Writing what you do not want feels like the direct way to remove it. In practice a list of exclusions performs worse than a description of what you do want, and the reason is mechanical.
Naming a concept activates it
An instruction to avoid something still contains that thing. The model processes the concept, and the negation is a smaller signal riding on top of a much stronger one.
Text and image systems both show this. Asking for a scene without a particular object frequently yields the object, because the object was the most concrete idea in the sentence.
The result is not the model ignoring you. It is the exclusion being weaker than the mention, which is the opposite of how the instruction reads to a person.
Exclusions do not specify a replacement
Saying what to leave out leaves the resulting space undefined, and the model fills it with whatever it would default to. That default is often close to what you excluded.
Asking for prose that is not corporate, without saying what it should be instead, generally produces slightly different corporate prose rather than a different register.
A positive instruction constrains the space directly. Requesting short declarative sentences with concrete nouns removes the unwanted register by occupying the room it needed.
Long exclusion lists dilute each other
Prompts accumulate exclusions over time as each unwanted output is patched with another line. Ten avoid-this instructions do not enforce ten constraints.
Attention is finite across a prompt, and a long list of prohibitions competes with the actual task description for it. Frequently the task itself gets less weight than the appended rules.
A prompt that has grown this way usually works better after being rewritten from scratch as a positive description of the target output.
The exceptions are structural exclusions
Negation does hold reliably when it concerns form rather than content. Instructions such as returning only the requested field, or omitting a preamble, work well because they describe output shape.
These are enforceable because the model can check them against what it is producing, rather than needing to suppress a concept while writing about it.
The practical rule is that negation about format tends to hold and negation about subject matter tends to leak.
Convert each exclusion into a requirement
The repair is mechanical. For every avoid instruction, ask what should be present instead, and write that instead of the prohibition.
Avoiding a cluttered background becomes requesting a plain wall. Avoiding hedging becomes requiring each claim to be stated in one sentence without qualifiers.
The prompt usually gets shorter as well as more reliable, because a specification of the target does the work several prohibitions were attempting.