What a negative prompt is
Every image model steers toward the words you give it. A negative prompt is a second list the model steers away from. In Midjourney it lives inline: --no fence, text. In Stable Diffusion it has its own box under the prompt. Same idea, same rules.
It is a correction tool, not a design tool. The positive prompt does the designing. The negative prompt removes the one thing that keeps showing up wrong after you have seen the output. That order matters: negatives written before the first render are guesses.
What belongs in one
Three kinds of things earn a place, and each one earns it by appearing in an image you already made.
- Recurring intruders. The model keeps adding a fence, a watermark, a second moon. Name the object:
--no fence. - Style bleed. You asked for a linocut and it keeps rendering photo texture.
--no photographpushes it back toward the print. - Format junk. Text, borders, frames, signatures. These four cover most of what a clean image needs excluded.
What doesn't belong
The forty-term boilerplate you see pasted everywhere: ugly, deformed, blurry, bad anatomy, low quality, worst quality
and thirty more. It survives because it is copied, not because it is tested. Modern models already avoid most of that list, and every extra term dilutes the steering the useful terms provide.
Negatives also cannot fix a vague positive. If the image is boring, the fix lives in the prompt: a sharper subject, one concrete style word from the styles wall, one lighting word from the lighting wall. Say what you want before listing what you don't.
One caution: a negative still puts the word in play. Naming a rare object you fear can remind the model it exists. If --no snake starts producing ropes and vines, drop the negative and rewrite the scene without the fear.
The one-term method
Treat negatives like wall words: one change at a time.
- Render with no negative prompt at all.
- Look at what actually went wrong. Not what might.
- Add one term for the worst offender. Render again, same seed if your tool allows it.
- Keep it only if the offender left. Repeat until the image is clean, then stop.
Most working prompts end up with zero to three negative terms. If yours needs ten, the positive prompt is doing too little work; take it to the builder and rebuild it slot by slot.
Published here first.