Creating Anki cards from PDFs: slides and lecture notes into cards

The lecture handout is a PDF, a hundred and twenty slides, half of them diagrams. And you sit in front of Anki copying sentence by sentence. After slide twelve you've had enough. We know that.

There are five ways from PDF to card, and each has a point where it hurts. Here is which one suits what, where the technology gets stuck, and what you should know about sharing decks made from lecture material.

Way 1: Copy and paste

The classic. Select text in the PDF, copy it, paste it into Anki's front and back fields, or edit the sentence as a fill-in-the-blank (cloze). The Anki manual documents both note types and pasting from the clipboard.

Good for: definitions, mnemonic sentences, anything where one sentence carries exactly one statement. Hurts with: tables, formulas and anything running across two columns (more on that shortly).

Honestly, the slowest method is not the worst one either. Anyone who formulates things themselves processes the material; more on that below under the generation effect.

Way 2: Diagrams with Image Occlusion

For slides full of figures, Image Occlusion is the best way. Since Anki 23.10 it is built in without an add-on. In the Add window you choose the note type "Image Occlusion", insert the image and draw masks: rectangle, ellipse or polygon. Anki turns each shape or group of shapes into a card.

There are two modes:

• Hide All, Guess One: All marked areas are covered, one is asked.
• Hide One, Guess One: Only the asked area is covered, the rest stays visible.

The catch: the process always starts with an image and masks that you set yourself. Image Occlusion cannot automatically turn an entire PDF into study questions. The older add-on Image Occlusion Enhanced does essentially the same thing; you should check yourself whether it works with your Anki version.

Way 3: Spreadsheet first, then CSV import

If you prefer working in a spreadsheet on your laptop: question in column A, answer in column B, save as CSV or text file, import into Anki. According to the manual:

• Fields are separated by comma, semicolon or tab.
• The file must be UTF-8. An Excel file must first be exported as text.
• The first line determines the number of fields. Missing fields stay empty, extra ones are ignored.
• HTML can be allowed. Images and audio go into the collection.media folder and you reference them in the fields.

Important: the import only takes over what is already structured. CSV does not spare you the actual work of turning the lecture handout into good questions. "Converting PDF to CSV" and "writing good cards" are two different tasks.

Way 4: AI in Anki or in the browser

This is where it gets convenient. The add-on AnkiBrain documents PDF import for AI analysis and automatic card creation from whole documents, according to its changelog now also with images from PDFs and automatic Image Occlusion. Web services such as AnkiDecks and Ankigenix promise PDF upload, automatic cards and export as .apkg.

These are manufacturer claims. According to our research, there are no independent quality comparisons of these services. What AI cards can fundamentally do and where they miss the mark is covered in our article Create flashcards with AI. In short: check every card against the source.

Why PDF text so often comes out broken

Anyone who has copied a table from a PDF knows the result: numbers in the wrong order, words without spaces, "fi" as a strange character. That is not your fault. The documentation of the PDF library pypdf describes the typical problems:

• Tables often consist of individually positioned text. Rows and columns must first be reconstructed.
• Reading direction, spaces, ligatures, footnotes and formulas regularly go wrong during extraction.
• Pure scans have no text layer. Normal text extraction finds nothing there. Scans already processed with OCR often have an invisible text layer, which is not necessarily correct.

And OCR is not error-free. If the text recognition and then the extraction both go wrong, the errors add up. No reliable error rate for German lecture slides exists.

On the iPhone, Apple helps in two places. With Live Text you can select and copy recognized text from images. Apps can recognize text in images via Apple's Vision framework, with language correction and custom technical term lists; according to Apple, this recognition runs on the device. Apple does not guarantee error-free formulas.

Way 5: With checkst directly on iPhone or iPad

checkst has a split view for exactly this case: at the top you see the PDF page (or an image), at the bottom the card editor. You move the divider depending on whether you want to see more slide or more card. With the crop selection you draw a frame around a diagram and take it as an image into the card. That way you write the question yourself without jumping between two apps, and never have to copy broken PDF text at all.

If you want it faster, checkst reads the text of a PDF, uses text recognition for images, and lets the AI suggest cards from it: Basic, cloze or mixed. You see each suggestion individually and save only what you want to keep. Finished decks can be exported as .apkg and used further in Anki.

Generation effect: is writing it yourself worth it?

The question behind all five ways: is it bad if a machine makes the cards?

Research on the generation effect says: what you produce yourself, you remember better than what you only read. The meta-analysis by Bertsch et al. (2007) across 86 studies and 445 effect sizes found an average advantage of 0.40 standard deviations. For text learning, Schindler and Richter (2023) across 20 studies with N = 3,551 found g = 0.41. Sounds solid, but: older studies showed larger effects than newer ones (0.63 versus 0.28 after the year 2000).

And then it gets uncomfortable. Schindler and Richter (2025) found no robust general advantage in seven laboratory experiments. Only in one experiment with 152 people was the generation condition better on a multiple-choice test, and only there. None of the experiments was preregistered.

Above all, none of these studies tests your real workflow of "lecture PDF, self-written versus AI cards". The statement "self-made cards are learned 40% better" is not supported by these numbers. A sensible middle path: use AI drafts, but shorten and reformulate them. In doing so you still process the material.

May I share the deck?

Now the part nobody reads until the email from the chair arrives. This is not legal advice, but what the law says.

• Slides can be protected. § 2 UrhG lists, among other things, literary works and scientific drawings and tables. The prerequisite is a personal intellectual creation. Whether your specific slide is one cannot be said across the board.
• § 60a, teaching and instruction: For illustration in teaching, for non-commercial purposes, up to 15% of a published work may in principle be used, primarily by teachers and participants of the respective course. Illustrations and works of small scope may be used in full under the conditions.
• § 60c, research: For one's own scientific research, up to 75% may be reproduced. However, this does not permit making it publicly available.
• A private copy is not publication. § 53 permits private copies under conditions, but in principle prohibits their distribution. A publicly accessible deck also touches the right of making available to the public under § 19a.
• Naming sources is required by § 63. However, a source citation alone does not make a use permitted.

"It's for studying" or "it's free" is in any case not sufficient justification for a public upload. In practice this means: cards in your own words are less problematic than copied slides, and a deck for your study group from the same course is something different from a public download for everyone. In checkst you share decks in groups that others can join only by invitation. What you upload there, you should still select according to these rules.

Which way for which lecture handout?

• Text-heavy lecture handout, digitally created: copying or AI suggestions, then shortening.
• Slides full of diagrams: image crops or Image Occlusion.
• Scanned lecture handout: OCR for text, or straight to image crops when it concerns figures.
• Formulas and tables: take over as images, not as text.
• Everything on iPhone, without laptop: checkst with split view.

Whichever way: the card is only finished when it asks exactly one thing and you understand it without the PDF.