How to Make Useful Flashcards From a Course PDF With AI—With Page Sources

You can start with one section and a small trial deck. This guide covers preparing different kinds of PDFs, writing questions that work without the surrounding chapter, checking calculations and missing conditions, and exporting cards to Anki or Quizlet. The six-page sample below gives you a source you can open alongside the cards.

Start with five cards from one section

If your PDF has readable text and you only need a small deck, start here. The later sections explain what to change for scans, diagrams, calculations, and longer reading assignments.

1. Choose a short section you have already read. Copy its text and add the document name and printed page number before each page. Preserve any condition, footnote, or formula needed to understand it.
2. Start a new chat and paste the labelled text with the prompt below. You can also attach a PDF in ChatGPT using its file-upload feature; specify the assigned pages and verify their numbering.
3. Read each question with the answer hidden. Attempt an answer, then open the cited page to check the wording, conditions, and any calculation. Rewrite questions that need missing context and remove claims you cannot find.
4. Keep the useful cards and use the export instructions below to put them in your study app. Work through the next section after you know this small deck is usable.

I am studying [topic]. My objective is [what I need to know or do].
Use only the course section below and my page labels.
Create up to five flashcards that cover distinct parts of this objective.
Return a table with Front, Back, and Source.

Each front must ask one clear question and make sense on its own.
Keep the answer brief, with any condition needed for it to be correct.
Use questions that require recalling or applying the material;
do not copy paragraphs onto the back.
Cite the document and page that supports each answer.
If the text is incomplete, unreadable, or cannot support a card,
list the problem separately. Do not invent an answer or page number.

[Paste the course text with a document name and page labels here.]

A table of sources is optional for this small start. Use one when definitions depend on several pages, a calculation has important assumptions, or you need to track coverage across chapters. You do not need to finish every step in this guide before studying five checked cards.

1. Choose an input method that preserves the material

Open the PDF before opening the AI chat. Try selecting a paragraph, copying it, and pasting it into a plain text editor. Compare the result with the page. Check a formula or table as well: readable prose does not establish that mathematical symbols, column order, or footnotes survived extraction.

For a scan in Acrobat, Adobe's OCR instructions use All tools > Scan & OCR, then In this file. Select the page range and recognition language, and run Recognize Text. Review the resulting text against the scan. OCR gives you searchable text; you still need to correct errors before turning it into answers.

File attachment and visual understanding are different capabilities. OpenAI's PDF visual retrieval documentation distinguishes ChatGPT Enterprise visual retrieval from text-only processing, including the treatment of GPT Knowledge and Project Files. Check the documented behaviour for your upload route. If a diagram supplies the answer, provide that diagram explicitly and verify the labels yourself.

Use material you are permitted to process in the chosen tool. If a packet contains student names, assessment records, or restricted material, use an approved environment or remove that content first. A locked file needs an authorized readable copy from its owner. Generating a personal deck and distributing the source pages are separate decisions.

2. Set the page convention and study scope

A citation such as "page 17" is ambiguous when the PDF viewer says 21 and the printed footer says 17. Record both when they differ. For example, a card source might read Biology packet, edition B, printed p. 17 (PDF page 21). That is an illustrative format, not the numbering of the six-page sample.

When you paste text, add labels yourself. Put a label before each page and preserve headings and footnotes. If you extract a chapter into a new file, retain the original page labels; the new file's first page may have been page 89 of the book. Keep the document title or a short document ID with every citation.

[Document: Queueing and Backpressure, version 1.0]
[PDF page 2; printed page 2]
[Paste the complete relevant passage here.]

[PDF page 3; printed page 3]
[Paste the next page, including its conditions and unit check.]

Then tell the model what the deck should cover: the assigned pages, learning objectives, expected answer language, and assessment task. A vocabulary quiz needs different questions from an exam that asks you to explain a mechanism or calculate a result. Use your syllabus and instructor's objectives to choose coverage. The AI cannot establish what will appear on an exam from the PDF alone.

I am studying [course/topic] for [assessment task].
Use only [document names and assigned page ranges].
My learning objectives are [paste objectives].
Use [language] for questions and answers. Preserve technical terms.

Before making cards, list the supplied pages or sections, their topics,
and any unreadable, missing, or image-dependent material.
Use my explicit page labels; do not infer page numbers from the text order.
Identify anything that needs another source or a clearer image.
Do not write cards yet.

Compare the inventory with the PDF. A page listed by the model only shows that it recognized a label; it does not prove that it processed every sentence on that page. Check the last section, a middle section, and any page with unusual layout. Repair missing input before proceeding.

