Tell It What to Do, Not What Not to Do

The previous part was about the shape of the output. This part is about the very sentences you write in the request — and one common habit that seems helpful but actually does damage.

The habit is this: when the model's answer contains something you don't want, instinct says forbid it. “Don't use technical jargon.” “Don't be long-winded.” “Don't scare the patient.” The list of don'ts grows, and you expect the answer to come out cleaner. It usually doesn't — and the reason is worth hearing.

When you tell the model “don't use technical jargon,” you've only said where not to go; you haven't said where to go instead. The model now has to guess what to put in place of the jargon: a lay explanation? an analogy? dropping that part entirely? Every “don't” creates a blank the model has to fill by guessing, and the model's guess isn't necessarily yours. A positive instruction doesn't create that blank: “Write in language a patient with no medical background can understand, and wherever there's a hard concept, work in an everyday example.” Here the model not only knows what not to do — it knows exactly what to do instead.

See the difference on a real task. You want to write a preparation explanation for a patient who is about to have flap surgery.

The don't-driven version:
“Write an explanation about flap surgery for the patient. Don't use technical jargon. Don't make it long. Don't scare the patient. Don't talk much about complications.”

The do-driven version, the same request:
“For a patient who has flap surgery next week, write a short preparation explanation: four or five sentences, in plain everyday language, in a calm, confident tone. Focus on what is going to happen and what care is needed afterward.”

Both have the same goal, but the first left four blanks the model has to guess at, while the second filled those same four with a clear instruction. “Don't scare” became “a calm, confident tone”; “don't make it long” became “four or five sentences”; “no jargon” became “plain everyday language”; “don't dwell on complications” became “focus on the procedure and the aftercare.” Notice that nothing was dropped from the request; it was only translated from the language of prohibition into the language of instruction.

There is also a strange side effect of don'ts worth knowing: sometimes the prohibition itself highlights the very thing you didn't want. When you write “don't scare the patient about complications,” the word “complications” is now sitting in the middle of your request and the model's attention is fixed on it; in trying “not to scare,” it may start reassuring the patient about the very complications that weren't supposed to be the focus at all: “Don't worry about bleeding, infection rarely happens…” — and your patient gets scared precisely by reading these. A positive instruction (“focus on the procedure and the aftercare”) never lets those words into the room in the first place.

So the working rule: whenever you catch yourself writing a “don't,” stop for a moment and ask, “So what should it do instead?” The answer to that question is your instruction. Write that, and throw the “don't” away.

One point that might look like an exception but isn't: “Answer based only on the article I just gave you.” This isn't one of the problematic don'ts — it's a positive instruction; you're saying where to draw from, not what not to say. Setting a boundary, just like setting the shape, is telling the model what to do. And as it happens, it's one of your most-used ones: every time you hand over an article and want the model not to step outside it, this single sentence closes the boundary.

This part fits in one line: instead of a list of don'ts, say what to do. There is almost always a positive form of what you want, and it is almost always more precise.

In the next part we turn to a technique you've probably heard about more than any other: give the model a role, tell it “you are a periodontist.” It is the most famous piece of prompt-writing advice, and precisely for that reason we should say honestly what it does and what it doesn't.