From PDF Textbook to 800 Flashcards in 12 Minutes: First Aid Ch.4 Walkthrough

This is a war journal, not a tutorial. I sat down at 9:47 PM, opened SmartRecall, dragged in First Aid for the USMLE Step 1, 2025 edition, Chapter 4 — Cardiovascular, and started a stopwatch. Below is what happened over the next twelve minutes, with the timestamps and the cards that came out the other side.

I'm using the same chapter I used to hand-author cards for during my first failed Step 1 attempt. Back then, working from this exact chapter took me roughly two weeks of evening sessions. Tonight it took twelve minutes plus a culling pass. I'll show you the cards that survived and the ones that didn't.

The input

• File: First Aid Step 1 2025, Chapter 4 — Cardiovascular
• Pages: 52
• File size: 14.0 MB
• Format: PDF, mixed text + tables + diagrams (the diagrams matter — I'll come back to this)

A meaningful subset of Step 1 cardiovascular yield lives in this chapter: cardiac cycle and Wiggers diagram, heart sounds and murmurs, the JVP waveform, ECG and arrhythmias, congenital defects, ischemic heart disease, cardiomyopathies, the entire cardiovascular pharm matrix, vasculitides. If you can hold this chapter in your head you have probably bought yourself thirty Step 1 points.

The timeline

0:00   Drag-drop First Aid Ch.4.pdf into SmartRecall (52 pages, 14 MB)
0:08   Upload complete, file accepted
0:32   Text extraction done — 47,300 tokens, 23 figures detected
0:45   Section parser identifies 23 subsections
1:15   First batch of cards streams back in the UI (cloze + Q&A + case)
3:20   ~300 cards generated, scrolling fast
6:10   ~600 cards, model is into Pharmacology now
8:45   Generation complete: 812 cards across 4 types
9:00   I start the manual culling pass
12:00  Culling done. 640 cards survived. Deck saved.

Twelve minutes of clock time. About three of those were me clicking through cards and pressing x on bad ones. The rest was the model working while I refilled my coffee.

The 23 subsections it found

I want to surface this because it's the part that surprised me the most the first time. The section parser does not just split by H1 headings. It reads the chapter and produces a logical section list:

1.  Embryology — heart tube derivatives
2.  Embryology — fetal circulation, fetal-postnatal shunts
3.  Cardiac anatomy
4.  Cardiac cycle (Wiggers)
5.  Pressure–volume loops
6.  Heart sounds (S1–S4, splits)
7.  Murmurs — systolic
8.  Murmurs — diastolic and continuous
9.  Jugular venous pulse
10. Cardiac output and Fick principle
11. Frank–Starling
12. Resistance, capacitance, MAP
13. Electrophysiology — pacemaker action potentials
14. ECG basics
15. Arrhythmias
16. Congenital heart disease
17. Hypertension and dyslipidemias
18. Atherosclerosis
19. Ischemic heart disease and MI evolution
20. Cardiomyopathies
21. Heart failure
22. Vasculitides
23. Cardiovascular pharmacology

Section 23 alone produced 187 cards. Pharm is what kills people on Step 1, and pharm is what hand-authoring kills you on, because every drug needs at least three cards (mechanism, indication, side effect) and ideally five.

Sample cards — what came out

I'll show four. These are unedited from the generator. Two are good, one needed a fix, one I killed.

Card 1 — Cloze (kept as-is)

Source paragraph (First Aid p. 295):

"Frank–Starling relationship: ventricular force of contraction increases with end-diastolic volume up to a physiologic limit. The Frank–Starling curve is shifted upward by positive inotropes (e.g., catecholamines, digoxin) and downward by negative inotropes."

Generated card:

Frank–Starling: ventricular force of contraction increases with \{\{c1::end-diastolic volume\}\} up to a physiologic limit. The curve is shifted upward by \{\{c2::positive inotropes (e.g., catecholamines, digoxin)\}\}.

This is the cloze I would have written by hand. Two deletions, both testable, hint at preload without giving it away.

Card 2 — Q&A (kept after edit)

Generated card:

Q: What does the a wave on the JVP represent? A: Atrial contraction. Absent in atrial fibrillation; large ("cannon") a waves seen in third-degree AV block and tricuspid stenosis.

I edited the answer down. The original generation included a fifth sentence about ventricular tachycardia that was correct but turned the card into a paragraph. My culling rule: if the answer takes more than two breaths to read aloud, split it or trim it.

Card 3 — Case (kept)

Generated card:

Case: A 62-year-old man with chronic stable angina is intolerant of beta-blockers due to bradycardia. He is started on a new agent that reduces myocardial oxygen demand without affecting heart rate or blood pressure. Q: What is the mechanism of action? A: Ranolazine — inhibits the late sodium current (late I_Na), reducing intracellular calcium overload during ischemia and decreasing ventricular wall tension. Does not significantly affect HR or BP.

Excellent card. Tests the why this drug exists question, which is the only Ranolazine question Step 1 will ever ask. I would not have written this card by hand — I'd have written a generic mechanism card and missed the clinical hook.

Card 4 — Q&A (killed)

Generated card:

Q: What is the most common cause of death in the first 24 hours after MI? A: Cardiac arrhythmia, most commonly ventricular fibrillation, is the most common cause of death in the first 24 hours after MI.

Killed. The answer leaks the entire question. This is the most common failure mode in AI-generated cards and you have to be ruthless about it. The fix would be to rewrite the answer as just "Ventricular fibrillation," but it's faster to delete and let the next pass through this section regenerate.

