The 70-page funding call, the rental contract with the small-print clauses, the minutes of the residents’ meeting: these are exactly the documents an AI assistant seems made for. You upload them, type “summarise”, and ten seconds later you have a clear, tidy page.
The trouble is that this clear, tidy page can be wrong in ways that are hard to spot. The model may have read only part of the file, misunderstood a table or “rounded off” a deadline. Not because it’s stupid, but because of how these tools work. Here’s the method I use, with prompts you can copy.
Why a plain summary is risky
- The document may not be read in full. Every model has a limit on how much text it can consider at once (the context window). Some services read only parts of very long files.
- Scanned PDFs are images. If the document is a scanned photocopy without text recognition, the AI has to “read” it like a photo, with errors on numbers and amounts.
- Tables and multi-column layouts break when the text is extracted: a figure can end up in the wrong row.
- The middle gets neglected. It’s a known behaviour: models tend to give more weight to the beginning and end of a long text.
- Summaries smooth over exceptions. And the exceptions (“except where…”, “no later than…”) are often exactly what you need to know.
Step 1: check the PDF contains text
Open it and try searching for a word with Ctrl + F (on a Mac, Cmd + F), or selecting a line with the mouse. If nothing is found or nothing can be selected, it’s a scan. In that case run text recognition (OCR) first: Adobe Acrobat does it with Scan & OCR, the Mac’s Preview recognises text in images, and Google Drive converts a PDF into a text document if you open it with Google Docs.
Step 2: ask for the table of contents first, not the summary
This is the most useful check and the least used. Ask:
Before summarising, list every section and chapter heading in the document with the page where it starts. Also tell me how many pages you read in total.
Compare the answer with the PDF. If the last sections are missing or the page count doesn’t match, the model didn’t read everything: you’ll need to go in chunks (step 4).
Step 3: ask for a summary with page numbers
Summarise the document section by section. For every important piece of information, give the page it comes from in brackets. Quote dates, amounts and percentages exactly as written, without rounding. If a point is ambiguous, say so.
Page references do two things: they let you verify quickly, and they “force” the model to anchor itself to the text rather than to its general idea of what similar documents look like.
Step 4: for very long documents, work in chunks
If step 2 showed the file is too long, split it. Many PDF readers let you extract a page range (in Chrome, Print > Save as PDF with the pages you choose is enough). Upload one part at a time, ask for a summary of each and only at the end ask for an overall synthesis of the partial summaries.
Step 5: ask targeted questions with mandatory quotes
The summary gives you the big picture. What really matters should be asked one item at a time:
What are all the deadlines in the document? For each one, quote the exact sentence in quotation marks and give the page. If there are no deadlines, write "no deadlines found" rather than inferring them.
Other questions that work well, depending on the document:
List the clauses that involve penalties, automatic renewals or costs for me, with a verbatim quote.What requirements must I meet to apply? Separate mandatory ones from those that earn points.What was decided and voted on? For each item, give the outcome and the votes, quoting the minutes.
Explicitly asking for “not found” is essential: without it, many models prefer giving you a plausible answer to giving none.
Step 6: spot-check
Pick three claims from the summary, preferably ones with numbers or dates, and check them on the pages cited. If all three are right, you can reasonably trust the rest. If even one is wrong or the quote doesn’t exist, check everything: it means the model was working “from memory”.
A note on confidentiality
Before uploading a contract or a medical report to an online service, ask yourself whether that data can leave your home: names, ID numbers, bank details and health data deserve caution. You can redact them before uploading, use a business plan that excludes your data from training, or try a local model, which often does well at summaries.
Following these steps, a summary costs you five minutes instead of ten seconds. But they’re five minutes well spent when the document is one where a wrong deadline is expensive.