Free tool · runs in your browser
PDF Data Extraction Prompt Builder
Pick the kind of document, tick the facts you need, and copy a prompt for ChatGPT, Claude or Gemini that asks for an exact quote and a page number for every value, and says NOT_FOUND instead of guessing. You build the prompt here; your document never comes to this page.
Build your extraction prompt
1What kind of document?
2What do you need from it?
3Rules and answer format
4Your prompt
- Open your AI app and attach the document (or paste the text where the prompt says so).
- Paste the prompt in the same message and send it.
- Check the answer using the list below the tool before you rely on it.
What a good answer looks like Example with made-up data
The companies and numbers are invented. Fields with no sample row show NOT_FOUND, which is the answer you want when the document is silent.
How to check the answer
A prompt lowers the chance of a wrong value, it does not remove it. Spend two minutes on these before you use any number:
- Search three quotes. Open the document, press Ctrl+F (Cmd+F on a Mac) and paste the quote. If you cannot find it word for word, treat that row as wrong.
- Read the surrounding sentence. A correct quote can still be about something else, such as a cap that applies only to one type of loss.
- Check page numbers. The page printed on the sheet and the page number in your PDF viewer often differ. The prompt asks for the printed number, or "PDF page N".
- Re-add the totals. For invoices and reports, add the line items yourself and compare with the total.
- Open every row marked REVIEW, CONFLICT or NOT_FOUND. Those are the ones the AI was not sure about, and NOT_FOUND on a clause you expected to see is worth a look.
Why the prompt has these rules
Each rule closes a gap that causes a specific kind of error when you simply ask an AI to "pull the key terms from this PDF".
| Rule | What goes wrong without it |
|---|---|
| Use only the document | The AI fills gaps from what contracts usually say, so a missing clause looks like it exists. |
| Quote every value | Without a quote you cannot tell a real value from a plausible one. A quote can be searched for in seconds. |
| Give the page or section | Lets you jump straight to the spot instead of searching the whole file. |
| Say NOT_FOUND when missing | Gives the AI a safe answer, so it does not invent one to avoid a blank cell. |
| Show conflicts | Documents often disagree with themselves, for example a main clause and a schedule. Both values should be visible. |
| Flag doubtful items for a person | A bad scan or an unclear sentence should be marked, not silently resolved. |
| Check each quote before answering | A second look inside the same answer catches values that the quote does not support. |
What this builder cannot do
It cannot read your document, check the AI's answer, or promise a perfect result. It does not interpret clauses, rate risk or give legal, financial or medical advice; the prompt says so to the AI as well. Treat the output as a fast first pass that a person confirms. If your document is confidential, check that the AI app you use is approved for it before you attach the file.
Related free tools and prompts
Prompt Analyzer
Score any prompt on task, context, format, limits and details.
Open Analyzer →Prompt Optimizer
Turn a rough instruction into a structured prompt, side by side.
Open Optimizer →Prompt Converter
Move a prompt between ChatGPT, Claude and other styles, or into JSON.
Open Converter →More ready-made prompts: ChatGPT prompts, Claude prompts and chat and writing prompts.
Questions, answered
Will this stop the AI from making things up?
It lowers the chance and makes mistakes easier to catch, but it cannot remove them. The prompt asks for an exact quote and a page for every value and says to write NOT_FOUND instead of guessing, so you can check each row against the document. Always check the rows that matter.
Is it safe to paste a confidential contract?
This page never sees your document: you only build a prompt here, and nothing you type is sent anywhere. The risk is the AI app you paste into. Use an app your company or client has approved for confidential files, and read its data settings first.
Does it work for scanned PDFs?
It can, but scans are where mistakes are most common because the app has to read the page image. Keep the "Flag doubtful items" rule on, and check every quote and number against the page.
Which AI app should I use?
Any that lets you attach a file or paste long text: ChatGPT, Claude and Gemini all do, with different file-size and length limits that depend on your plan. The prompt is plain text and does not depend on one app.
How long can the document be?
Check your app's limit. For a long document, run the same prompt on one section at a time (for example 30 to 50 pages) and join the tables. Say which pages you attached so the page numbers stay right.
Is this legal or financial advice?
No. It helps you pull facts out of a document so a person can read them faster. It does not tell you what a clause means or whether to sign. For decisions that matter, ask a qualified lawyer, accountant or adviser.