You have the PDF open: a chapter, a paper, a deck of lecture slides. You know rereading it will not hold, and the next move is flashcards. So you face the same two bad options everyone does.
You make the cards by hand, typing a front and a back for every idea until somewhere around page four you quietly give up. Or you run the PDF through a "PDF to flashcards" converter, get a pile of cards in seconds, and discover they are flat, generic, and sitting in yet another app you will stop opening.
Making cards from a PDF without typing them is a solved problem. Making cards from a PDF that you actually remember is the part still worth getting right.
Why "PDF to flashcards" usually disappoints
The hand-built route fails on effort. Writing a clean question and answer for every idea in a dense PDF is slow, and the slowness is the reason the back half of the document never gets cards at all.
The generic-converter route fixes the speed and breaks two other things. The cards come out as flat definitions lifted from headings, "What is X? X is...", because that is the easy thing to extract. And the deck arrives detached from the PDF it came from, usually as an export bound for Anki, so the card and the paragraph that explains it live in different applications. A faster way to make recognition cards is still a way to make recognition cards, which is the same trap we pulled apart in why notes don't stick.
The card has to test recall, not echo the page
Before any converter, what decides whether a PDF card works is the question on it.
A weak card echoes the document back at you: "Define diffusion." You read the back, it matches the page you just read, you feel fine, and you have proven nothing. A strong card hides the source and makes you produce the answer: "A cell needs to move a molecule against its concentration gradient. Can diffusion do it, and if not, what has to happen?" One is recognition. The other is retrieval, and retrieval is the part that builds memory. This is the same distinction behind generated cards versus graded recall: a card the AI writes is only useful if answering it makes you retrieve, and something grades what you said.
How to make flashcards from a PDF, step by step
1. Import the PDF so the text is actually readable
The first requirement is that the tool can read the document, including scanned pages and figures. A clean text PDF is easy; a scanned chapter or a photographed page needs OCR to become text at all. Import the PDF so its content becomes real, searchable material the AI can work from, rather than an image it guesses around.
2. Let the AI pull the load-bearing claims, not every heading
The mistake is asking for "cards from this PDF" and accepting whatever comes back, which is usually one shallow card per heading. Most of a document is scaffolding; a fraction of it is the claims that carry the topic. Ask for the ideas worth testing, the ones you would be annoyed to forget, and let the rest go. A dense chapter should yield a focused set of sharp cards, not eighty thin ones you will never finish.
3. Make each card grade your recall against the source
Push every card from "define" toward "explain why," "predict what happens," or "apply it here," then answer from memory in full sentences and have that answer checked against the PDF, not against your own lenient flip. Because the card keeps the source attached, the grade is measured against what the document actually said, instead of your memory of skimming it.
4. Keep the card linked to the page it came from
When the deck lives in a separate app from the PDF, the source becomes orphaned and the card loses its context. Keep each card attached to the passage it came from, so reviewing a card and rereading the relevant paragraph are one motion, and so a card you got wrong sends you back to the exact spot that explains it.
5. Skip the export; let the spacing be built in
The only reason to export PDF cards to Anki is to get spaced repetition. If the cards already sit on an FSRS schedule next to the source, there is nothing to export and nothing to keep in sync. Each card comes back right before you would forget it, with the gaps widening as you prove you remember, which is why the scheduling algorithm matters.
By hand vs PDF-to-Anki converter vs in-notebook recall
| By hand | PDF-to-Anki converter | PDF that tests you back | |
|---|---|---|---|
| Speed | Slow; you quit by page four | Instant | Instant |
| Card quality | As good as your patience | Flat definitions from headings | Recall you produce, not recognize |
| Handles scans | You retype it | Often not | OCR, so scanned pages work |
| Grading | You self-rate the flip | You self-rate the flip | Typed answer checked against the source |
| Source attached | No | No; exported and detached | Linked to the page it came from |
| Coming back | A schedule you maintain | Export to Anki, then maybe | Spaced schedule, built in |
A worked example
Say the PDF is a ten-page paper on how a vaccine trains the immune system.
- Import: the paper becomes a page you can search, even if it is a scan, instead of a file you reopen and reread.
- Extract: the AI pulls the handful of claims that carry it, what an antigen is, what memory cells do, why a second exposure responds faster, what an adjuvant changes, and leaves the citations and methodology boilerplate alone.
- Recall, not recognition: instead of "What is a memory cell?", the card asks "why does a second exposure to the same pathogen produce a faster, stronger response?" You answer in your own words and are told what you missed, checked against the paper.
- Schedule: the memory-cell card returns in a few days, then a week, then a month, while you move on.
A month later you do not have an exported deck you forgot to open. You have a paper that still answers when it is closed. The same approach works whether your source is a PDF, a page of notes you turn into recall, or a lecture.
Where this fits
You can assemble this from parts: an OCR tool, a card generator, a flashcard app, and a scheduler. It works, and the assembling is the reason the deck dies by the second chapter.
Anti-Agent does it in one place. You import the PDF, it becomes a readable, searchable page, the load-bearing ideas turn into graded recall instead of flat definitions, and everything sits on an FSRS schedule with no export step. It is the same bet as an AI Anki alternative that grades your answer, aimed at the documents you are already studying from.
Frequently asked questions
How do I make flashcards from a PDF automatically? Import the PDF so its text is extracted, including scanned pages, have the AI pull only the ideas worth testing, and turn each into a question you answer from memory. The useful version also grades your answer against the source and schedules the reviews, so the automation covers the parts you would otherwise skip.
Can I make Anki cards from a PDF? Yes, several converters export PDF cards to Anki. The catch is that they tend to produce flat definition cards detached from the source, and you still self-rate the flip. If the cards already live on a spaced schedule next to the PDF, there is no need to export to Anki at all.
Do PDF flashcard converters work on scanned documents? Many do not, because a scan is an image, not text. You need OCR to turn a scanned or photographed page into something a generator can read. Check that whatever you use actually extracts text from scans, or the cards will be guesswork.
Are flashcards made from a PDF actually effective? Only if they make you retrieve. Cards that echo a heading back at you are barely better than rereading the PDF. Cards that hide the answer, make you produce it, and grade you against the source are among the most effective ways to study.
How many cards should I make from a PDF? Fewer than the document seems to demand. A dense chapter usually has a handful of load-bearing ideas worth testing. A small set of sharp recall questions you finish beats a huge export you abandon.
The bottom line
Making flashcards from a PDF by hand is the chore that ends the deck, and generic converters fix the speed while handing you flat cards you still flip and self-rate, detached from the source. Neither one gets you to the thing that matters, which is recall you have to produce and an honest check on whether you produced it.
So use AI for the extraction, and build the recall and the spacing in on purpose: pull the load-bearing ideas, make each card grade your answer against the page it came from, and let the schedule bring it back. Turn your first PDF into graded recall, and see what you actually kept a week later.
FSRS vs SM-2: why the modern spaced-repetition algorithm beats the classic
SM-2 is the spaced-repetition algorithm from 1987 that ran Anki for decades. FSRS is the machine-learning model that replaced it, fit to more than 700 million real reviews. Here is how each one works, why FSRS schedules fewer reviews for the same retention, and when the difference actually matters.
NotebookLM has no spaced repetition. This alternative is built around it.
NotebookLM is great at chatting with your sources, but nothing comes back and nothing tests you. Here is a NotebookLM alternative with flashcards and spaced repetition, where your sources turn into graded recall that resurfaces on a schedule.