Why does a prohibition work badly?
Because the model picks the most likely continuation, and a prohibition requires it to bring to mind the very thing you are forbidding. When you write "don't use jargon", jargon is now in the surrounding context, so its probability did not go to zero — it went up.
There is a second reason on top of that: a prohibition does not say what to do instead. "Don't be bureaucratic" — fine, but then what? The model is left guessing, and guessing usually falls back to the default, which was exactly what you did not want.
The rewriting rule
Every prohibition has a positive twin. That one works.
| Prohibition | Statement |
|---|---|
| Don't write at length | At most three paragraphs |
| Don't use jargon | Write so a non-specialist understands it |
| Don't invent data | Work only from the material supplied; if it is not in there, say so |
| Don't use bullet points | In flowing prose, in connected paragraphs |
| Don't apologise three times | Apologise once, in the first paragraph |
The pattern is the same throughout: put the thing you want where the forbidden thing was, in a measurable form. "At most three paragraphs" is checkable; "don't write at length" is not.
The reason that improves things a lot
There is one technique the providers' guides single out: say why.
Not out of politeness. The reason narrows the range of possible interpretations. Compare:
Never use ellipses.
Your answer will be read aloud by a text-to-speech engine, so never use ellipses.
The second avoids not only ellipses but everything else that would go wrong when spoken — tables, emoji, parenthetical asides. With half a sentence you have set constraints you never listed.
What it will honour as a prohibition
Not every prohibition is bad. Anything specific, closed and checkable it follows reliably:
- A closed list: "the category can only be one of these four values"
- A specific field: "no phone number should appear in it"
- A specific name or product: "do not mention competitors"
- A specific figure: "the price must not appear in the letter"
The difference from a general matter-of-taste prohibition is that you can tell at a glance whether it was honoured. If you cannot decide, neither can the model.
The quantitative limit worth knowing about
It follows five rules reliably. It does not follow forty with exceptions and precedences between them — and worse, it does not tell you it dropped three. That is the fifth point in What AI Is Not Good For.
Three practical conclusions follow:
- Rank them. What really matters goes first and gets highlighted. What would merely be nice, leave out.
- Break it into steps. Rather than one prompt handling forty rules, several steps each handling fewer.
- Whatever a program can check, let a program check. Character counts, required fields, forbidden words — don't leave those to the model; a simple check afterwards is more reliable.
Where do constraints belong?
Whatever is the same for every task — the role, the tone, the forbidden topics, the spelling of the company name, the required format — belongs in the system prompt, where you write it once.
Whatever changes per task belongs in the prompt. And one practical rule: put the important constraint at the end of the prompt rather than in the middle — in a long text, the model uses information in the middle less reliably.
One rewrite, worked through
Before:
Write a customer reply. Don't make it too long, don't use technical terms, don't promise anything, and don't be so formal.
Four prohibitions, none of them measurable. The result will be unpredictable.
After:
Write a customer reply, at most 120 words, in three paragraphs. Direct, factual tone — the way you would explain it to a colleague. State as fact only what appears in the material below. We are not offering a deadline or compensation; if the customer asks for that, say you are passing it to a colleague.
The same four intentions, but each as a statement, a size or a closed rule. This it will honour — and you can check whether it did.
What to take away
Don't write what it should not do, write what it should do instead — measurably where possible, and with a reason if half a sentence covers it. What remains a prohibition should be specific and checkable. Everything else should not be left to the prompt at all.