How to make flashcards with AI: the rules that separate useful decks from forgettable ones

What 'good' AI flashcards actually look like

A good AI-generated flashcard is atomic, recallable in under eight seconds, and tagged. Atomic means one fact, no compound questions. Recallable in eight seconds means the cloze hides a single high-information word, not a clause. Tagged means future you can filter your deck by topic without opening every card. Miss any of these and you have a flashcard you will leaf-press through on autopilot, which teaches you nothing and inflates your review count.

The most common failure mode is the compound card. 'The femoral nerve supplies quadriceps, sartorius, and pectineus' as a single cloze teaches you the order of the muscles, not which nerve supplies them. Three separate atomic cards beat it every time. The second most common failure is clozing the easy word. 'The {{c1::brain}} is in the skull' is not a flashcard, it is filler. Cloze the word a clinician would need to recall under pressure.

The four rules to bake into every prompt

1. One fact per card

Tell the model explicitly: 'no compound facts'. If a sentence in your notes contains three drugs, three side effects or three branches, instruct the model to split it into three cards. Anki's documentation calls this the Minimum Information Principle and it is the single rule most students skip.

2. Cloze by default, Q&A only for definitions

Cloze deletions are faster to write, faster to review, and let you hide multiple fields in one sentence (c1, c2, c3) without writing new cards. Reserve Q&A format for definitions or for cards where the answer is a list. Cloze handles everything else more efficiently.

3. Cloze the high-information word

The information value of a cloze equals how hard the word is to guess from context. 'The treatment for {{c1::anaphylaxis}} is 0.5 mg IM adrenaline' tests nothing because the dose gives it away. Flip it: 'The treatment for anaphylaxis is {{c1::0.5 mg}} IM {{c2::adrenaline}}.' That is two atomic clozes on one card, both testing what matters at 3am on a ward.

4. Tag at generation time, not after

Append tags inside the prompt: 'tag::cardiology::pharmacology::beta-blockers'. Anki uses :: as a tag hierarchy separator, so tagged cards become filterable by module, system and topic later. Nobody retrofits tags onto 600 cards. Generate them or lose them.

The prompt that bakes all four rules in

Paste your notes after the prompt below. The output imports cleanly into Anki via File then Import then Text Separated by Semicolon. Tweak the module and subtopic before each batch.

"Convert the following notes into Anki cloze deletions. Rules: (1) one fact per card, no compound facts. (2) Cloze the high-information word, the one a clinician would need to recall under pressure, not trivial words. (3) Use multiple clozes (c1, c2, c3) on a single sentence where it tests separate facts. (4) Append ' tag::medicine::[module]::[subtopic]' to each card. (5) Output one card per line, plain text, no numbering, no quotes, no commentary. Module: [paste]. Subtopic: [paste]. Notes: [paste]."

Which AI to use to generate the cards

ChatGPT, Claude and Gemini all follow the prompt above competently on the free tier. Claude (Sonnet 4.5 and above) tends to write the cleanest medical cards because it follows formatting instructions most strictly. ChatGPT is the most accessible, and the GPT-5 family handles longer note dumps without truncating output. Gemini is fine but more likely to add unrequested headers and emoji. Pick whichever you already have open.

All-in-one apps: convenient, but the algorithm matters

Knowt, Quizlet's AI features, RemNote and a growing list of newer apps generate and review in one place. Convenience is real. The trade-off is the spaced-repetition algorithm. Anki uses FSRS (Free Spaced Repetition Scheduler), which beat SM-2 by roughly 20 percent on workload-equivalent retention in the 2023 benchmark by Jarrett Ye, and SM-2 already beat most alternatives. Most all-in-ones use a simplified Leitner-style schedule, which is fine for languages and poor for high-volume medical content.

What to feed the model, and what to keep out

• Your own lecture notes work best. They are already filtered for what your course examines.
• A NotebookLM study guide built from your slide deck. Source-grounded, so the model cannot drift.
• A specific textbook section, pasted in. Not the whole chapter.
• Guideline summaries from NICE, the BNF, or local trust protocols.
• Past-paper answers. Clozes built from worked rationales are unusually high yield.

• Avoid raw lecture transcripts. Filler sentences become filler cards. Summarise first, cloze second.
• Avoid 200-card sessions. Generate 30, audit the format, then scale.
• Avoid auto-import without reading. Bin bad cards at generation, not three weeks into review.
• Avoid patient identifiable data. Even anonymised case notes can breach GMC guidance if pasted into a consumer model. Strip names, MRNs and dates.

A worked example: generating beta-blocker cards

Take this note: 'Beta-blockers reduce mortality post-MI by 23 percent (ISIS-1, 1986). Bisoprolol and carvedilol are first line in heart failure with reduced ejection fraction. Avoid in severe asthma and decompensated heart failure.' Run it through the prompt above. The output is six atomic cards, not one paragraph card.

1. Paste notes into ChatGPT with the prompt above (module: cardiology, subtopic: beta-blockers).
2. Skim the output. Bin any compound clozes; reject any that hide trivial words.
3. Copy the cleaned list into a .txt file, one card per line.
4. In Anki: File then Import then choose Cloze note type then Import.
5. Set new cards per day to 20 in deck options. Review daily, ideally at the same time.
6. After two weeks, suspend cards you have failed three times running and rewrite them by hand.

Deck hygiene: the habit that keeps the system alive

Most students abandon Anki not because the cards are bad but because the queue becomes unmanageable. The single best preventative habit is a weekly 15-minute audit: open the browser, sort by 'lapses' descending, and rewrite or suspend the top 20. Cards that lapse repeatedly are almost always compound, ambiguous, or testing the wrong word. Rewriting them is faster than reviewing them forever.

The second habit is honesty on the Again / Hard / Good / Easy buttons. FSRS calibrates to your inputs. Press Good on cards you guessed and the algorithm will assume you knew them; the result is a schedule that feels easy until exam week, when the cards you faked come due in bulk. Press Again whenever recall took more than a few seconds. The short-term cost is more reviews this week. The long-term saving is enormous.