Negative Prompting — When It Helps and When It Hides the Bug

Article summary

Negative prompts ("do not include emojis," "no introduction," "do not invent facts") work when the model has a strong default you need to override. They fail when they are used as a substitute for a clearer positive instruction. Rule: if you find yourself adding a fourth "do not," your positive instruction is too vague.

Negative prompting is the pattern of telling the model what NOT to do. "Do not start with 'As an AI.'" "Do not use emojis." "Do not invent dates." It sounds obvious. It works only sometimes, and overusing it is a sign your positive prompt is doing the wrong job.

When a negative is actually useful

A negative prompt does real work when:

The model has a strong default you need to suppress. Chat-tuned models start with throat-clearing phrases ("Certainly!", "Of course!"). A direct "do not include any preamble" removes them faster than trying to reword the positive instruction.
You are overriding a specific failure pattern. "Do not include the word 'crucial.'" "Do not use the em dash." These are sniper-level corrections aimed at a known tic.
The "no" sets a boundary the positive could not. "Do not answer questions outside [scope]." This is policy, not style.

When a negative is a band-aid

A negative is hiding a problem when:

It is your fourth or fifth "do not." That means your positive instruction is vague and you are spot-fixing failures one by one. Rewrite the positive.
It mentions something you do not want the model to think about. "Do not consider X" puts X in the conditioning. The model now thinks about X. Either remove the constraint or restructure the prompt so X never comes up.
It tells the model not to do something obvious. "Do not invent facts" sounds good but does not actually stop the failure mode it points at. Use retrieval and "say not stated" instead.

A worked rewrite

A common pile of negatives:

Write a summary. Do not be too long. Do not include bullet points. Do not be too short. Do not introduce yourself. Do not be vague. Do not skip the recommendation.

This is a positive instruction screaming for help. Rewrite:

Write a 120–180 word summary in continuous prose.
Open with the headline finding. Close with one specific recommendation.

One positive line replaces six negatives. The output is more reliable because the model has a target instead of a minefield.

The honest rule

Use negatives sparingly, as scalpels, after the positive instruction is solid. If you are reaching for them as the main tool, the positive instruction is the problem.

Frequently asked questions

Should I write what NOT to do, or what TO do?

Always start with what to do. Add negatives only when the model has a strong default that fights you (e.g., starts every answer with "Certainly!"). A prompt with one positive instruction and zero negatives is usually clearer than a prompt with three of each.

Do "do not hallucinate" instructions help?

Marginally and unreliably. A better fix is to give the model the source material and tell it explicitly to say "not in this source" when the source does not cover the question. That removes the temptation to invent rather than asking the model to suppress it.

Why do some "do not" instructions seem to make things worse?

Because LLMs are next-token predictors and the tokens you mention in the prompt enter the conditioning. "Do not write a list" can sometimes increase the chance of a list because the word "list" is now in scope. Rephrase as "respond in continuous prose, no bullet points" and the bias flips.