I keep a scrap of questions that actually returned the clause, and a scrap of questions that returned a lecture. Same contract. Same tool. The difference was the sentence I typed.
Questions that hit, and questions that wandered
| What I typed | What I got |
|---|---|
| What is the termination notice period, and who must receive it? | Thirty days, written notice to the address in section 14. I found it. It matched. |
| Is there a cap on liability? | A paragraph quoting the cap and the carve-outs. One carve-out was paraphrased loosely. I fixed the paraphrase from the page. |
| What should I worry about? | Four themes from the recitals. None of them were the indemnity, which is what I actually worry about. |
| Summarize this contract. | A summary. Use the summarizer if that is what you want. Chat is worse at “everything.” |
Chat with PDF is instructed to answer from the supplied document and to say when the document does not contain the answer. That last behavior is the feature I pay attention to. A refusal is a result. It means I should search the PDF myself or admit the topic is not in this extract. A tool that always answers will answer from the recitals when the operative clause is missing, and it will sound the same.
The extract is the contract, as far as the model is concerned
I had split the signature pages off to send them separately and then asked about the notice address. The model said the document did not specify an address. The address was on the signature page I had not included. My question was fine. My file was incomplete. I put the signature page back in the working copy and asked again. Now it cited the address. I opened the page and checked the suite number. One digit differs from the letterhead. The PDF was right. The letterhead was old. I would have “corrected” the PDF if I had trusted the letterhead. Don’t.
Scanned contracts need OCR first. Chat on an image-only PDF is chat on nothing. You will get a polite emptiness or, worse, a guess if some header text extracted and the body did not. I look at whether I can highlight a clause. If I cannot, I am not ready to ask questions.
I ask for the words, then I ask a second question only after I have checked the first answer on the page. Stacking five questions in one prompt produces a blended memo I cannot map back to lines. One question. One check. The next question can use a sentence of context I type myself: “Using only the limitation of liability section, is consequential damage excluded?” Naming the section stops it from wandering into the warranty.
Things I do not ask it to be
I do not ask whether the contract is enforceable, whether I should sign, or what a court would do. Those are not in the document. A careful answer says so. A sloppy prompt invites a tiny legal essay that cites nothing you can point to. I am not your counsel and the model is not mine. The useful output is a pointer to a sentence. The sentence still needs a human who is allowed to interpret it.
Math in tables: I ask it to quote the cell, and I add the cells myself. I have watched a total that was a sum of a subset get described as a sum of the whole. The arithmetic took ten seconds. The argument it would have caused would have taken a week.
Page numbers in the answer are only as good as the text. If you need “this is on page 12,” look at page 12. I type the page into my notes after I see it. I do not let the chat message become the pin cite in a letter to the other side.
The request goes to the Worker. A contract with a confidentiality clause that forbids third-party processors does not become acceptable because the question was narrow. Narrow helps accuracy. It does not rewrite the clause. If I cannot send it, I read it.
When the answer is “the document does not say,” I search the PDF for two synonyms before I believe the gap. “Terminate” might be “cancel.” “Notice” might be “notify.” If both searches miss, the gap is probably real and I write “not found in pages 1–22” in the note, including the range, so nobody thinks I searched the schedules I had split off.
I do not paste the model’s paragraph into the redline. I paste the contract’s paragraph, or I describe the issue in my own words with the section number. The model’s paragraph is a map. Putting the map in the document is how a paraphrase becomes the draft. I have seen that happen in a shared doc named final. The other side quoted our paraphrase back to us as if we had proposed it. We had, accidentally. The fix was to revert to the sentence that was actually on the page.
Schedules, and the question that only works if you include them
The main body said pricing was “as set out in Schedule B.” I asked for the unit price. The model, correctly, said the body did not state a unit price. I had not attached Schedule B. That felt like a failure and it was a success. A worse tool would have invented a price from a round number in the recitals. I added the schedule, asked again, and got $14.20 per unit with a volume break at 500. I checked the table. The break was at 5,000. The model had dropped a zero in the prose while the digits in its quotation were right. I used the quotation, not the sentence around it. Then I stopped trusting sentences around quotations.
Definitions sections poison casual questions. “What does delivery mean?” returns the definition, which may exclude collection by the buyer, which the commercial team calls delivery anyway. I ask “quote the definition of Delivery” and I read the exclusions myself. Asking “does delivery include pickup?” gets a yes or no that compresses three lines into a coin flip. I want the three lines.
I keep a running note of section numbers I have personally opened. If it is not on the note, I have not verified it, even if chat sounded sure. The note from the last review was twelve lines. The contract was forty pages. That ratio is fine. The other twenty-eight pages were boilerplate I was not asked about. Pretending to have chatted my way through all forty is how boilerplate surprises you at signature. If the question is “any unusual indemnity,” I still open the indemnity heading myself. Chat can point. Unusual is a judgment, and the judgment is the part I am paid for, or the part counsel is paid for. The pointer just saves me from reading the recital twice.
Cross-references are the other trap. “Subject to section 9” may be a casual pointer or the clause that deletes the warranty you just quoted. I open section 9 before I repeat the warranty to a client. The chat answer about the warranty was not wrong. It was partial. Partial, forwarded, becomes the whole story. I type “see also §9, read” next to any answer that contains “subject to.”