Ask anyone who quit Anki why they quit. It’s almost never the reviewing — reviews take minutes. It’s the card-making: an hour of copy-pasting, formatting, and second-guessing for every lecture, before a single review happens. Card-making is the tax you pay to enter the system, and it’s why most people’s spaced-repetition career lasts three weeks.
The tax is now optional. Here’s the workflow, and — just as important — what still needs your judgment.
What a good card looks like (this part hasn’t changed)
AI or no AI, the rules for effective cards are old and well-tested — SuperMemo’s Piotr Woźniak wrote them down decades ago as the “20 rules of formulating knowledge.” The ones that matter most:
- One card, one fact. “Describe the Krebs cycle” is not a card, it’s an essay prompt. “What does the Krebs cycle produce per acetyl-CoA?” is a card.
- Cards must force retrieval. A card you can answer by recognition (“yes, I’ve seen this slide”) teaches you nothing. The question has to make you produce the answer.
- Short answers win. If the back of a card is a paragraph, you’ll grade yourself generously and the scheduler gets garbage data.
The reason hand-made decks fail is that following these rules for 300 facts is genuinely laborious. The reason AI-generated decks work is that a model can apply them uniformly at fact number 300, when a human has long since started pasting whole sentences onto card backs.
The workflow
In Repeto, you drop material onto a deck and it comes back as cards. What counts as material is deliberately broad:
- Pasted text — notes, a chat log, anything on the clipboard
- PDFs and documents —
.pdf,.docx,.txt,.md - URLs — any article or Wikipedia page; the text is fetched for you
- YouTube — the transcript becomes the source material
- Audio and video — recorded lectures are transcribed first, then carded
- Photos — a textbook page, a slide, a whiteboard after class
Two details make this usable for real course loads rather than demos. First, big sources run as a background job with visible progress — start a 90-minute lecture upload, close the tab, come back to a finished deck. Second, every generated batch lands tagged, so a bad run never pollutes a deck you’ve been building for months: filter by the batch tag, keep the good cards, delete the rest in one sweep.
The deck itself shapes the output. Each Repeto deck carries a generation profile — what a good card looks like in that subject — so a Spanish article becomes vocabulary cards with gender and an example sentence, while the same article dropped on a “Spanish culture” Q&A deck becomes questions about content. If the output needs steering, deck settings take plain instructions: “answers under 10 words”, “always ask for the mechanism, not the name”.
The 10 minutes that are still your job
Generated cards should be triaged, not trusted. Models occasionally mangle a number or flatten a nuance, and no generator knows what your professor emphasized. So after a batch lands, skim it once and do three things:
- Delete freely. A 60-card batch from a lecture should often become 40. Deleting a card costs nothing; reviewing a useless one costs forever. Aim your pruning at duplicates and trivia.
- Fix the front, not the back. If a card is weak, it’s usually the question that’s vague. Sharpen the cue.
- Add the 2–3 cards only you know to make. The thing the lecturer said twice, the distinction that confused you — that’s the highest-value card in the deck, and it takes 20 seconds now that it’s the exception rather than the job.
Ten minutes of triage on a one-hour lecture, versus an hour of manual card-writing that most people simply never do. That’s the actual trade.
If you have years of Anki work
Nothing here asks you to abandon it. Repeto imports .apkg files — decks, cards, media — and exports your whole collection back to .apkg whenever you want, so old decks and generated ones review side by side under the same FSRS scheduler, and nothing is locked in.
Generation from full documents, links, and lectures is part of Pro; the free tier includes 10 AI generations a week, which is enough to run this workflow on something real — one article or one chapter — before deciding it’s how you want to study. Drop the thing you’re actually supposed to read this week into app.repeto.space and see what comes back.