What it is
A limit names something to keep out of the answer: a topic, a kind of claim, a source, a promise, a name. On the ten checks this is Limits, the sixth check. Typical ones at work are “leave out customer names”, “no health claims”, “use only the attached policy” and “do not mention the reorganization, which has not been announced”.
The reason is the part that does the work. A model given a bare rule applies it to the exact words you used. A model told why can apply it to cases you did not list, because it knows what you are protecting against.
When it helps
It helps most with the mistakes you have already seen. If replies to customers keep offering a discount, if summaries keep adding industry background nobody asked for, if every product description mentions a competitor, name that thing in the prompt and say why it is unwanted.
It helps wherever an unwanted line would cause real trouble: a promise the company cannot keep, a claim the legal team has not cleared, a confidential project in a document that will be forwarded, a figure from last year’s report in this year’s. And it helps keep scope. “Leave out pricing, which finance is reviewing separately” stops the answer from duplicating someone else’s work.
When it does not
For format and style, a ban is weaker than a description. “Write in plain paragraphs” works better than “no bullet points”, because it tells the model what to produce instead of only what to avoid. Save limits for content, and ask for the format you want.
A long list of don’ts is hard to follow and hard to check. Ten rules compete with each other and with the task. Keep the few that matter, and state each one once: a limit repeated in three places can start to outweigh everything else in the prompt.
A limit without a reason can be applied too narrowly or too widely. “Do not mention price” might stop the model from saying the product is affordable, or it might not. If the boundary matters, give the reason and an example of what crosses it.
And a limit does not check itself. The model can break it in a way that reads fine, so read the answer for the thing you banned before it goes anywhere, especially in the first few runs of a prompt you plan to reuse.
Before and after
Three requests where the after version adds limits with their reasons. Each score comes from the Prompt Grader’s free ten-check estimate, written for any chat model.
Write a reply to the customer below about their delayed order. """ [the customer's email] """
Write a reply to the customer below about their delayed order. Do not promise a delivery date: the carrier has not given us one, and a missed promise would make this worse. Leave out the reason for the delay, which our legal team is still reviewing. """ [the customer's email] """
A model asked to calm a customer may reassure them with a date. The second version gets a reply that apologizes and says what happens next without a date, and without an explanation that legal has not cleared.
Write a product description for our new standing desk for the website. Specs: [the spec sheet]
Write a product description for our new standing desk for the website. Avoid health claims, such as better posture or less back pain: we have no studies to back them, and the product page is reviewed for advertising rules. Do not mention other brands by name. Specs: [the spec sheet]
Standing desks are often sold on health benefits, so that is where an unguided description can go. With the limit and its reason, the description sticks to what the spec sheet supports, such as height range and load, and the reason covers health claims the prompt did not list.
Summarize what our customers said in the survey results below for the board. """ [the survey results] """
Summarize what our customers said in the survey results below for the board. Use only the survey results. Leave out industry trends or anything you know from elsewhere, because the board asked what our own customers think. """ [the survey results] """
Summaries of survey data can drift into general claims about the market that sound informed and come from nowhere in the data. The limit keeps every sentence traceable to a response, which is what lets the board trust it.
Each after version gains two points: one for the limit and one for Fit, because the limit sits in its own paragraph and the prompt now has separate parts.
The line to add
Leave out {{what_to_leave_out}}, because {{reason}}.
Write one line per limit, and keep the list to the limits that would really cost you something. If you find yourself writing a fifth, check whether some of them are really about format or tone, and describe what you want for those instead.
In ChatGPT, Claude, Gemini and Copilot
Google’s guide for Gemini treats constraints as part of clear instructions, and says you can tell the model what to do and what not to do. Anthropic’s guide for Claude makes the case for the reason. Its example replaces a bare “never use ellipses” with an explanation that the response will be read aloud by a text-to-speech engine, and notes that Claude can generalize from the explanation. The same guide recommends saying what to do instead of what not to do when the subject is formatting.
OpenAI’s guide to its latest models, the ones behind ChatGPT, advises keeping a set of rules in one place and stating each rule once, and warns that repeating an instruction such as “ask first” can make the model ask when it does not need to. Microsoft’s four parts of a Copilot prompt have no separate slot for limits; they belong with expectations, how Copilot should respond.
Key takeaways
- Name what to leave out and why. The reason lets the model cover cases you did not list.
- Start with the mistakes you have already seen in past answers.
- Use limits for content. For format and style, describe what you want instead.
- Keep the list short and state each limit once.
- Read the answer for the banned thing before it goes anywhere.
Questions
Do negative instructions work with ChatGPT and Claude?
For content, yes, especially with a reason: “do not mention the price change, which has not been announced.” For format and style, describing what you want works better than banning what you do not, so ask for “plain paragraphs” rather than “no bullet points”. Keep the list short and say each limit once.
How do I stop AI from making promises to customers?
Name the promises it must not make and why: no delivery dates the carrier has not confirmed, no refunds beyond the policy, no discounts. Then give it the line it can use instead, such as “we will email you as soon as the carrier confirms a date.” Read replies before they go out until the prompt has proved itself on real cases.
Why does AI ignore my instruction not to do something?
Common reasons: the limit was buried in a long prompt, it had no reason so the model drew the line in a different place, or it was one of so many rules that some lost out. Put the limit near the task, give the reason, cut the list to the few that matter, and test it on a case that would tempt the model to break it, for example in Prompt Tester.
Prompts to use with this
- Check a draft's claims against your sourcesBefore a report, article or proposal goes out, you want every factual claim matched to a source and anything unsupported pulled out for checking.
- Organize your evidence for a small claims caseA customer has not paid or a dispute is heading to small claims court, and you need your facts, amounts and evidence in order before you file or attend.
- Group a keyword list into topic clusters and pagesYou exported a list of search keywords and need them grouped by topic, with one page planned per group, so you write fewer, better pages.