PDF to Anki: Four Routes That Work, and Where Each Stops

The job, defined. PDF to Anki is the work of getting a fixed document, whether a vocabulary list, a textbook glossary, lecture slides, or a scanned reader, out of its read-only container and into cards a scheduler will resurface on a calendar. One number frames every route: Anki's own importer reads no PDFs at all, so a 40-page glossary at 20 rows a page has to become 800 lines of plain text somewhere outside Anki before Anki sees any of it. A small two-column list belongs in ChatGPT plus a CSV; a scanned chapter needs a converter; a long exam list needs selection before conversion; and if you don't specifically need Anki, the PDF can become cards where you import it, with no roundtrip. Same disclosure as every guide in this series: I build Piccard, a flashcard tool whose PDF import is the fourth route below. Grade my homework accordingly.

Can Anki import a PDF directly?

No, and the manual is blunt about it. The page documenting text imports opens by narrowing what it accepts:

> "The files must be plain text (myfile.txt)."

The next line rules out the office formats on the way past:

> "Other formats like myfile.xls, myfile.rtf, myfile.doc must be saved as a plain text file first."

(Anki Manual, Text Files)

A PDF is not on that list because it sits further from text than a .doc does. It is a print layout that happens to live on your screen, sometimes with a real text layer inside, sometimes as photographed shapes that merely look like words. So every "PDF to Anki" route is secretly two jobs: extraction, meaning get rows out of the layout, and import, meaning get text into Anki. The import half is free and boring. The extraction half is where the routes disagree, and where the money and the mistakes sit.

What do the converter sites give you?

The sites ranking for this search are small, independent tools, and they run the extraction for you, then hand back an .apkg deck that opens in Anki on a double-click. Their differences live in what happens when your PDF turns out to be a scan.

The name a lot of people type is gone. pdf2anki is the brand a chunk of this search is navigating toward, and typing pdf2anki.com today lands you somewhere else: the domain returns a permanent redirect to Memrizz, an AI flashcard site. The tool still exists in the way a tool absorbed by another company still exists. The name alone no longer tells you what you will get, so read what you are looking at before uploading anything.

Text layer or image cards. 2anki's PDF converter is explicit about the fork:

> "2anki reads the text into question-and-answer cards, falls back to image cards when a page has no text"

That fallback is the honest handling of scanned pages. A photograph of a page carries no text to extract, so the options are OCR first or cards whose front is the page image. Image cards are real cards; whether they are useful cards depends on whether you wanted to read whole pages during review.

Free tiers measured in pages. AnkiDecks states its cap directly:

> "The free plan processes up to 4 pages per PDF."

The same page allows 4 decks a month on the free plan and sends anything bigger to Pro. Four pages is a handout. A 30-page textbook chapter is eight four-page uploads spread across two months of allowance, or the paid plan, so run that arithmetic while the PDF is still closed.

What none of them do is choose. A converter will happily deck all 800 rows of your glossary, and an 800-card deck built in one afternoon usually drowns in its own review backlog inside the first month. The fuller version of that arithmetic, run on lecture transcripts, is in Lecture to Anki. The selection problem there is the same problem here, one artifact over.

Can ChatGPT make Anki cards from a PDF?

Yes, in the sense that ChatGPT writes the text file Anki wants. It never touches Anki itself, and the gap between those two sentences is the whole route. The steps: paste your rows into the chat (attaching the PDF also works for documents ChatGPT can read, but pasting shows you exactly what the model saw), ask for one line per card in front,back form with no header row, save the output as a .txt, and import. On the format you are aiming at, the manual says:

> "Anki tries to guess which field separator (commas, tabs, etc.) your file uses."

Save as UTF-8, which the same page requires, and the guess resolves on the first try.

It breaks on a bare word list, with no definitions on the rows: ChatGPT invents the back of each card, and its inventions are plausible in the specific way that survives a review session and then fails an exam. A two-column glossary is the good case: the meaning is already on the row, and all the model does is reformat. Long lists are the other bad case: ask for 2,000 rows and you will often receive a few hundred correct ones, a summary of the pattern, and an apology. The practical ceiling is roughly the list you would have typed by hand anyway, minus the typing.

The prompt that selects which rows deserve cards is a different prompt from the one that formats them, and confusing the two is how 800-card decks get built in an afternoon. Formatting is what the chat is good at. Selection it will only do if you ask, and it will do it on your behalf with no memory of what you actually know.

What does skipping the roundtrip look like?

The fourth route does not end in Anki, so the .apkg step disappears: the PDF becomes cards inside the tool that read it. This is Piccard's import path; what follows is the same flow I ran with a real glossary.

The import. Drop in the PDF (a .docx or pasted text takes the same door). The mining step hands back a ranked checklist rather than your rows echoed: twenty-odd leading candidates off a list-sized file, as many as a hundred off a long one that the importer splits four ways. Each row carries a study-worth tag, top tier pre-checked, select-all within reach. The tags rank what deserves study; they cannot know which words you came in with, so strike those yourself.

The confirm. A confirm enrolls at most 40. A ceiling round takes three presses, a list round one. Feeding the same text through again surfaces the same leaders, so the way down a long list is pasting in what comes next. Energy runs about one unit per kept card; a 1,600-row glossary kept whole costs about 1,600 energy end to end.

The schedule. Reviews run on FSRS-6 — the scheduling algorithm Anki's own versions now carry — with the defaults welded in place. Nothing leaves for Anki: no .apkg, no export step. Deck-list loyalists should take one of the first two routes instead.

Where cards can go that a PDF cannot follow. The same tool's other door is captioned video: a YouTube upload with a caption track shows its lines in the extension, and the line you pick is the only text that gets proposed from. Twenty cards per confirm at most, and every card born there carries its moment — in review, the card summons its source, the video at the second the line was spoken. Captionless video leaves nothing to select, and the tool says so. For an exam with a listening half, that attachment is the thing a printed glossary row can never carry. The vocabulary-list version of that argument, with one real exam's numbers, is in TOPIK Vocabulary.

Which route for which PDF?

Match the route to the artifact before opening any upload dialog:

The short version

Anki reads text and nothing else, per the manual's own first line, so "PDF to Anki" always means somebody extracts, then Anki imports. A small clean list goes through ChatGPT and a CSV — check row 20, check the last row. A scan needs a converter, which means accepting image cards or running OCR first; remember that pdf2anki now redirects to Memrizz before judging either by the other's reputation. Long lists die of backlog. Select before converting, however you convert. And if Anki was only ever the means, the PDF can become cards directly: twenty-odd candidates from a list, up to a hundred from a long document, 40 a confirm, FSRS-6 underneath, no roundtrip. The border-crossing miseries that come with file pipelines into Anki have their own write-up in Exporting Language Reactor to Anki — the .csv you just built crosses the same customs border, one country over.

The free tier is 150 energy (about 150 cards), then prepaid top-ups at $5 / $10 / $20; no subscription, and topped-up energy never expires. To start: Piccard on the Chrome Web Store.

All guides: Piccard Blog.