3. Choose questions that match what you need to learn

Choose material from the course objectives and questions you struggled to answer after reading. Good candidates include a term you must define, two concepts you confuse, an assumption that changes when a formula applies, and a step you cannot explain. Skip repeated introductions, decorative examples, and facts outside the assigned scope. A page does not need a card just because it appears in the PDF.

The five-card start lets you check the question style before generating more. For a full section, add cards until its selected objectives are covered without repeating the same task. The 15-card sample below illustrates several question types; its count is not a target for your chapter.

Reverse cards need their own wording. Turning "What is backpressure?" around into "What is signalling upstream producers to slow down called?" can work. Reversing a paragraph-length answer usually produces a vague question with several possible terms. Generate the reverse only when that recognition task matters to the course.

A cloze card hides part of a sentence. Anki's Cloze note type uses syntax such as {{c1::answer}}. For this sample, In the simplified single-stage model, utilization is {{c1::λ/μ}}. tests the formula. Put the page source in the answer-side extra information. Use a normal question when a surrounding sentence would reveal the answer, or when the task is to explain why the formula applies.

4. Keep a source table for dense or long material

For more complex material, record the passage and condition supporting each selected concept before making cards. This helps when a definition has exceptions or a worked example draws on earlier pages. A spreadsheet or plain table is enough. Give concepts IDs only if you need to track them across a larger deck.

For a comparison that needs two pages, cite both. A definition page cannot support a numerical threshold simply because it discusses the same concept. When the passage only lists a term, hold any card that asks for an explanation absent from the material.

Extract the learning targets for the objectives I supplied.
For each target return: target ID, document/version, page label,
supporting passage, proposed meaning, and required qualification.
Separate independently testable facts.
Keep formulas, units, exceptions, and scope limits.
Mark targets that require unavailable images or additional material.
Use only the supplied source. Do not draft cards yet.

Review the table before card generation. Fixing one mistaken source statement here prevents it from becoming several wrong cards. A card's source entry should let you reopen the supporting passage without searching the whole chapter.

5. Draft a trial deck with complete, answerable questions

Each front should let you attempt an answer without opening the PDF or remembering the previous card. Include the model, process, or scenario when the same term has several meanings. Put the citation on the back where possible; a chapter heading or page title on the front can accidentally reveal the answer.

Using the approved source table, draft a small trial deck.
Cover [selected objectives] with [appropriate question types].
Return: card_id, type, front, back, document/version, page_source,
supporting target IDs, and tags.

Give each card one retrieval task. A comparison may test one distinction.
Split independent definitions and separate diagnostic questions.
Make each front understandable on its own. Avoid answer leakage.
Keep the essential answer concise; put necessary explanation after it.
For calculations show formula, substitution, result, units, and conditions.
For unsupported targets, return a separate unresolved list, not a guess.
Do not invent thresholds, exam predictions, or source page numbers.

These edited examples come from the sample PDF. The packet uses a simplified queueing model, so the questions name that scope and keep its conditions.

A question such as "What are λ, μ, and ρ?" combines three concepts. Split it so you can tell which definition you missed. A question asking you to diagnose two unrelated failures has the same problem. Check whether an earlier card already tests either diagnosis before adding more. The 15-card sample deck gives each variable a separate question and includes definitions, comparisons, calculations, and questions about the model's limits.

Keep calculation cards and formula-recall cards separate when both skills matter. A front that supplies the formula tests substitution or arithmetic. A front that asks which relation applies tests formula selection. Label your learning target accordingly so a correct numerical answer does not stand in for understanding when the relation is valid.

6. Check the answer, the source, and the question

Open the cited page for every trial card. Read the front with the back hidden, attempt an answer, then compare the back with the passage. Check the full statement: an answer may quote a correct formula while dropping its stability condition. A page citation is useful only when that page supports the answer.

For a second AI review, open a fresh chat and supply the exact source passages and deck again. Ask for the supporting passage and any specific problem. A fresh conversation avoids carrying over the drafter's explanation, but it is still another model response; verify disputed points against the PDF yourself.

Review this deck against the supplied PDF passages.
For each card, give its ID, supporting passage and page, and any problem.
Check unsupported claims, wrong pages, missing conditions, unit errors,
ambiguous fronts, leaked answers, and semantic overlap with other cards.
Flag bundled independent questions for splitting.
Suggest keep, revise, split, merge, or remove, with a specific reason.
Do not use the draft answer as evidence. Do not fill source gaps.

