Medical students spend 2-3 hours after every lecture manually making Anki flashcards. They sit with slides on one monitor and Anki on the other, grinding through dozens of concepts, deciding what's worth testing, writing cloze deletions, tagging everything. Then they do it again tomorrow.
My girlfriend is in her first year of medical school. I watched her do this ritual night after night and thought: this is a solved problem, someone must have automated it. Turns out — not really.
AnKing exists, and it's great, but it covers board prep material. Nobody's making pre-made decks for your school's specific cardiology lecture from Dr. Martinez on a Tuesday. That content is unique to your program, your professor, your exam. You're on your own.
So I built SlideToAnki. Upload a PDF or PowerPoint, AI generates cloze deletions and Q&A cards, you review and edit them, then export a .apkg file ready for Anki. Free, no sign-up required.
But the tool itself isn't the interesting part. The interesting part is what happens next.
The quality problem
Here's the thing about AI-generated flashcards: most of them are bad.
Not obviously bad — they look like flashcards. They have a front and a back. They test something. But anyone who's studied with Anki seriously knows the difference between a card that builds understanding and a card that just tests surface-level recall.
A bad cloze: "The {{c1::heart}} pumps blood to the body."
A good cloze: "{{c1::Digoxin}} toxicity is exacerbated by {{c2::hypokalemia}} because both compete for the same binding site on Na+/K+ ATPase."
The first one is trivia. The second one tests a mechanism, links two concepts, and is the kind of thing that shows up on Step 1. Getting the AI to consistently produce the second kind is hard. Prompt engineering gets you 70% of the way there. The last 30% requires something else entirely.
Enter the flywheel
Every card SlideToAnki generates has a thumbs up and thumbs down button. That's it — simple binary signal. But that simple signal is the foundation of everything.
Here's how the flywheel works:
Step 1: Generate. User uploads slides, AI produces cards. Some are great. Some are garbage. That's fine.
Step 2: Feedback. User reviews cards, keeps the good ones, downvotes or deletes the bad ones. This takes 2 minutes instead of 2 hours — editing is always faster than creating from scratch.
Step 3: Learn. Those thumbs up/down signals become training data. We can analyze patterns: what makes a card get upvoted? What makes it get deleted? Which cloze deletions feel natural vs. arbitrary? Which question styles actually test understanding?
Step 4: Improve. Feed those patterns back into the prompt. Adjust the generation parameters. The next batch of cards is slightly better. Repeat.
This is a data flywheel. The more people use SlideToAnki, the better the cards get. The better the cards get, the more people use it. Each cycle compounds.
Why this matters more than the AI model
Everyone building AI products right now is obsessed with which model they're using. GPT-4o vs. Claude vs. Gemini — it's the new "what framework do you use?" debate.
But here's what I've learned from building two AI products now: the model is a commodity. OpenAI ships a better model every few months. Your prompt engineering advantage has a half-life of maybe six months before the base model just does it by default.
What's not a commodity is domain-specific feedback data. Nobody else has thousands of med student judgments on what makes a good pharmacology cloze deletion. That data is unique to SlideToAnki, and it gets more valuable with every card that gets rated.
This is the same pattern you see in every successful AI product. Tesla doesn't win because they have the best neural net architecture — they win because they have billions of miles of driving data feeding back into the system. Spotify's Discover Weekly isn't magic — it's years of skip/save signals from millions of users.
The AI is the engine. The data flywheel is the fuel. Without the fuel, you're just running the same generic model as everyone else.
What's actually planned
I'm not going to pretend I have a 5-year roadmap on a whiteboard somewhere. But here's what's coming based on real feedback from the first wave of users:
Multi-language support — Already shipped this one. A user uploaded German lecture slides and got English cards back. Embarrassing. Fixed it in a day. SlideToAnki now generates cards in whatever language your slides are in.
Image occlusion — This is the most requested feature. Medical education is extremely visual — histology slides, anatomical diagrams, pathway charts. Right now we handle text well but can't do "label this diagram" style cards. That's next.
Smarter slide parsing — PDF extraction is messier than you'd think. Tables get mangled. Diagrams become gibberish text. We're improving the parser to understand slide structure better, not just extract raw text.
Cohort sharing — If 150 students in your class are all uploading the same lecture, we should be able to pool that signal. One student's edits improve the deck for everyone. This turns a solo tool into a collaborative one.
Professor-specific optimization — Over time, we'll have enough data to know that Dr. Smith's pathology lectures emphasize mechanisms while Dr. Jones focuses on clinical presentations. The AI can adapt its card style to match the teaching style.
AnKing integration — The holy grail. Cross-reference generated cards against AnKing's existing decks so you're not duplicating what's already covered for board prep. Your SlideToAnki cards fill in the gaps that AnKing doesn't cover.
The honest state of things
SlideToAnki is early. Really early. The card quality is good but not great. The parsing works on most PDFs but chokes on some. The export is solid but basic.
But the flywheel is turning. Every upload, every thumbs up, every deleted card makes the next generation slightly better. That's the bet — not that the V1 is perfect, but that the system improves faster than any static tool can.
If you're a med student (or dental, nursing, PA, pharmacy — anyone drowning in lecture slides), give it a try. It's free. And if the cards are bad, hit that thumbs down button. You're not just complaining — you're making it better for the next person.
That's the whole point.