Why the raw output is never the deck
Ask any generator for flashcards from a 40-page PDF and you'll get several hundred. They will include: definitions worth knowing, the same definition three times in different words, questions about the document rather than the subject ("What does Figure 3 show?"), and cards whose answer is a paragraph. The first category is your deck. The rest is what turns a review queue into a chore.
The cost of a bad card is not zero, it is compounding. A card you can't answer gets marked wrong, so the scheduler shows it more often, so the queue grows, so you skip a day, so the queue grows further. Decks die from junk cards, not from difficulty. This is the failure mode described in AI flashcard quality.
The five-test cull
Run every generated card past these. It takes about four seconds per card and it is the difference between a deck you keep and one you abandon.
| Test | Delete the card if… |
|---|---|
| One answer | The question has more than one defensible answer, or the answer is longer than about fifteen words. |
| Standalone | It only makes sense with the PDF open — "According to the text…", "In this chapter…", any reference to a figure number. |
| Not free | You can answer it without knowing anything, by grammar or elimination. |
| Not a duplicate | Another card already tests the same fact from the same direction. |
| Worth the interval | You would not care if you forgot it. Ruthlessness here is what keeps the queue small enough to clear daily. |
Splitting the fat cards instead of deleting them
The most common bad card is one whose answer is a list: "What are the causes of X?" with six bullets. It fails the one-answer test, but the content is genuinely examinable. Don't delete it — split it.
- Turn the list into cloze cards: "The causes of X are A, B, C, D, E and ___" — one card per hidden item, each with a single answer.
- Add a count card: "How many recognised causes of X are there?" This is the one that saves you in the exam when you've listed four and can't remember whether there were five.
- Keep one whole-list card for the week before the exam if the exam genuinely asks you to produce the whole list — but tag it so it isn't in your daily queue.
The full workflow
- 1
Check the PDF is actually text, not pictures of text
A scanned handout or a photographed page is an image. If you can't select the words in a PDF reader, no extractor can read them either without OCR. Tools differ sharply here — some silently produce nothing useful, which looks like a bad tool and is actually a bad input.
- 2
Cut the PDF down before you generate
Drop the title page, the reading list, the acknowledgements and the appendices you were never assigned. Every page you leave in produces cards you'll have to delete. Ten seconds here saves ten minutes of culling.
- 3
Generate, then walk the list once with a delete key
First pass is deletion only — don't edit, don't improve, just delete anything failing the five tests. Expect to remove half to two thirds. If you're deleting fewer than a third, you're being too kind.
- 4
Second pass: split the list-shaped survivors
Any card whose answer has commas in it. Turn it into clozes as above. This is where the deck goes from mediocre to good.
- 5
Third pass: add the questions the PDF didn't ask
The document explains; the exam applies. Add three or four "why" and "what would happen if" cards per topic yourself. These are always the highest-yield cards in the deck and no extractor generates them well.
- 6
Put it on a real scheduler and review daily
A deck reviewed on a proper interval algorithm costs a few minutes a day and holds for months. A deck reviewed by flipping through the whole stack costs an hour and holds for a week.
Textbook chapters versus lecture slides versus papers
| Source | Cards per 10 pages | The trap |
|---|---|---|
| Lecture slides | 25–40 — dense, already distilled | Bullet fragments make questions with no single answer. Expect heavy splitting. |
| Textbook chapter | 15–25 after culling | Enormous volume of context prose that generates trivia. Cull hardest here. |
| Journal paper | 5–10, mostly method and conclusion | Almost all of a paper is not examinable. Cards about the specific study rarely earn their place. |
| Your own handwritten notes | 20–30 | Needs handwriting recognition first; abbreviations you invented won't be understood. |
When not to bother
Flashcards are for facts you must produce from memory, quickly, without reasoning. They are a poor fit for material where the exam asks for an argument, a derivation or a judgement. A law student's case-name-to-principle cards are excellent; "cards" summarising an essay's structure are wasted effort that would be better spent writing the essay.
If the PDF is a reading you need to *understand* rather than *know*, summarise and self-explain instead — self-explanation buys more per minute on that kind of material than any deck will.