In the sample exercise, we deliberately added two wrong answers to show what a source check should catch: a p95 latency below 200 ms that the PDF never gives, and an answer that calls Little's Law backpressure. The ChatGPT review rejected both. For your deck, ask where each disputed claim appears and compare it with that passage.

Recalculate numerical answers yourself. In the sample, 12 requests/min × 0.25 min = 3 requests. If you rewrite the question with the same time expressed as 15 seconds, first convert 15/60 = 0.25 min. Multiplying 12 by 15 without conversion gives the wrong result for a per-minute rate. Keep the conversion on the back if the card tests unit handling.

Two fronts can use different words yet require the same answer. Compare the task each card asks you to perform, rather than only looking for identical text. A formula-recall card and a calculation using that formula test different skills; two reworded requests for the same definition usually do not. Software that checks formatting or repeated text cannot replace this review.

7. Adjust the deck for different course material

Fact-heavy chapters and terminology

Select the definitions, distinctions, and exceptions your objectives require. Keep qualifiers such as a time period, classification system, or edition when they change the answer. Avoid a card that asks for every item in a long paragraph. If the course requires a fixed list, state the expected set or count and decide whether you need a whole-list card, individual items, or both.

For bilingual study, preserve the course's technical term alongside its translation on the back. Specify whether you want to recognize the original term, produce it from a translation, or explain the concept in your study language. Keep those tasks separate. Check translations against the course glossary before importing them.

Mathematics, formulas, and procedures

Use separate questions for choosing a method, stating an assumption, performing a calculation, and explaining a step. Preserve symbols and units exactly. When a formula's derivation spans several pages, use a question about one transition with enough context to reproduce it, and cite the pages needed for that transition.

For an ordered procedure, ask what to do at a specified stage and why, or which condition makes the next step valid. Keep one complete practice problem outside the deck if the assessment requires sustained reasoning. A set of short answers cannot show whether you can organize a full solution without prompts.

If you ask AI to write a new numerical practice question, label it as additional practice. Its numbers are not quotations from the PDF. Keep the original source for the method, show the changed inputs, and work through the answer yourself. Reject a variation that violates the method's assumptions or silently changes what the variables mean.

Diagrams and visual relationships

A text card can test a relationship described in a caption. Identifying a part by appearance needs the image. Crop only enough to remove irrelevant surroundings; retain orientation, scale, and labels needed to interpret it. Put the figure number and page on the back. Keep the image file with the deck rather than relying on a picture that remains in the AI chat.

Provide the figure, its caption, and the explanation it refers to in the same message. Ask for cards about only the labels and relationships that are legible and supported by that explanation. If an arrow, legend, or axis label is unclear, supply a clearer crop or transcribe it from the source before generating the card. A model's description of a blurry image is not enough to establish the answer.

For label identification, Anki's Image Occlusion notes can hide parts of an image. Mask the answer without concealing the context needed to identify it. For a question about a relationship, write a front that names the relationship being tested; hiding a label alone may only test its name.

Essay courses, explanations, and code

For essay-based assessments, use cards to retrieve the elements of an argument: a claim, its stated evidence, an objection, or the connection between two ideas. Attribute a position to the author or lecture that presents it. Keep a practice outline or short written response alongside the cards to check whether you can assemble those elements into an argument.

For code examples, preserve indentation and enough surrounding code to make the question answerable. Specify the language or version if the source does. A prediction card should include its input; a debugging card should ask about one demonstrated fault. Verify output by running the authorized example when practical, or hold the card if the source does not establish the answer.

These question templates show how to adapt the same method beyond the queueing sample. Replace the bracketed text with details from your course; the answer must come from the cited passage or figure.

8. Handle long textbooks, multiple files, and conflicting sources

Work through assigned sections in batches that you can check. Split at a section boundary where possible, and include the preceding definition when a later page refers back to it. Avoid splitting halfway through a worked example, table, or figure explanation. There is no universal page count per batch: a dense equation page may need more attention than several pages of prose.

Keep a coverage table with document/version, original pages, objective, generated card IDs, and unresolved targets. After each batch, reconcile it with the table of contents or assigned reading list. Ask the model to report which supplied targets have no cards, but compare that report with your own list. Covers and reference pages can be intentionally excluded; record that decision rather than demanding cards for every page.

When two files disagree, preserve both passages and identify their sources. A newer lecture may clarify an older packet, or the difference may be unresolved. Do not let the model silently choose one. Use the course owner's clarification to determine which answer belongs in the study deck. Until then, keep the disputed card out of active review.

