Study notes and flashcards from a PDF: a loop that works

Open Study notes from PDF, add a text PDF and pick one of three kinds. Each run costs 1 Gem on GPT-4o mini, and you see the cost before you start.

• Notes. Headings that follow the document's own topics, the main points under each, and key terms with short definitions.
• Flashcards. 15 to 25 cards, one per line, written as "Q: question A: answer".
• Quiz. 10 multiple-choice questions with options A to D. The answers are listed at the end, not next to the questions, so you can cover them up.

Every result has Copy and Download .md buttons. That is the whole export: Markdown text you paste or save wherever you study.

Export the deck from your presentation software as a PDF, then run Notes on it. Slides exported that way carry real text, so the model can read titles and bullets.

The headings in the result follow the topics of the deck, which is handy: you get an outline that matches the order your lecturer taught in. The key terms list at the bottom of each section is the part to read twice.

Slides are thin on words. A bullet like "Cost of capital" tells the model little, and the notes will be thin there too. Fill the gaps yourself, or add the textbook chapter and ask a follow-up question about that topic.

A big deck is a problem of size. Split it by lecture first with the free Split PDF tool, which runs in your browser. One lecture per file gives sharper notes and keeps each PDF within about 5.7 MB.

Text is extracted from the PDF. There is no OCR. That has three consequences for slides:

• Image-only slides. If a slide is a screenshot or a picture of text, there is no text layer, and the model sees nothing there.
• Scanned handouts. A scan of a printed handout is a photo of a page. If you cannot select the words in your viewer, the tool cannot read them.
• Diagrams. Charts, flow diagrams and annotated figures are not read. If a slide's meaning sits in a diagram, the notes will skip it. Write a line about it yourself.

A quick test before you run anything: try to select a sentence on the slide. If you can, the model can read it.

The example below is fictional. Say a finance lecture deck has a slide titled "Cost of capital" with three bullets: debt, equity and a weighted average. Notes on that slide come out short, as warned above.

Here is how the loop goes on that one lecture:

• Run Notes on the lecture file. The result has a heading for cost of capital, the three bullets under it and a key terms list with a line for each term.
• Run the Quiz with the notes closed. Say you miss the question on the weighted average, because the slide never explained how the weights are chosen.
• Ask in chat for cards on that gap. For example: "Make flashcards only for the weighted average. Use the wording of the document." If the deck is silent on the weights, the cards will be silent too, and the model should say so.
• Add the textbook chapter as a second file. Ask how the weights are chosen. Now the answer has real text to stand on.

The gap showed up in the quiz, not while you were reading, which is the point of the loop. Write the missing line on the slide yourself, then make the card from it.

Most weak cards come from how the run is used, not from the tool.

• Running a whole deck in one go. A big file gives a thin spread of cards across many topics. One lecture per run gives cards that cover it properly.
• Cards that ask for a list. A question like "Name the stages" is hard to grade when you answer it from memory. Ask for harder cards with why and how, one fact per card.
• Skipping the check. A card can state a wrong fact in a confident tone. Compare each card with the slide or page before you learn it.
• Making cards for everything. A full set feels thorough and wastes time. Let the quiz show which topics need cards.
• Importing without a look. After the Anki workaround, open the first few cards in Anki and check that the question and the answer landed in separate fields.

And one habit worth keeping: put the lecture title in the file name before you add it. When you have several results saved as .md, you will know which is which.

Reading the notes feels productive and teaches less than you think. Testing yourself does the work. Here is a loop that uses all three outputs:

• Notes first. Run Notes on one lecture or chapter and read them once, slowly, with the slides open.
• Quiz without notes. Close the notes. Run the Quiz and answer all 10 questions before you look at the answers at the end.
• Cards for what you missed. For each wrong answer, ask for flashcards on that topic only. Study those cards, not the whole deck.

Then repeat with the next lecture. A day later, take a fresh quiz on the same material. Because a run costs 1 Gem, asking again is cheap.

One caution: a model can write a confident card with a wrong fact. Skim the cards against the slide or page before you memorize them.

There is no Anki export, and no other flashcard export. You only get Copy and Download .md. The workaround is a few minutes of text editing:

• Copy the cards from the result.
• Turn each line into "question, a tab, answer". Replace " A: " with a tab, and strip "Q: " from the start of each line.
• Save the file as .txt.
• In Anki, choose File > Import, pick the file and set Tab as the separator.

So a line like "Q: What does a notice period set? A: How long before leaving you must tell the other side" becomes the question, a tab character, then the answer.

Any text editor with find and replace does this in two passes. You can also ask the model in chat to output the cards in that layout, but check that it used real tabs.

A Study run is 1 Gem on GPT-4o mini. New accounts start with 30 Gems, and pricing lists packs and plans. Follow-up questions in chat are charged by the model you pick.

The PDF's extracted text is sent to that model. Splitting a deck with Split PDF does not upload anything: it runs on your device and costs no Gems. For a quick overview before you make notes, see how to summarize a long PDF. For your own questions about the material, see how to chat with a PDF.