AI
Prompting for Designers: How to Direct AI Like an Art Director
Vague in, vague out. References, constraints and iteration — the briefing skills designers already have, applied to generative tools.
Contents
The designers getting the most from generative tools aren't writing magic spells. They're doing what good art directors always did: giving precise direction with references, constraints and a clear definition of done. Prompting is briefing — and briefing is a design skill you've been practicing for years.
References Beat Adjectives
'Clean', 'modern' and 'premium' mean nothing to a model — they're averages of everything. '1970s Swiss poster, two inks, generous margins' is a direction. Collect a swipe file of eras, materials, photographers and movements; your reference vocabulary is your prompting vocabulary.
Constrain Ruthlessly
Every unconstrained dimension becomes average. Specify palette, format, lighting, lens, texture — and crucially, what to exclude. Negative constraints ('no gradients, no photorealism, no text') remove more badness than positive ones add goodness.
Iterate One Variable at a Time
Changing subject, style and composition between rounds teaches you nothing about what worked. Lock two, vary one. Five disciplined rounds beat fifty slot-machine pulls — and build a reusable sense of how the tool responds.
Know When to Stop Prompting
The model's job ends where your judgment begins: compositing, typography, brand fit, final polish. Ship-ready work is finished by hand, always. If you've spent an hour fighting the tool for something you'd draw in ten minutes, close the tab.
Frequently Asked Questions
Frequently asked questions
What makes a good AI image prompt?
Subject + era/material reference + composition + lighting + palette + exclusions. Lead with the most constraining element first, and keep a log of prompts that worked so successful patterns compound across projects.
Why do my AI generations all look the same?
Adjective-only prompting converges on the model's average. Add specific references, unusual constraints, and anti-references ('unlike typical SaaS illustration') to pull output away from the mean.