Style language — say it so a model can act on it
There are two vocabularies in design conversation. One is for clients: clean, modern, premium, eye-catching, it should really pop. The other is the one you use across the desk from another designer: risograph grain, 2px uniform stroke, duotone, halftone at 45 degrees, shot on 85mm at f/1.8. The first vocabulary describes a feeling about the output. The second describes the output.
Image models only hear the second one. "Clean and modern" has appeared next to ten million wildly different images in training data; it selects for nothing. "Flat vector, 2px stroke, isometric grid, no gradients" selects for a look you could pin on a moodboard.
The fix is the same translation you already do when a client says "make it feel premium" and you quietly write down increase whitespace, tighten the palette to two colors, switch to a serif with high contrast strokes. Do that translation before the prompt, not after the disappointing render:
| Client says | You prompt |
|---|---|
| clean, minimal | flat vector, single-weight line, generous negative space, two-color palette |
| warm, artisanal | risograph print, visible grain, terracotta and cream, hand-drawn irregularity |
| premium, editorial | photograph, 85mm lens, shallow depth of field, muted palette, soft window light |
| retro, playful | screen-print texture, halftone dots, off-registration, 70s poster palette |
Three token families do the heaviest lifting:
- Medium and process tokens — risograph, screen print, linocut, collage, 3D render, photograph. Models trained on captioned images know print processes shockingly well, because humans caption them.
- Craft-spec tokens — 2px stroke, isometric, duotone, halftone, lens and lighting language for photo work. These read like your production notes because they are production notes.
- Named-palette tokens — terracotta, deep teal, off-white, or literal hex codes. "Colorful" is not a palette. Your kit (next lesson) makes this section automatic.
One warning from the other direction: piling on forty tokens is its own failure mode. A prompt with three contradictory style families ("risograph photoreal 3D render") averages them into mud. Pick one medium, one process, one palette — the same discipline as a style tile.