Negative instructions, and why they half-work

“Do not mention the price”

You write it. The answer mentions the price.

This is one of the most reliably annoying behaviours in everyday prompting, and it has a mechanism worth understanding, because the fix follows from it directly.

Every word in your prompt is in the context. “Do not mention the price” puts price in front of the model, in a prompt about your product, immediately before it writes about your product. The negation is a small amount of text asking it to avoid a topic that the rest of the prompt has just made salient. Sometimes the negation wins. Sometimes it does not.

You can watch the same effect in yourself. “Do not think about the last argument you had” is not a workable instruction.

Positive beats negative, consistently

The reliable fix is to say what to do instead of what to avoid. Compare:

• Do not use jargon → Write so a 15-year-old could follow it.
• Do not be too long → Under 150 words.
• Do not make things up → Use only facts from the document below. Where it does not say, write “not stated”.
• Do not sound like AI → Short sentences. No lists. No “moreover”, no “delve”, no “it is important to note”.
• Do not include the price → Cover the specification, the warranty and the delivery time. Those three only.

Each rewrite replaces an absence with a target. A target is something the model can produce; an absence is something it has to check for after the fact, and there is no checking step.

Notice the last pair. “Cover these three only” implies the exclusion and gives positive work to do. It is the same instruction with a shape.

When a negative is the only honest form

Some constraints are negative and cannot be rewritten. Do not contact the client directly. Do not use any figure not in the table. Do not include patient names.

For these, three things help.

Be concrete rather than categorical. “Do not use marketing language” is unenforceable; the model does not have a stable definition. “Do not use the words: revolutionary, seamless, unlock, empower, game-changing” is a list it can check against.

Put it last, and separate it. A short list of prohibitions immediately before the model answers, under a heading like MUST NOT:, survives better than the same words buried in paragraph two.

Verify in a second pass. Ask separately: “Does the draft below contain any of the following? Quote each occurrence.” Checking is a much easier task than composing under prohibition, and it produces evidence rather than assurance.

The special case: “do not invent”

This is the negative instruction people rely on most, and on its own it does very little. A model does not know it is inventing. Fabrication and recall are produced by the same process, which is why “do not make things up” is closer to a wish than a constraint.

What works is giving it an alternative to inventing:

Answer only from the document below. For anything the document does not state, write “not stated”. If you are relying on a sentence, quote it before interpreting it.

Three positive instructions, one of which (the quote) is verifiable in seconds. This is much stronger than any number of prohibitions, and it is the single most useful pattern in document work.

Rejected options are context, not instruction

A related habit worth having. When you have already tried something and it did not work, say so as material rather than as a ban:

I have already tried moving the deadline (the client refused) and adding a second developer (no budget). Suggest options that do not depend on either.

That gives reasons, not just exclusions, so you do not get a slightly reworded version of the thing you rejected. It also stops the most tedious loop in AI-assisted work, where you reject an idea and get it back three turns later with a different adjective.

The summary

Say what you want. Where you must say what you do not want, name it exactly, put it at the end, and check the output against it rather than trusting that the instruction held.