview first: if the file has a header row, it will show up there as a card. Then import five cards as a trial before you bring in the whole deck. Browser extensions such as NotebookLM Tools and AnkiNLM add their own export buttons, but they are unofficial, and you may not need them. Gemini Notebook's own viewer tracks which cards you got right or missed, but Google's help page says nothing about spaced repetition. For scheduling, you still need Anki.
Choose the chatbot prompt when you want cloze cards and a page number on every card. Choose NotebookLM when your sources are long and you would rather stay inside them.
Do You Need an Anki Add-on for AI Cards?
Probably not at first. The prompt and text file cost nothing and show you every card before it reaches your deck. If you make cards every day and the copying gets old, two tools at least keep you in charge of what gets added.
Anki-GPT is an add-on that turns pasted notes or a slide screenshot into one to five cards you can edit before adding. It includes 75 free AI answers, then costs from $3 a month as of November 2026. anki-llm is a command-line tool that shows you candidate cards and imports only the ones you select. You bring your own API key, and commands that touch your collection need the AnkiConnect add-on.
Avoid anything that promises to push a whole PDF into your collection in one click. That skips the only step that makes this method work. If you are comparing general study tools instead, our overview of AI tools for students covers the wider field.
When You Should Write Anki Cards Yourself
Write a card yourself when making it is the part that teaches you. Generating information from your own head tends to beat reading it: a meta-analysis by Bertsch and colleagues pooled 86 studies and found an average benefit of 0.40 standard deviations.
Flashcards show the same thing. Steven Pan and colleagues ran six experiments with self-made and premade digital cards, holding total study time equal, and found self-made cards ahead on a test 48 hours later by an estimated d = 0.45. How the cards got made mattered. Students who typed passages out word for word did no better than students with premade cards, and the typing ate nearly half their session. Paraphrasing helped, and so did plain copy-and-paste, a result the authors say they can't fully explain. Their best guess is that the gain comes from working through the material while you build the card, which is the step an AI takes off your hands. None of those cards came from an AI, so applying this to AI drafts is an inference.
The closest AI warning comes from a 2025 randomized trial by André Barcaui with 120 business undergraduates at a Brazilian university. On a surprise test 45 days later, students who had studied with unrestricted ChatGPT scored 57.5%, against 68.5% for students who had studied without AI. Only 85 of the 120 came back for that test. The ChatGPT group also spent less time studying (3.2 hours against 5.8), though a smaller gap in scores remained after the analysis adjusted for that. The trial covered chatbot use in general at one university and did not involve flashcards. It still points the same way: effort you skip at study time tends to come back later as forgetting. Our guide to using AI to study smarter has more ways to keep that effort in place.
A practical split:
| What you are learning | Who writes the first draft |
|---|---|
| Vocabulary, drug names, anatomy, dates | AI, then cull |
| Lectures you attended, exam close | AI, from your own notes |
| A concept you can't explain yet | You, after the Feynman technique |
| Ideas for your own work or writing | You, with your own examples |
| Procedures and problem-solving | Neither: do practice problems |
The first two rows only work for material you already understand in context. For lecture content, draft from your own notes and check every card against them. For procedures, spend the time on practice problems and keep cards for the facts underneath.
There is a middle route that keeps the writing with you. Draft your own cards, then ask a model to look for weak spots without rewriting anything:
Here are Anki cards I wrote myself. For each card, tell me only what could go wrong in a review six months from now: ambiguous wording, more than one valid answer, or more than one idea. Do not rewrite the cards. Do not add facts.
[paste your cards]Flashcards are one form of active recall, and they suit facts better than reasoning. Keep them inside a wider plan for studying effectively, with practice problems and explaining ideas out loud doing the rest.
Common Mistakes When Using AI for Anki
Asking for as many cards as possible. Every card you import adds reviews for months. Ask per highlight and cap the number.
Generating from a chapter you haven't read. You can't judge a card on a topic you don't understand, so the errors go straight into your memory.
Importing first and culling later. Later rarely comes. Do the Keep, Fix, Delete pass while the cards are still text.
Assuming your own PDF makes the cards safe. The 4.8% hallucination rate above came from cards built on specific lectures. Check anything you will be tested on.
Letting the model grade its own homework. AI judges in the benchmark approved plenty of bad cards. Use a critique as a hint and decide for yourself.
Start with Ten Cards Tonight
Pick one chapter you have already read and highlight ten passages. Run the drafting prompt, cull the output, and import what survives with the ai-draft tag. Two weeks from now, search for that tag and count how many cards you later edited or deleted.
That count is better evidence about your own subject and your own chatbot than any benchmark. If most cards held up, keep drafting with AI. If they didn't, write more of them yourself. And if you deleted none of the ten tonight, read them again.
Frequently Asked Questions
Can AI make Anki cards?
Yes, but treat them as drafts. In the 2026 Memory Machines benchmark, the best model tested still wrote unusable cards 36% of the time, and newer models added to its ranking since have not beaten that. Plan to review every card before it enters your deck.
What is the best Anki ChatGPT prompt?
The best Anki AI prompt names the passages you want to remember, limits each card to one idea, requires one correct answer, restricts the model to your source, and specifies the import format. The full prompt in step 2 does all of that, and you can paste it into Claude or Gemini as well.
How do I import AI flashcards into Anki?
Save the cards as a UTF-8 plain text file with Anki's header lines at the top, then click File, then Import. A pipe separator is safer than tabs for text copied from a chatbot, because tabs can turn into spaces and put the whole card on the front.
Can NotebookLM export flashcards to Anki?
Not directly, but it can download flashcards as a CSV file that you can import into Anki. NotebookLM is now called Gemini Notebook, and the Download option sits in the flashcard viewer. In Anki, click File, then Import, choose the CSV, and map its columns to Front and Back.
Are AI-generated Anki cards accurate?
Mostly, but not reliably enough to skip checking. One medical school study found one hallucination per 21 flashcards even after its team tuned the prompts, so compare each card with your source, especially before an exam.
Is it better to make your own Anki cards?
Usually, when you have the time and the material is new to you. Across six experiments by Pan and colleagues, self-made digital cards beat premade ones overall, though no trial we could find has compared them with AI-made cards directly. AI drafting makes the most sense for facts you already understand.
You may also like
Active Recall Technique
25 Aug 2026Spaced repetition Technique
21 Mar 2023Feynman Technique
29 May 2023How to Study Effectively
15 May 2024