How Nomi builds a prompt

The sentence you type is only part of what the drawing model reads. Nomi appends the rules of the draw at the last step, at the one point nothing downstream can drop them.

Last updated 20 August 2026

Your words go first

Whatever you wrote, or whatever Nomi wrote for you in Agent, leads the prompt. Everything else is added underneath, and all of it says the same thing about precedence: it fills in what the request did not say, and never overrules what it did.

A short brief ends up a small part of the whole. One measured case: 134 characters typed, 1,719 characters sent.

The rules that ride along

Which of these appear depends on the shape you asked for and on what is attached.

  • The safe zone, only for the tall story and reel shape: keep text, logos and the call to action in the middle, because the platform's buttons cover the top and bottom bands.
  • Nothing attached: every word in the picture must be one the prompt named, and any surface it gave no words to stays blank rather than getting invented lettering.
  • A product picture: its shape, packaging, label text, colours and logos are copied exactly, and the rest of that photograph is left behind: the light, the background, the props beside it.
  • A picture that is only a reference: the ad borrows layout, lighting, mood and styling, and never the other brand's logo, packaging or on-image text.
  • A reframe or a translation replaces all of the above: the same ad in a new shape, or in a new language with that market's own price and offer.

What the brand contributes

The brand's look goes last, and only the part a still picture can act on: the general look, the colour palette, the photography direction, the composition preferences, the visual identity rules and the forbidden styles, in that order.

The typeface, the icon style, the corner radius and the logo placement are not sent. They mean nothing to an image model, and belong to the editor instead.

The prompt has a hard ceiling. The brand's look takes whatever room is left, dropping whole fields from the bottom of that list rather than cutting one in half. A long, detailed brief therefore gets less brand, which is the right way round.

The Image lane holds two fields back

Photography direction and composition preferences describe a picture rather than a look, and in the Image lane your one sentence is the whole request. Two sentences of art direction will beat it: a request for a screenshot on a phone once came back as the brand's own stock product shot instead. So those two are withheld whenever the words are yours, and kept in Agent, where the model writes a long scene of its own.

If an ad comes back generic, the brand is usually thin rather than the sentence. The look fields are most of what a prompt is made of.