Create slides from a PDF without losing the citations
To create slides from a PDF and keep the citations, work from a claim list rather than from the PDF's headings: read the document, write down each claim with its page or section, order the claims, and put one claim per slide with the location in the footer. If an AI tool does the reading, require that every slide names the page it came from, and check a sample against the PDF.
Most "PDF to slides" tools skip the location step. The deck they produce looks finished and cannot answer "where does this come from?", which is the first question in any classroom or review. This guide is the version that keeps that answer.
What goes wrong when a PDF is summarised into slides
A PDF's structure is the author's, not the presenter's. Converting section headings into slide titles produces a deck that walks through the paper in the author's order, with the author's emphasis, which is rarely the order a room needs. And a summary loses the attachment between a statement and its evidence: "accuracy improves by 12%" on a slide is a claim with no page, no table, no condition.
Both problems have the same fix: extract claims with locations, then build the deck from the claims.
Step 1: decide what the deck is for
The same paper produces different decks for a lecture (teach the method), a journal club (critique the evidence) and a project meeting (decide whether to adopt it). Write the purpose in one sentence before opening the PDF. It decides which claims matter.
Step 2: read with a claim list open
Go through the PDF once and write claims as you go, each with a location:
The method needs no labelled data. §2, p.3
Accuracy improves 12% over the baseline on dataset A. Table 2, p.6
The improvement disappears on dataset B. §5.2, p.8
Training took 40 GPU-hours. §4.1, p.5
Locations are page numbers for a paper, section numbers for a chapter, and figure or table numbers whenever the claim rests on one. Ten to twenty claims cover most papers; a textbook chapter may give more, and you will select.
Step 3: order and group the claims for the room
Reorder the list for the audience, not the author. A lecture usually goes problem, idea, evidence, limits. A journal club goes claim, evidence, doubt. Group into three to five sections and title each with a claim.
Step 4: one claim per slide, location in the footer
Each slide: the claim as the title, the minimum evidence (a redrawn figure, one row of the table, one number with its condition), and the location in the footer. In Marp Markdown:
---
# Accuracy improves 12% over the baseline, on dataset A only
- Table 2: 0.81 vs 0.69 (dataset A)
- §5.2: no improvement on dataset B
<!-- Source: paper.pdf, Table 2 p.6 and §5.2 p.8. Ask: what differs between A and B? -->
If the deck footer names the paper once, the per-slide footer can be as short as Table 2, p.6. Full references go on the last slide.
Step 5: check a sample against the PDF
Pick five slides at random and open the PDF at the cited location. Is the claim what the page says? Are the conditions on the slide? This takes ten minutes and catches the rounded number and the dropped caveat, whether a person or a tool wrote the deck.
Doing it with an AI assistant
An assistant can do steps 2 and 4 quickly if it is given the actual PDF and told what to produce. What to require, and what to check:
- Input is the document, not a description of it. A tool that generates slides from a prompt ("make slides about paper X") is not reading the paper.
- Every claim carries a location. Ask for the claim list first, with page or section for each, before any slide is written. If locations are missing or vague ("throughout"), the extraction is not usable.
- Sources stay attached to the output. The deck should name the PDF, and the PDF should be next to the deck, so the citation can be opened.
- Nothing outside the sources. Instruct the assistant to write only from the attached documents; general knowledge added silently is how invented claims appear.
- Check the sample. Step 5 applies to the assistant's deck exactly as it applies to your own.
This is the workflow Lecture Studio is built around: the PDF goes into the unit's sources folder, the assistant reads it and proposes an outline with the location of each point, and writes a Marp deck that cites those locations; the deck and the PDF stay together in your folder. It runs on macOS and needs an Oberik project key. The general method, without any tool, is in how to make lecture slides from your readings.
What about scanned PDFs, slides-as-PDF and long reports?
- Scanned PDFs need OCR before any tool can read them; check a page of the text layer before trusting it.
- A PDF that is already a slide deck (someone's lecture) is a poor source for citations because it has none itself. Go to the primary text it was made from if you can.
- Long reports are best split: extract claims chapter by chapter, and let the purpose from step 1 discard whole chapters.
FAQ
Can AI create slides from a PDF?
Yes, and several tools do it in seconds. The question to ask is whether each slide names the page or section it came from and whether you can open that location; most tools do not, and a deck without locations cannot be checked.
How do I cite a PDF on a slide?
Name the document once in the deck footer or on the first slide, then put the location (page, section, table or figure) in the footer of each slide that uses it. Full references go on the last slide.
How many slides should one paper become?
Count the claims that serve your purpose, usually ten to twenty, and give each one slide. A journal club deck for one paper is often twelve slides; a lecture that uses the paper as one of several sources may use five.
Does Lecture Studio read PDFs?
Yes. PDFs, Markdown, Word and Excel files in the sources folders are read and previewed; the assistant cites them by location in the decks and notes it writes.