Most bad AI images are bad briefs.
“A picture of a coffee shop” is not a brief. It does not say what is in the frame, where the camera is, what kind of picture it is, or what it must not contain — so the model decides all of that, and you spend four more generations discovering you disagreed.
The fix is not a longer prompt. It is a prompt with the six things that actually determine the output, and nothing else competing with them.
The six components
Not a template to fill in mechanically — a checklist to run against a prompt before you spend a credit on it. Missing components are where the model gets to improvise.
Subject
What is in the frame, concretely
a single ceramic pour-over cone
Framing
Camera position and crop
top-down, centred, generous negative space
Medium
What kind of picture this is
flat vector illustration, thick uniform strokes
Light
Direction and quality
soft studio key from the left, gentle shadow
Palette
Named colours, counted
two colours only: warm cream and near-black ink
Exclusions
The failures you keep hitting
no text, no letters, no watermark, no gradients
all six, assembled
“A single ceramic pour-over cone, top-down and centred with generous negative space, flat vector illustration with thick uniform strokes, soft key light from the left casting a gentle shadow, two colours only — warm cream and near-black ink. No text, no letters, no watermark, no gradients.”
One reference image beats a paragraph of adjectives
This is the single highest-leverage technique on this page, and most people discover it after producing forty inconsistent images. “Warm, minimal, editorial” means something different on every generation. An actual image means exactly one thing.
So the working method is: iterate on one image until you like it, then pass its URL in referenceImages for everything else. Style becomes something you carry rather than something you re-describe.
Describing the style
“warm, minimal, editorial, modern, clean”
Five adjectives. Twenty possible outcomes.
Carrying the style
referenceImages: ["…/approved-hero.png"]
One image. One outcome, repeatedly.
Five failures, and what to write instead
The image has garbled text in it
why
You asked for a sign, a label, a poster — or you did not forbid text at all.
write instead
Add "no text, no letters, no words, no watermark" to every prompt. Overlay real type in HTML afterwards.
Everything comes out looking like a stock photo
why
The prompt named a topic, not a picture. "Teamwork" resolves to the average of every teamwork image ever captioned.
write instead
Describe an object and a composition, not a concept. "Two hands passing a paper crane over a desk" beats "collaboration".
The subject is tiny in the middle of the frame
why
No framing instruction, so the model defaulted to a safe centred composition.
write instead
State the crop and the camera: "close crop, subject filling the frame, shot from slightly above".
Colours change on every generation
why
Adjectives instead of colours, and no reference image.
write instead
Name the palette and count it: "three colours only — deep teal, bone white, burnt orange". Then add a reference.
The cutout has ragged edges
why
The prompt described a scene, so the matting model had no obvious subject to isolate.
write instead
For anything you intend to cut out: "one subject, centred, plain white background, isolated".
Iterate cheap, finish expensive
Prompt length does not affect cost — the model does. So the sensible loop is to do all your wording experiments on the fast tier and only spend premium credits on the version you already know you want.
Explore
8 credits a shot
Four or five fast generations to find the composition and palette.
Lock
free
Pick the winner. It becomes your reference image for everything downstream.
Finish
30 credits a shot
Re-run the winning prompt on a premium model, once.
Prompting questions
How long should an image prompt be?
Long enough to specify subject, framing, medium, lighting and palette — usually one or two dense sentences. Beyond that you get diminishing returns and, past a point, dilution: a prompt with thirty competing details gives the model permission to ignore most of them.
Do negative instructions work?
Partially, and they are worth including anyway for the ones that matter. "No text, no letters, no watermark" is the highest-value negative you can write because baked-in text is the failure that most often forces a regeneration.
Why do my images drift in style across a batch?
Because words are a lossy description of a look. Two prompts with the same adjectives can land in very different places. Pass an approved image in referenceImages and the model has the thing itself rather than a description of it — this fixes more drift than any prompt rewriting.
Should I ask for a specific artist or brand style?
Naming a living artist to imitate their style is a bad idea both ethically and practically — it produces derivative work and may breach the terms you are operating under. Describe the qualities instead: the medium, the palette, the stroke weight, the era. You get more control and a result that is yours.
Is a longer prompt worth more credits?
No — cost depends on the model, not the prompt length. That means iterating on wording is cheap on a fast model, which is exactly where you should be doing it before you re-run the winner on something expensive.