Why I built SmartRecall: three months of USMLE prep that broke me
Some products start with a vision. SmartRecall started with a notebook full of unanswered Anki cards and the slow realization that I was studying less every week, not more.
This is the founder story I wish someone had told me before I spent three months of USMLE Step 1 prep building card decks instead of learning medicine.
The Anki promise
If you've prepped for any heavy-memorization exam in the last decade — USMLE, MCAT, the Bar, CFA, 法考, 考研 — someone has told you to use Anki. The pitch is simple and accurate: spaced repetition is the most evidence-backed memorization technique in cognitive science. Pioneer studies by Cepeda et al. (2008) showed retention improvements of 200% over massed practice. Piotr Wozniak's SuperMemo SM-2 algorithm, published in 1987 and embedded in Anki since 2006, schedules each card at the precise moment your memory of it is about to decay.
The science works. I had no doubt about that.
What I doubted, by week six of my Step 1 prep, was whether I could keep up with the workflow Anki required to make the science work.
What actually went wrong
I was using First Aid for USMLE Step 1 — about 700 pages of dense, high-yield content. The "right" way to use Anki for Step 1 is well-documented: every concept becomes a cloze card, every drug becomes 3-4 cards (mechanism, side effects, indications, contraindications), every diagram gets occlusion cards.
A reasonable estimate for First Aid alone: 8,000 to 12,000 cards.
At my pace — and I'm not slow — that meant about 2-3 minutes of card authoring per testable concept. Multiply that out:
10,000 cards × 2.5 minutes each = 416 hours of pure card-making
That's 17 full days where you do nothing but type cards. Before you've reviewed a single one.
I was ten weeks out from my exam. I had Sketchy videos to watch. UWorld questions to grind. Pathoma to rewatch. And I was supposed to spend nearly one third of my remaining prep time just typing flashcards?
It got worse. Around card 1,800 I realized half the cards I'd made early were poorly worded. The cloze deletions were too long. The hints were too obvious. The card was testing trivia instead of the actual concept I'd flagged in the chapter. I started rewriting old cards — adding hours to the backlog while the exam date didn't move.
By week six I was making 50 cards a day and reviewing 200 — both numbers below where I needed to be — and the gap was widening every day.
I did not pass on the first attempt. I'll spare you that story.
The realization
What broke me wasn't Anki. The review experience is fine. The algorithm works. I still recommend SM-2 to anyone who asks.
What broke me was the unstated assumption baked into the entire spaced-repetition pedagogy: that the bottleneck is reviewing cards, so we should optimize the review experience.
For 99% of high-stakes exam preppers, the bottleneck is creating cards. The review side is solved. The authoring side is a wall of unpaid work between you and the science that's supposed to save you.
That's why most people who try Anki for an intense exam quit within a month. They don't quit because they hate the algorithm. They quit because the labor cost of feeding it is higher than the time it saves.
What I built instead
After Step 1 I went deep on the question: why hasn't anyone fixed the authoring problem?
The answer turns out to be timing. Until very recently, you couldn't ask a model to read a chapter of a textbook and emit good flashcards. The cards would be too literal, too long, miss the testable concepts, hallucinate facts that weren't in the source. By 2024-2025 the frontier models got reliable enough at structured extraction that you can hand them a PDF chapter and get back cards a real teacher would have written.
So that's what SmartRecall does. You upload a chapter — pharmacology, micro, biochem, whatever — and the model reads it, identifies what's testable, and emits four card types: cloze, basic Q&A, multiple choice with realistic distractors, and case analysis. The same SM-2 scheduler that powers Anki schedules your reviews. You spend your time reviewing and learning. Not typing.
A few things I deliberately kept out of the design:
No image occlusion (yet). It's a great technique, but the AI generation pipeline for it is still unreliable. Better to ship nothing than ship a half-baked version that loses people's trust.
No community deck marketplace. Shared decks have a reputation problem in medical education — the most popular community decks (Anking, Zanki) are great, but they're also so large that they reintroduce the original problem in reverse: now you're drowning in cards you didn't choose. SmartRecall makes your deck from your source, which means you actually know what's in it.
No Anki migration. I considered it for months. The cost-benefit doesn't work. People who already have a 5,000-card Anki deck for an exam don't need SmartRecall — they're already past the authoring wall. The people SmartRecall is built for are people like me at week one, who haven't started yet, and whose decision is "build cards by hand for 400 hours, or generate them in 4 hours and start reviewing today." That's the choice the product is designed around.
What I hope you take from this
If you're prepping for an exam right now and you're losing the authoring race — it's not your discipline. It's the workflow.
Try SmartRecall for free, or don't. The 20 free credits are enough to test it on one chapter and see if the cards it generates are the cards you would have written if you had time. If they are, the math gets very simple very fast. If they aren't, no harm done.
Either way: don't skip spaced repetition because the front-loading killed you the first time. The science is too good to give up on. Find the workflow that lets you actually use it.
I'm building this in public. You can find me as @alexrchen on GitHub and @alexrchen on Product Hunt — I write up what I learn about memory, exam prep, and where AI helps (or hurts) studying. If you want to ask questions or push back on any of this, support@smartrecallai.com reaches me directly.
— Alex
