Prompt hygiene for document questions means writing each question so the AI can only answer from the file in front of it, shows the exact passage it relied on, and says plainly when the document is silent. Vague questions produce confident-sounding summaries that blend the document with the model's general knowledge. Tight questions produce answers you can verify in about thirty seconds. HiDocument, which analyzes PDFs, Word documents and images and lets you chat with them, works best when you ask in the second style.
Prompt hygiene is the discipline of phrasing questions about a document so that every answer is bounded, quotable, and checkable. If you skip it, the cost is quiet: a missed termination notice period, a misread cap on liability, or a date that never appeared in the file. The seven rules below are the ones that matter most in contract and compliance work.
Why does the way you ask change the answer?
A language model predicts plausible text, and a loosely worded question gives it room to fill gaps with plausible-sounding material. When you ask Is this contract fair?, there is nothing in the document to anchor the answer, so the reply leans on general patterns about contracts.
When you ask What does section 9 say about termination for convenience, and who may invoke it?, the model has a place to look and you have a place to check. The model has not become smarter. The question has simply removed the room for improvisation.
- Open questions invite opinion, which is not verifiable against the file.
- Bounded questions point to a clause, a party, or a defined term.
- Quote-demanding questions make a wrong answer visible immediately.
For the underlying failure mode, see our explanation of how AI can hallucinate dates and facts in document analysis.
What are the seven rules of prompt hygiene?
The seven rules are: scope the question, demand a quote, name the defined term, separate extraction from judgment, pin down dates, allow the answer to be "not stated," and verify before you rely. Each one closes a specific gap where errors enter.
- Scope the question. Name the section, party, or topic. Ask about one thing per prompt.
- Demand the quote. Ask for the exact wording and its location, such as section number or page.
- Use the document's own defined terms. If the contract says "Services" or "Confidential Information," use those words.
- Separate extraction from judgment. First ask what the text says. Then, as a second step, ask what it means for you.
- Pin down dates. Ask for the date as written, and ask separately how it was calculated if it is derived.
- Allow "not stated." Tell the model that a missing clause is a valid answer.
- Verify before you rely. Open the cited passage and read it yourself.
How do you scope a question to the document?
You scope a question by telling the model to answer only from the uploaded file and by naming exactly what you want located. A good scope sentence is short and repeatable, such as: Using only this document, find the clause that governs payment timing and quote it.
Scoping matters most for long files, because a model has more places to drift. A 300-page agreement behaves differently from a 3-page one, as we cover in our piece on why long PDFs analyze differently from short ones. For long files, ask about one section at a time and name the section.
Also tell the model what to do when it cannot find the answer. Without that instruction, many models will produce something rather than say nothing.
What does a good prompt look like next to a bad one?
A good prompt names the target, limits the source, and requests evidence. A bad prompt asks for a verdict. The table below shows common document questions rewritten for hygiene.
| Weak prompt | Why it fails | Hygienic version |
|---|---|---|
| Is this NDA safe to sign? | Asks for a verdict with no anchor | Using only this document, list the term, the exclusions from confidentiality, and the governing law. Quote each with its section. |
| When does this contract end? | Mixes stated and calculated dates | Quote the sentence that sets the term. If an end date is not written, say "not stated" and give the start date and duration separately. |
| What are the risks? | Unbounded, invites generic risks | Find any clause that lets the other party change price or terms without my consent. Quote it, or say none was found. |
| Summarize the liability section. | Summaries can drop caps and carve-outs | Quote the liability cap, any exclusions from the cap, and any carve-outs. Do not paraphrase numbers. |
| Does this comply with our policy? | The policy is not in the prompt | Here is our rule: [paste rule]. Does clause 14 meet it? Quote the clause and explain in one sentence. |
How should you handle dates, numbers, and defined terms?
Treat every date, dollar amount, and percentage as a value to be quoted, not paraphrased. A summary that says "about 30 days" is not the same as a clause that says "within thirty (30) business days of receipt." The difference between calendar and business days can move a deadline by a week.
Defined terms deserve the same care. Contracts give capitalized words precise meanings that can differ from everyday usage. If you ask about "confidential information" in lowercase, the model may answer with the everyday meaning rather than the one the contract defines.