When you revise a source or change an answer, keep a versioned master file and a short change record. Retain the card ID for a wording correction to the same learning target. Give split cards distinct IDs. After a factual correction, reread the affected page and review that card deliberately, since you may already have practised the earlier answer.

9. Export the cards to Anki or Quizlet

Keep a master file with the question, answer, and source. Add IDs and tags if you need to update or organize a larger deck. The sample CSV, Anki Basic tab-separated text, and Quizlet import text contain the same 15 cards.

For your own checked cards, ask the model to convert the approved table into two-column import text using the prompt below. Copy the output into a plain text editor and save it as UTF-8 text. Each line should contain one question, one tab, and its answer with the source attached. Check that the conversion has kept every answer and page reference.

Convert only these approved cards into plain tab-separated text.
Column 1: Front.
Column 2: Back, followed by "Source: [document and page]".
Use one actual tab between the two fields and one line per card.
No heading row, Markdown table, or commentary.
Keep the approved wording, numbers, formulas, and citations unchanged.
Replace internal tabs or line breaks inside a field with spaces.
[Paste the checked cards here.]

Anki: a simple front-and-back import

Anki accepts UTF-8 text with comma, semicolon, or tab separators and lets you map columns to note fields. For the supplied Basic file, use a Basic note type and map column 1 to Front and column 2 to Back. It has two tab-separated fields and no ordinary header row. Source information is already included on the back.

Preview the separator and mapping before importing. Leave HTML handling off for this plain-text file. If you add <br> or other HTML formatting, enable HTML handling deliberately. Anki uses the first field for duplicate matching during text imports; review its duplicate or update setting, especially when importing into an existing collection.

Anki: retain IDs for repeated updates

If you plan to update the deck repeatedly, the ID-based tab-separated text uses ID, Front, Back, Source, Tags as its column order. Create a custom note type with the fields ID, Front, Back, and Source; map the last column to Anki's tags. Keep ID as the first field for matching updates. Put your own IDs in that field, not in Anki's internal GUID field. For a one-off deck, the Basic file is simpler.

Use the following field templates so the ID does not replace the study question. The back shows the answer and its source.

Front template:
{{Front}}

Back template:
{{FrontSide}}
<hr id="answer">
{{Back}}
<div>{{Source}}</div>

A CSV row is an Anki note; the note type determines which cards it produces. Keep a Basic pilot separate from Cloze or Image Occlusion notes. A literal cloze string imported as a Basic question will not create the intended hidden-answer card.

Quizlet: import two columns on the website

Quizlet's import instructions use the website: select Create, then Flashcard set, and choose Import. Paste the supplied text, use tabs between terms and definitions, and newlines between cards. Each row becomes one card. The definition includes its source because this file has only two columns.

Check the preview and languages before saving. Quizlet's instructions describe Create as saving and publishing the set, so check the set's visibility and your permission to share the material before completing that step. Keep tabs and embedded newlines out of individual fields so they do not split a card unexpectedly.

Check a pilot before importing the whole deck

Import a small selection into a temporary deck or set first. Include a definition, a calculation, and the longest answer so you can inspect different content. Check the question with the answer hidden, reveal the back, and confirm that the source is visible. Look at λ, μ, ρ, decimal values, and minus signs on the device you will study with.

For image cards, include the image files and check their display. Anki's text-import manual explains placing referenced media in collection.media and enabling HTML for image references. Do not assume that a CSV contains the images themselves. For equations, start with readable plain text; introduce richer formatting only after a pilot renders it correctly.

10. Fix common failures at the point they occur

11. Use the deck and revise it from actual mistakes

Hide the back and attempt an answer before revealing it. For a calculation, write the formula and units as well as the number. For an explanation, compare the meaning and required condition, rather than demanding the exact wording of an AI-generated sentence.

When you miss a card, distinguish a forgotten fact from a confusing question. If several answers would be reasonable, narrow the front. If you know the term but cannot explain its role, return to the passage or add a separate explanation task. If you cannot understand the answer after reading it, study the source before adding more cards.

Set the number of new cards to a load you can review alongside your coursework. If unfinished reviews accumulate, reduce new additions and remove redundant cards. Continue practising full problems, diagram interpretation, or written arguments when those are the assessed tasks.

Keep the source PDF, master deck, and import file together. Record the document edition and the date you changed an answer. Before sharing a deck, check its visibility, source permissions, and unresolved cards; a model's passing review cannot make those decisions for you.