Prompting for Designers: The 5 Instructions That Change Everything

Most bad AI generations are not the model's fault. They're the brief's fault. After watching designers run thousands of prompts inside Koha, the same five fixes show up over and over. Get these right and the hit rate stops feeling like a slot machine.

None of this is about magic words or secret syntax. It's about treating the prompt the way you'd treat a brief to a junior designer: specific, visual, and easy to act on.

1. Name the medium, not the vibe

“Cinematic and moody” means nothing. “35mm film, soft window light from camera left, eye-level, shallow depth of field” means something. The model has seen millions of images tagged with concrete production language and almost nothing tagged “moody.”

Pick from the vocabulary a photographer or DP would use: lens (50mm, 85mm, anamorphic), light (overcast daylight, hard rim light, practical neon), angle (low-angle, top-down, three-quarter), and stock or finish (Kodak Portra, matte print, polished render). Stack two or three of those. That's your style.

2. Anchor with a reference, then describe the delta

If you can show it, don't spell it out. Drop a reference image into the prompt and the model already has 90% of what you mean by “that warm editorial look.” Now your job is just to describe the change you want from the reference.

“Same lighting and palette as this, but swap the subject for a ceramic vase” outperforms a 200-word style description every time. In Koha, references sit right next to the prompt box, so the delta-from-reference workflow is the default, not a workaround. If you're new to dropping references into image, video, or 3D generations, the generation walkthrough shows the exact flow.

3. Say what shouldn't be there

Designers spend ages adding more positive description when the real problem is something the model keeps inserting. Text on packaging. Logos. A second hand. Lens flare. A bowl of fruit nobody asked for.

Be explicit about exclusions: “no text, no signage, no people in background, no lens flare.” This single move probably fixes more failed generations than any other. The model can't read your mind about restraint, only about presence.

4. Lock the composition before the content

A good prompt reads like a shot list, in order: frame, subject placement, then content. “Wide shot, subject in lower-right third, negative space top-left, horizon at bottom quarter” is a composition. Whatever you put in that frame will look intentional.

Most prompts skip straight to content (“a hiker on a mountain”) and leave composition to the model. That's how you end up with ten generations of the same centered hero shot when you needed something for a banner crop.

5. Change one variable per iteration

When something is close but not right, the temptation is to rewrite the whole prompt. Don't. Change one thing: the lighting, or the palette, or the pose. If you change three things and the next generation is worse, you have no idea which change caused it.

This is where Koha's prompt history panelearns its keep. Every prompt and every generation stays pinned in a side panel, so going “back two steps and try a different light” is one click, not a memory test. Iteration becomes a real feedback loop instead of a guessing game.

The shape of a good prompt

Put the five together and a strong prompt looks something like this: medium and lighting first, then composition, then subject, then exclusions, with a reference image attached. “Editorial product photo, soft overcast daylight, 50mm, three-quarter view, subject lower third, ceramic vase with single dried branch, neutral linen backdrop, no text, no props.”

That's not a prompt the model is going to misread. And once you have that baseline, you change one variable at a time until it's exactly the thing in your head.

Try the five on your next generation. The difference is usually obvious by the second iteration. If you're curious how this fits into a wider team workflow, here's how Koha pulls generation, editing, and review into one place.