The better the input, the better the output

The most common misconception about generative AI: that it removes the need for a briefing. The opposite is true. The model does exactly what the input allows. No more.

For the GINESTA Immobilien campaign, the prompts ran to about 2,500 characters each. Not because longer is better, but because precision is: composition rules, light logic, material behaviour, style constraints. Everything a good art director would brief a photographer on, translated into language a model can execute.

The workflow that has proven itself:

1. Translate the briefing. An agency briefing describes intent. A prompt briefing translates intent into measurable image rules: what must be true in every frame, what may vary.
2. Spar before you render. I use a language model as sparring partner to sharpen structure, wording and variants before a single image is generated. Cheap iterations first, expensive ones later.
3. Version your prompts. Prompts are code. Version them, compare outputs side by side, keep what worked and know why.
4. Fine-tune for consistency. A campaign is not four good images. It is four images that belong together. The last mile is always about consistency, not beauty.

The principle scales down, too: even a single packshot gets better when the input is treated like a briefing, not like a wish.