- Ask for the definition section first, then ask questions using those terms.
- Request numbers exactly as written, including the words and the figures.
- When a date depends on a trigger (signature, delivery, notice), ask for the trigger and the period separately.
- Do the arithmetic yourself, or ask the model to show each step so you can check it.
Our guide on what AI document verification can and cannot confirm goes deeper on why a model can read a date but cannot confirm it is correct in the real world.
Where does manual checking become too slow?
Manual checking gets slow when you have to run the same ten questions across a stack of files. Typing the same hygienic prompts into each document, copying the quotes into a sheet, and opening each cited passage can take longer than the reading it was meant to save.
This is where a tool earns its place. HiDocument lets you upload a PDF, Word document or image, chat with it, and extract structured details so that the same questions run consistently. The bulk mode handles several files at once, and the free tier includes 10 analyses per month with a 5 MB file cap, which is enough to test your prompts on real documents before committing to anything.
The tool does not replace your verification step. It makes the setup faster and the quotes easier to locate.
When should a person stay in the loop?
A person should stay in the loop whenever an answer will drive a signature, a payment, a filing, or a deadline. The model can find and quote text. It cannot know your negotiating position, your jurisdiction's quirks, or whether the document you uploaded is the final version.
The NIST AI Risk Management Framework, published at nist.gov, treats human oversight as a core control for AI systems, and the same logic applies at the scale of one contract review. HiDocument is a tool, not counsel, so anything with legal consequences should be checked by a qualified professional.
A simple verification routine works well:
- Open the page or section the model cited.
- Confirm the quote matches the source word for word.
- Check that the quote actually supports the claim, not just that it exists.
- Look one paragraph above and below for exceptions that change the meaning.
- Record the clause reference next to your decision.
What if you cannot afford the extra step?
The most common objection is that quoting and verifying doubles the work. In practice it moves the effort to a place where it is cheap: confirming a quote takes seconds, while finding a clause from scratch takes minutes. You are checking a pointer rather than reading the whole file.
The other objection is that a careful prompt is tedious to write. Save your best prompts as a short template once, and reuse them for every contract of the same type. If you want a baseline first, our post on preparing documents for AI analysis covers file quality, which affects every answer.
People also ask
Should I paste the whole contract into the prompt?
Not usually. Upload the file so the tool reads it directly, and ask about specific sections. Pasting excerpts is useful only when you want to add outside material, such as your own policy, for comparison.
Does asking for a quote guarantee accuracy?
No. It makes errors easier to catch, which is different. You still need to confirm the quote against the source and read the surrounding text.
Is a confident answer more likely to be right?
No. Tone reflects how the text was generated, not how well it matches the document. Check the citation, not the confidence.
What is the next step?
Pick one contract you have already reviewed by hand, and run three hygienic prompts against it: one for the term, one for payment timing, and one for liability. Compare the quoted passages with what you found yourself. You can create a free HiDocument account, upload the file, and see within a few minutes whether the quotes line up with your own reading. If you later need larger files or batches, the pricing page lists the Pro limits.
Frequently Asked Questions
What is prompt hygiene in document analysis?
Prompt hygiene is the practice of writing questions that limit an AI to the uploaded document, ask for exact quotes with locations, and allow the answer 'not stated.' It reduces guessing and makes every answer easy to check against the source text.
How do I stop an AI from making up answers about my contract?
You cannot eliminate it, but you can reduce it. Ask narrow questions, require a quote with its section number, and tell the model that a missing clause is an acceptable answer. Then open the cited passage and confirm it says what the model claims.
Should I ask an AI for a summary or for exact quotes?
Ask for exact quotes first, then a summary if you need one. Summaries can drop caps, carve-outs, and conditions. Quotes keep numbers and wording intact, so you can see precisely what the document says before you interpret it.
Why do defined terms matter when I question a document?
Contracts give capitalized terms specific meanings that can differ from everyday English. If you use the document's own defined terms in your question, the model is more likely to answer with the contract's meaning instead of a general one.
Can I rely on an AI answer without checking it?
No. Any answer that drives a signature, payment, filing, or deadline should be verified against the source passage. HiDocument is a tool, not legal counsel, so consequential decisions should also be reviewed by a qualified professional.