Prompting Is Art Direction, Not Magic

There is a genre of content built entirely on the idea that somewhere out there is a magic sentence. Buy the prompt pack, paste the incantation, receive professional imagery.

People buy them, paste them, get something that looks impressive in isolation, and then discover they cannot produce a second image that matches the first. Which is the moment it becomes obvious that a single good output is not the same as a usable visual system.

The reason prompt packs do not work is simple. A prompt is a brief. And most people write bad briefs, which I have written about at length in How to Write a Design Brief That Gets You What You Wanted, where the failure mode is identical: describing the output you imagine instead of specifying the conditions that produce it.

What a prompt actually is

When I art direct a photographer, I do not say "take a nice photo of a bottle." I specify the subject, the treatment, the lighting, the camera position, the background, the mood, the color handling, and the things I do not want in frame. Then the photographer applies their judgment inside those boundaries.

A prompt is that conversation, typed. The model is a very fast, very literal collaborator with no memory, no taste, and no ability to ask a clarifying question. Everything it does not know, it invents. Every dimension you leave unspecified is a dimension it will randomize.

"A photo of a bottle on a table, professional, high quality" leaves roughly nineteen decisions unspecified. That is why the result is generic. It is not because the model is weak. It is because you did not direct it.

The direction framework

Six dimensions, in the order I specify them. This is the same list I would give a photographer.

Subject. What is in frame, concretely. Materials, scale, quantity, condition, position. "A matte black ceramic bottle, unlabeled, standing upright" rather than "a bottle."

Medium and treatment. Photograph, illustration, render, collage. If photographic, what kind: editorial, product, documentary, studio. If rendered, what surface qualities.

Light. The most underused parameter and the one that changes the most. Direction, quality, and temperature. Hard directional light from the upper left. Diffused overcast. Single window source. Practical light in frame. Almost all of the difference between an amateur image and a professional one lives here.

Composition and camera. Framing, angle, focal length, depth of field, negative space, where the subject sits. "Shot at 85mm, shallow depth of field, subject in the left third, generous negative space to the right for type."

Color and mood. The palette, the grade, the contrast range. Muted and desaturated, or high-saturation and punchy. Warm shadows or cool ones.

Exclusions. What must not appear. This is the important one and it gets its own section.

Exclusions matter more than additions

This is the thing that separates people who get usable output from people who get impressive one-offs.

Adding descriptors makes an image more elaborate. Removing possibilities makes it consistent, and consistency is what a brand actually needs. Ten images that share a treatment are worth more than one spectacular image that matches nothing else you own.

The most useful example I have is from my own studio. Every generated asset for Infinite Design runs against a locked set of parameters and a fixed exclusion list, and one of the excluded items is red. Red and crimson never appear in anything the studio generates.

That started as a system decision rather than a taste one, and it had a consequence I did not initially plan. Because the studio brand can never produce red by accident, red became available as the accent color for my personal brand. The two identities separate themselves automatically, with nothing to enforce by hand.

That is what a good exclusion does. It is not a restriction on creativity. It is a structural decision that removes an entire category of future inconsistency.

Standard exclusions worth having in most brand work: text and lettering, since models still produce garbled type. Watermarks and logos. Clutter and busy backgrounds. Oversaturation. Lens flare, unless you asked for it. Specific colors that fight your palette. Whatever cliché the model defaults to in your subject area, and there is always one.

Iterating like a designer

The behavior that separates the two approaches is entirely about how you handle the second attempt.

Change one variable at a time. If you rewrite the whole prompt after a bad result, you learn nothing, because you cannot tell which change did what. This is the slot machine approach: pull, discard, pull again, and never build any knowledge.

Keep a log. Prompt, parameters, seed, result, and one line on what you would change. Within a few weeks you have your own reference for what actually works in your visual language, which is worth vastly more than any prompt pack because it is calibrated to your brand rather than to a stranger's.

Lock your parameters. Style settings, version, aspect ratios, and stylization values should be fixed across a project. Varying them between images is the fastest way to end up with a set that does not cohere.

Use seeds and references for consistency. If a set of images needs to look related, this is not optional.

Select ruthlessly. Generating a hundred options is free. Choosing the right one is the job, and it is the same judgment that has always been the job. Most people generate too much and select too little.

Treat outputs as raw material. The generated image is rarely the final asset. It is a plate. Color, crop, retouch, composite, and integrate it into the layout the way you would with a photograph you commissioned.

Where it still fails in brand work

Being honest about this matters, because overselling the tool is how people end up disappointed in a way they blame on themselves.

Logos and marks. It cannot do this. A logo needs to be vector, geometrically precise, reproducible at any scale, and defensible as an original. Generated marks are none of those things. Anyone selling AI logo generation to a business is selling them a problem.

Type. Still unreliable, still garbled, and even when it is legible it is not a licensed typeface you can use consistently elsewhere.

Exact repeatability. Getting the same product, the same person, or the same room across a series remains difficult. Improving, but not solved.

Precise product accuracy. If the image needs to show your actual product, generation is the wrong tool. Photograph it.

Anything requiring real specificity. A real location, a real person, a real event. It will produce something plausible and wrong, which is worse than nothing.

Legal clarity. Rights, ownership, and training-data provenance remain unsettled and vary by jurisdiction. For a client's core brand assets, that is a risk worth taking seriously rather than assuming away.

The short version

There is no magic sentence, and the search for one is the reason prompt packs sell.

What there is instead is the same skill that has always separated good visual work from mediocre visual work: knowing what you want, being able to describe it precisely, knowing what to exclude, and being able to tell which of a hundred options is the right one.

The tool got faster. The job did not change.