Skip to content
Prompt Corpus
Technique12 August 20263 min read

Why "don't mention pricing" makes it mention pricing

Negative instructions are the most reliably disappointing thing beginners write. The reason is mechanical, and the fix takes one sentence.

You write a careful prompt. Somewhere in it you add: do not mention pricing. The response mentions pricing. You add emphasis — do not, under any circumstances, mention pricing — and now it mentions pricing twice and apologises for it.

This is not the model being contrary. It is a predictable consequence of how the thing works, and once you see why, you stop writing prompts this way.

What a prohibition actually does

A language model does not evaluate your instruction and then apply it as a filter over its output. There is no stage where a finished response gets checked against your rules. There is only next-token prediction, conditioned on everything in the context window.

Your prohibition is part of that context. The tokens for "pricing" are now sitting in the sequence, and every subsequent prediction attends over them. You have made the concept more available, not less. The word "not" is a weak signal spread across a long generation, competing against the much stronger signal of a topic you just introduced.

It is closer to "don't think about an elephant" than most people expect. The comparison gets made casually and it happens to be mechanically apt.

The fix

State the positive. Describe the space you want the output to occupy, rather than fencing off a piece of it.

Instead of don't be too formal, write write as if explaining to a colleague over coffee. Instead of don't mention pricing, write cover features and availability only. Instead of don't make it too long, write under 120 words.

Each replacement does the same job without putting the unwanted concept into the context. And each one is more specific, which is a second benefit: "not formal" covers an enormous range, while "explaining to a colleague over coffee" is one recognisable register.

When negatives do work

This is not an absolute ban, and it would be sloppy to present it as one.

Negative instructions about form are generally reliable:

  • No preamble
  • No bullet points
  • No closing summary
  • Do not use bold

These work better for two reasons. They describe the shape of the output rather than its subject, so there is no unwanted topic being seeded. And their violation is immediately visible in the first few tokens, where the instruction is still close by in the context.

Negative instructions about subject matter are the unreliable ones, because naming a subject is exactly how you make a subject likely.

The reliable pattern

When you genuinely need something excluded, pair it:

Cover features and availability. If the reader asks about cost, direct them to the sales team rather than quoting figures.

That gives the model somewhere to go. A bare prohibition leaves a hole, and the most probable continuation frequently fills it with the very thing you excluded. Naming the alternative means there is a better path available than the one you were trying to close.

If you want the mechanism behind this properly, it is in what a language model actually is and why prompting works at all. The short version is that your prompt selects from behaviour the model already has. A prohibition narrows badly. A positive description narrows well.

All notes