How to summarize a long contract so you can check it page by page

AI assistants summarize long documents quickly, but they are less reliable about where each point came from, and a contract summary you can’t trace to the page is of little use. This method asks for a table instead: every party, obligation and date with its clause, its page and the exact words, so you can open the document and confirm each row. The steps are the same in Claude, ChatGPT, Gemini or Grok, and they work for any long document.

What you’ll be able to do

• Get a table of the parties, obligations, deadlines and payments in a long document, with the clause and page for each one, and confirm every row against the original.
• Catch the usual errors (wrong or invented page numbers, missed clauses, pages the assistant couldn’t read) before they reach your notes.

What you need

• A chat assistant. Since the method uses only the chat, a free account that lets you attach a PDF or paste text is enough.
• A public document in PDF, 20 pages or longer (see the next section).
• A PDF reader with a search box, open next to the chat.

Why practice on a public document

Anything you attach or paste goes to the provider and is handled under its terms and your account settings. Before a client’s contract goes anywhere near a chat assistant, you need to know what happens to it and whether your duty of confidentiality allows it. Our article Before you upload a client document lists the questions to ask. Until you have the answer, practice on documents anyone can read, where a mistake costs nothing:

• a contract a government agency has published on its transparency portal, or one a company listed in the United States has filed with the SEC (free on EDGAR, as Exhibit 10 to its reports);
• the terms of service of an online service you already use.

Steps

1. Open the PDF in your reader and try to select a sentence, or search for a word you can see. If neither works, the file is probably a scanned image: read the section Scanned PDFs below first.
2. Check whether page 1 of the file matches the number printed on it. A cover or a table of contents can throw this off. You will ask for the file’s page number, the one your reader shows.
3. Start a new conversation and attach the PDF. If you can’t, or the assistant says the file is too long, paste the text instead, in parts if needed. Put a line such as --- Page 7 --- before each page, so the assistant knows where every page starts. Then make two changes to prompt 1: replace “The attached PDF” with “The text pasted above”, and replace the PDF page rule with “PDF page: use the page markers in the pasted text.”
4. Paste prompt 1 (below) and send it.
5. Check the table row by row. Search your reader for five or six words from each quote (a longer search can fail on the PDF’s line breaks) and see whether they land on the page and clause given. Mark each row as confirmed, wrong page or not found. Check every row, including the ones that look obvious.
6. Paste prompt 2 to look for what the first pass left out, and check the new rows the same way.
7. Compare the list of headings from prompt 2 with the document’s own. A heading with no row in the table deserves a read.
8. Keep only the version you checked and corrected by hand. If the assistant fixes a page number, check the fix too.

Two prompts to copy

Prompt 1, the table:

The attached PDF is a contract. Read all of it before answering.

Make a table with these columns:
Item | Type | Clause | PDF page | Exact quote

- Type: party, obligation, date or deadline, payment, termination, or other.
- PDF page: count the first page of the file as page 1, whatever number is printed on the page.
- Exact quote: copy the words as they appear, up to 25 words. Do not paraphrase in this column.
- If you are not sure of a page, write "unsure" instead of guessing.

Before the table, tell me whether any pages could not be read (for example, scanned pages) and which ones.

Prompt 2, what’s missing:

Go through the document again and look only for what the table is missing, in particular:
- obligations in annexes or schedules
- definitions that change an obligation
- clauses that override or qualify others ("notwithstanding", "subject to")
- renewals, notice periods, penalties and conditions

Add each one as a new row in the same format, marked "new".
Then list every clause heading in the document with its PDF page.

For other kinds of document, change the first line and the types.

What can go wrong and how to check

Wrong or invented page numbers

An assistant can give a page number with complete confidence and still be wrong: thrown off by a cover, taken from the number printed on the page instead of the file, or simply made up. That is why the prompt asks for an exact quote. If your search finds the quote, you have the right place, whatever the table says. If it finds nothing, treat the row as unconfirmed, because the quote itself may have been reworded. Clause numbers come from the text and tend to hold up better than page numbers. Use them as a second reference and to choose between two matches when the same words repeat.

Missing clauses

A table looks complete even when it isn’t. Annexes, definitions and the last few pages are easy to skip; prompt 2 and step 7 are there to catch them.

Scanned PDFs

If you can’t select the text, each page is a picture. Whether an assistant can read those pictures depends on the tool, and it may answer from whatever it managed to read without mentioning the rest. Misread figures do the most harm: dates, amounts, percentages. If your PDF software has text recognition (OCR), run it first; otherwise, practice on another document. OCR makes mistakes too: if you work from a scan, check every date and amount against the page.

Long documents read only in part

An assistant may read only part of a long file and summarize that part without telling you. Watch for a table that thins out toward the end. Then ask for the document’s last clause, word for word, with its page, and compare the answer with the original. If it doesn’t match or comes back vague, paste the rest in parts, with a marker before each page as in step 3.

If you practice law in Spanish, our book Inteligencia Artificial Para Abogados is for you. It is in preparation, and our Books section will say when it is out.