Deck stats panel

After the culling pass:

┌─────────────────────────────────────────────────────────┐
│  Deck: First Aid Ch.4 — Cardiovascular                  │
├─────────────────────────────────────────────────────────┤
│  Total generated:           812                         │
│  Killed in culling:         172  (21%)                  │
│  Edited then kept:           58  (7%)                   │
│  Kept as-is:                582  (72%)                  │
│  Final deck size:           640                         │
├─────────────────────────────────────────────────────────┤
│  Card type breakdown                                    │
│    Cloze:                   311                         │
│    Basic Q&A:               201                         │
│    Multiple choice:          82                         │
│    Case analysis:            46                         │
├─────────────────────────────────────────────────────────┤
│  Subsection coverage         23 / 23                    │
│  Time elapsed:               12 min 04 sec              │
└─────────────────────────────────────────────────────────┘

What I killed and why

The 172 cards I killed broke down roughly like this:

• Answer leakage (~70 cards). The question contains the answer, like Card 4 above. Most common failure mode by a wide margin.
• Low specificity (~50 cards). "What is hypertension?" "Elevated blood pressure." Technically correct, useless for Step 1. These come from the model trying to be exhaustive about the chapter's introductory paragraphs.
• Duplicates (~35 cards). The model generated a cloze and a Q&A for the same fact. I keep one or the other, not both.
• Out-of-scope (~17 cards). Mostly cards generated from the figure captions that weren't actually testable content — labels on a diagram, footnotes about edition changes.

That's a 21% kill rate. It's higher than I'd like and it's lower than it was six months ago. The model is improving faster than I'm learning new culling heuristics.

The honest framing: assume one in four generated cards is mediocre. If you can't bring yourself to delete them, the workflow doesn't work for you. The whole point is that deletion is cheap when generation is cheap.

The comparison nobody publishes

The widely-cited estimate for high-quality flashcard authoring is about 60 seconds per card for a domain expert working from a source they know well. Wozniak's own numbers and the Anki community's 20 rules of formulating knowledge both put a careful student around that ballpark. Realistically, for a med student writing First Aid cloze cards on a chapter they're learning for the first time, it's worse — closer to 90 seconds.

So at the conservative 60s/card industry estimate:

800 cards × 60 seconds = 13 hours 20 minutes of pure card-making

Twelve minutes versus thirteen hours and twenty minutes. That's a 66× speedup, and the comparison is the conservative one — for a first-pass learner the real-world ratio is closer to 100×.

Even if I assume my 21% kill rate eats some of the gain, the post-culling math is still 12 minutes generation + 3 minutes culling (already counted) versus 800 × 60s = 13.3 hours. The order of magnitude doesn't move.

Week 1 review schedule (real SM-2)

Here's what the SM-2 scheduler did with the 640 cards over the next seven days. These are real numbers from my account this week.

Day 0 (Wed)   New cards introduced            812 → 640 after culling
Day 1 (Thu)   Due for review                  ~270
Day 2 (Fri)   Due for review                   ~95
Day 3 (Sat)   Due for review                  ~220   (Day-1 lapses returning)
Day 4 (Sun)   Due for review                  ~110
Day 5 (Mon)   Due for review                   ~80
Day 6 (Tue)   Due for review                   ~70
Day 7 (Wed)   Due for review                  ~180   (Day-3 graduations)

The two visible bumps (Day 3 and Day 7) are SM-2 doing its job: cards I rated Good on Day 0 graduate to a 3-day interval, then a 7-day interval. Cards I rated Hard or Again come back faster and stack onto the next short interval. By Day 14 the daily review load stabilizes around 60–90 cards, which is the steady state I can actually sustain alongside UWorld.

If you're new to spaced repetition: the pile looks punishing for the first three days, then the curve flattens. The first 72 hours are where most people quit. Don't.

Caveats I'm not going to hide

Diagrams don't fully work yet. The Wiggers diagram, JVP waveform tracings, and ECG strip examples in Chapter 4 are some of the most testable visuals in all of Step 1. The model reads the captions and the surrounding text, but image occlusion cards from the figures themselves are not in this product yet. I still draw a Wiggers diagram by hand once a week. There's no AI shortcut for that yet that I'd trust.

The 21% kill rate is real and you will feel it. Some chapters generate cleaner — biochemistry pathways, for instance, generate at maybe 12% kill. Pharmacology generates dirtier because the model loves to pad answers with mechanism trivia. If you're not willing to triage, this workflow is not faster than hand-authoring. It's just a bigger pile.

You still need the textbook open while you cull. Two of the cards I kept on first pass turned out to have subtle factual drift — one swapped "anterior" for "inferior" in an MI lead distribution. I caught both because I was reading along in the chapter. If you cull on vibes alone, errors slip through. Cull with the source open.

Try it on a chapter you know

If you have First Aid or Pathoma or BRS Physiology on your laptop right now, drop one chapter into SmartRecall and see what comes back. The free credits cover one full chapter generation. If the cards aren't the cards you would have written if you'd had thirteen hours, fine — you've spent four minutes finding out.

If they are, the math is the math.

I'd love to see what you get out of a chapter I haven't tested yet. Send me the deck stats panel — [email protected] reaches me directly, and I read every one. I'm especially interested in non-USMLE chapters: 法考 criminal law, CFA Level 1 ethics, GRE vocab lists. The generator is exam-agnostic but I haven't run timed walkthroughs on those yet.

— Alex