Lucero

Negative Prompting: Telling AI What Not to Do

By the Lucero Editorial Team \u00b7 Updated 2026-08-07 \u00b7 4 min read

TL;DR

Telling an AI what to avoid, like do not use bullet points or do not mention pricing, works well for format and tone constraints. It works less reliably for suppressing specific facts or ideas, which the model may still reference indirectly.

Where negative prompting works well

Format and style constraints, like avoiding a bulleted list or overly formal tone, are reliably respected when stated explicitly.

Where it gets unreliable

Asking a model to avoid a specific topic or fact entirely sometimes still lets it slip through indirectly, since suppression is harder than direction.

Positive framing often works better

Rephrasing a negative constraint as a positive instruction, like write in plain conversational language instead of do not sound corporate, tends to be more reliable.

Combine both when needed

Pairing a brief negative constraint with a positive alternative usually outperforms either alone, building on the prompt engineering guide.

Frequently asked questions

Why does the AI sometimes ignore a negative instruction?

Suppressing a concept is harder for the model than following a direct positive instruction, so it occasionally leaks through.

Is it better to list many things to avoid or just one?

Fewer, clearer constraints tend to be followed more reliably than a long list of things to avoid.