The logic of the visual brief #
The biggest mistake is starting with the subject. If you tell an LLM to "create a beautiful watch," it has no context. If you tell it to "create a square ecommerce hero image for a stainless steel field watch," you've given it a functional goal and a framing constraint. This narrows the probability space and stops the model from hallucinating random artistic choices.
When dealing with products, be obsessive about what must remain recognizable. Don't use fluff; use technical descriptors. Mention the silhouette, the material finish, and the exact placement of labels. If you're doing a reference-guided edit, explicitly separate the "constants" from the "variables."
Composition and Lighting Hierarchy #
A cluttered prompt leads to a noisy image. If you describe the subject, the background, and the lighting all as "dramatic," the AI usually over-saturates everything. I prefer using concrete cinematography terms to direct the eye. Instead of "cool style," try:
Centering:"subject on the right third" or "centered subject"** Depth:"softly blurred background" or "close-up detail" Light Source:**"hard midday shadows" or "cool rim light against a dark background"
Lighting instructions are almost always more effective than broad style labels. "Soft window light from the left" is a directive the model can actually execute, whereas "elegant lighting" is subjective and unpredictable.
Handling the Typography Trap #
LLMs are getting better at text, but for professional work, it's still a gamble. My workflow is to treat text as a separate design step. I prompt for negative space so I can add the actual copy in post-production.
Leave the upper half low-detail for final event text. Do not render words, dates, logos, or decorative borders.
The Framework #
If you want a repeatable AI workflow for images, stop winging it and use a skeleton. This forces you to define the constraints before you hit generate.
Create a [image type] of [main subject] [action or state].
Place it in [setting]. Use [composition and viewpoint], with [lighting].
Make [key detail] the focal point. Preserve [constraints].
Leave [position] clear for [optional copy].
The key to prompt engineering here is isolated iteration. If the lighting is off, change the lighting sentence only. If you change three variables at once, you have no idea which word actually fixed the image. This is a practical tutorial in control—stop guessing and start directing.
Next Stop letting your AI agents leak their internal logic to anyone →