The real time cost of making Anki cards from one lecture deck.

For a 90-slide lecture with about 200 testable facts and 14 labeled figures, a conservative honest number is 24 to 30 hours. That breaks down as 6 to 8 minutes per multiple-choice card (read the slide, distill the fact, write the stem, write three plausible distractors, sanity-check) times roughly 200 cards, plus 10 to 15 minutes per image-occlusion card (export the figure, place masks, label the regions) times 14 diagrams. Strip out the image-occlusion work and you're at 20 to 26 hours for the text cards alone. Skip the distractor work (cloze only) and the per-card minute number drops to about 3 to 4, which is roughly 10 to 13 hours for 200 cloze cards.

Around 10 to 12 MCQ cards per hour, sustained over a multi-hour session. Above that rate the distractors start looking obviously wrong, which makes the card a recognition test rather than a recall test. Cloze cards run faster (roughly 18 to 20 per hour) because there are no distractors. Image-occlusion runs slower (4 to 6 per hour at best) because the masking has to be placed precisely on each labeled region. The bottleneck is not the typing; it's the few seconds you spend looking back at the slide to confirm what you remember is what the slide actually says.

It's the in-between work. The actual typing of a multiple-choice stem takes 30 to 45 seconds for a fluent writer. The fact extraction (reading the slide carefully enough to be sure what's testable on it) and the distractor writing each take 90 seconds to 3 minutes. A bad distractor (too obviously wrong, too obviously a synonym of the right answer, drawn from a different topic) makes the card useless the moment you've seen it once, which makes the writing of distractors the part most students cut first when they're tired. That's also when card quality collapses.

Sometimes yes, often no, and the dividing line is whether you actually study what you carded. The hand-carding-as-encoding argument only holds if the cards then enter a spaced-repetition rotation that catches the forgetting curve. If the carding session was followed by zero or one passes through the deck, the encoding effect is gone within a week, and the hours spent carding produced no durable gain. The honest math: hand-carding gives you encoding benefit during the writing window plus the cards themselves. Generated cards give you only the cards. If you'd use the saved 25 hours for retrieval practice on the generated cards, you come out ahead.

On a real run with a 90-slide brachial-plexus anatomy deck containing 14 labeled figures: 58 seconds to produce 204 multiple-choice cards plus 14 image-occlusion cards. The terminal output is reproduced on this page. Add 30 seconds for the .apkg export, another 30 seconds for the Anki File > Import flow, and you're at about 2 minutes from PDF on disk to a deck sitting in your collection next to AnKing. The product behind that number scored 81.3 on a held-out three-document eval where Turbolearn scored 57.8.

The image-occlusion gap is roughly 4x wider than the MCQ gap. A hand-made MCQ takes ~7 minutes; a generated MCQ takes ~0.25 seconds (200 cards in 50 seconds is 4 per second). Ratio: about 1,700x. A hand-made image-occlusion card takes ~12 minutes; a generated one takes ~0.5 seconds (14 cards in 8 seconds is roughly 2 per second). Ratio: about 1,440x. But the absolute hours are bigger on image-occlusion: 14 labeled figures saved manually is 2.8 hours of work; 200 MCQs is 23 hours. Both are real time. Most students cut image-occlusion first, then later wonder why they keep missing tagged-diagram questions on the practical.

Yes, downward. Cloze cards skip the distractor-writing step, which is roughly 40 to 50 percent of the per-card time on an MCQ. A fluent cloze writer can sustain 18 to 20 cards an hour from a lecture deck. A 200-card cloze pass on a 90-slide deck lands at 10 to 13 hours. Cloze loses one thing the MCQ has, though: distractor-quality is what enforces conceptual discrimination between near-synonyms. If you mostly study from cloze, you're trading 50 percent fewer hours for somewhat shallower testing on the same fact set.

Worse, materially. A 60-minute YouTube lecture takes 60 minutes just to watch before you start carding. A scanned PDF with no OCR layer either needs OCR (5-10 minutes for a 90-page handout) or has to be transcribed manually as you go (an extra few seconds per card looking at the page). For the automated path, both sources pass through the same pipeline; a YouTube lecture transcript is converted with timestamps preserved on the explain panel, and a scanned PDF goes through OCR first. End-to-end automated time is roughly 90 to 180 seconds depending on length, not 75 to 100 minutes manual.

No. Two real exceptions. First, for board content (Step 1, Step 2 CK, Step 3) the premade decks (AnKing, Zanki, Pepper) exist because thousands of students iterated on the same material for years. No generator beats that yet. Second, for a class where you genuinely use the carding session as your encoding pass and follow it with spaced retrieval, the hand-carding is buying you something. The honest rule is: when the alternative to generating is 'I will not card this lecture at all because I don't have 25 hours this week,' generated cards win every time. When the alternative is rigorous hand-carding followed by daily review, it's a closer call.

The held-out three-document eval scores each tool on factual correctness, stem clarity, distractor quality, and question-type coverage across three independent lecture documents. 81.3 is Studyly, 78.0 Unattle, 68.0 Gauntlet, 57.8 Turbolearn. It matters for time because a 57.8 card set forces you back in to edit obvious failures (wrong distractors, off-topic stems), which adds 30 to 60 seconds per bad card and eats much of the time you thought you were saving. An 81.3 card set still has cards you'll edit, but the edit rate is roughly one in fifteen rather than one in three. Methodology is on studyly.io/quality.