How to Study a PDF with AI Without Fake Quotes
A grounding skill for people who study a PDF with AI. Claude or Gemini, file uploaded. Every quote needs a page. If it is not in the file, the model must refuse. Outline, then quiz. No invented citations.
Anyone who needs to study a PDF with AI can do it in one sitting without fake quotes: upload to Claude or Gemini, demand page-plus-sentence, refuse anything not in the file, outline the document, then quiz from the outline. Spend 25 minutes on grounding before you let a model "explain" a paper it has already started inventing.
If you cannot open the page and see the sentence, the answer is fan fiction.
This is a skill with a stop, not a clever chat. If you need the pattern for reusable loops, read what a skill prompt actually is and then lock the rules below.
Grounding rules before you study a PDF with AI
Set the law in the first message, before the first question. If you ask a content question first, Claude and Gemini will answer, then retrofit a citation. Retrofit is how fake quotes get a page number.
1. Upload the file. Name it. State the law.
Upload the PDF to Claude or Gemini. In the same message:
"This file is the only source. Every claim must include a short quote and a page number from this file. If the file does not contain the answer, reply REFUSE and say what is missing. Do not use training knowledge. Do not paraphrase a quote as if it were a quote."
Name the file in the instruction so the model cannot wander into "the literature." There is no literature in this chat. There is one PDF.
2. Quote with a page, or it does not count
A valid answer looks like this:
- Claim in one sentence.
- Quote in quotation marks, under 25 words.
- Page number.
- One-line explanation that does not add facts.
Your instruction: "If you cannot quote, you cannot claim. Paraphrase is allowed only after a quote, never instead of one."
Then spot-check the first answer. Open the page. If the quote is off by more than a function word, the whole session is contaminated. Restart the chat. Do not "correct as you go." Corrections teach the model that you negotiate with citations.
3. Refuse if it is not in the file
REFUSE is a success state. Treat it like a green check.
Ask a question you know is outside the PDF on purpose, once, as a drill. You want to see REFUSE. If Claude or Gemini answers anyway, your law did not land. Restate it. Do not proceed to outline.
4. Outline the file, still quoted
Only after two grounded answers and one successful REFUSE do you ask for an outline.
"Outline this PDF. Each heading must include one quote and a page. Maximum eight headings. No heading for a topic that lacks a quote."
Verify two headings at random, not the first one. Models front-load honesty and get sloppy in the middle. If either heading fails, you do not quiz. You restart the outline with the failed heading pasted back as a counterexample.
5. Quiz from the outline, not from the void
Now: "Quiz me with five questions. Each question must be answerable by one of the outline quotes. After I answer, you quote the page again. If I ask something off-file, REFUSE."
When the five questions are done, stop. Do not ask "what else should I know." That prompt is a fake-quote generator with a self-help smile.
Save the grounding law as a skill in the Promptcrates tools for study loops so the next PDF does not start from vibes.
Privacy fork: do not upload what you cannot leak
Claude and Gemini uploads leave your machine. If the PDF is personal, client, or unpublished, do not use this hosted loop.
Run the same quote-and-refuse rules on a local model instead. The Qwen 3.8 27B local coding-agent note is the privacy pattern: keep the file on your box, still demand pages, still treat REFUSE as success. Local does not mean grounded. You still write the law.
Reusable skill card
Trigger: You dropped a PDF on Claude or Gemini and you are about to ask "what is this saying."
Input: The uploaded file, the grounding law (quote plus page, REFUSE if missing), and one study goal in a single sentence.
Output: Two grounded answers, one deliberate REFUSE, an eight-heading outline with quotes, then a five-question quiz tied to those quotes.
Stop: After the quiz. No "what else should I know." No training-data definitions. No second PDF in the same chat.
Takeaways
- Upload to Claude or Gemini, then state the law before any content question.
- No quote, no page, no claim. Paraphrase never replaces a quote.
- REFUSE is success. Drill it once on purpose.
- Outline with quotes, then quiz from the outline. Stop there.
- Sensitive PDFs stay local. Hosted models are not a vault.
Frequently Asked Questions
Can I paste a few pages instead of uploading the file?
Only if you also paste page numbers in the text. Without pages, you cannot enforce the quote rule. Upload is cleaner.
What if Claude or Gemini cites a page that looks real?
Open that page. If the sentence is not there, mark it FAIL and demand a refuse. A plausible page number is still a fake quote.
Is ChatGPT allowed in this loop?
Only if the PDF is in the same chat as the questions and you still require page-plus-quote. This skill is written for Claude and Gemini uploads because the file is then the ground, not a memory.
Why outline before quiz?
A quiz on a fake outline trains you on fan fiction. The outline has to survive a page check first.
What about a sensitive PDF I should not upload?
Do not upload it to Claude or Gemini. Use a local model. See the Qwen local coding-agent note for the privacy pattern, then apply the same quote-and-refuse rules on your machine.