Why not a keyword list
Two years ago the accepted trick was to pile up comma-separated keywords: "cat, oil painting, golden light, 8k, award-winning". That worked because the text-reading part of the models was a short encoder trained on keywords.
It is not that any more. A full language model now sits in front of today's image models, and that changed what is worth writing: a proper sentence beats a keyword list, because it carries the relations as well. "Woman, red dress, beach, sunset" states four separate things; "a woman in a red dress walking along the shore at sunset" also states who is doing what and where. One model vendor sets exactly those two examples against each other in its own guide. Why this changed is unpacked in the post on how image generation works, down at the level of the mechanism.
The four layers
The vendor guides were written independently of one another, yet they recommend almost the same structure. Boiled down to four layers it goes like this — and most bad images are bad because the second, third or fourth layer is missing.
- Subject. What is in the picture, and what it is doing. Not "a dog" but "an old terrier asleep on a windowsill". The action is what makes a picture a picture; without it the model hands you a catalogue shot.
- Style. Photo, illustration, painting, 3D — and what kind within that. This layer decides the most, and it is the one people leave out most often. If you do not say, the model gives you the average for that subject.
- Light. Where it comes from, how hard it is, what colour it is. "Soft window light from the side", "hard midday sun with sharp shadows", "golden hour, backlit". Light is the difference between an amateur and a professional image, and the model follows it accurately.
- Composition. Where we are looking from, what is in the foreground, how much air there is around it. "Shot from below", "close-up", "subject in the left third, empty space on the right". There is a separate post on that one: setting the composition, because in 2026 it is no longer only a matter of asking nicely.
Before and after
The same image, phrased two ways. The difference is not the length but whether all four layers are present.
BEFORE
café, laptop, morning, cosy
AFTER
An open laptop stands on a wooden café table, beside it a half-finished
latte and a notebook. Documentary photograph, natural colours, slight
grain. Soft morning window light from the side, with long shadows across
the table. Shot from slightly above, close framing, the laptop on the
right, the left third left empty.
The "after" version is four sentences, and each one adds a different layer. What is not in it: not a single "8k", "award-winning" or "highly detailed". Those words counted for something on the older models; today the concrete description does the work.
The order matters
This is the part fewest people know, even though one vendor spells it out in its documentation: the model pays more attention to whatever comes first. If your prompt opens with the mood and ends with the subject, you get a weaker result than the other way round.
So the usable order is: main subject → action → style → light → composition → incidental detail. If something does not show up in the picture, the first thing worth trying is not a louder phrasing ("it is VERY important that…") but moving it earlier in the sentence.
How long should it be
Here the vendors disagree, and it is worth knowing that before someone sells you one of them as a rule. One open-weight family's documentation gives three bands: 10-30 words for a simple idea, 30-80 words for most scenes, above 80 only for multi-character or technical jobs — and it adds: "start short, and only add what changes the image". Another large provider recommends the opposite: for them "short and simple prompts give the best image". A third builds explicitly on detailed, multi-clause description.
The practical resolution is simple: your own tool's guide decides, not a general rule. What holds everywhere: start shorter, look at what is missing, and add exactly that. The biggest problem with an eighty-word prompt written in one go is not its length but that when the image comes out wrong, you have no idea which half-sentence spoiled it.
The camera words, and what they actually mean
"35mm, f/1.4", "Kodak Portra 400" and "three-point softbox lighting" really do work — the models learned from photographic descriptions, and these phrases reliably summon a look.
But there is an important caveat, which one vendor states outright: treat camera settings as cues about appearance, not as physical simulation. "f/1.4" does not compute optics; it summons the way a picture shot at f/1.4 looks. It follows that stacking them buys you nothing: "85mm f/1.2 softbox golden hour bokeh cinematic" is not five settings but five mood cues contradicting one another. One lens, one lighting situation, one film reference — that is enough.
What not to write in
Four habits that no longer help today, and on some models actively hurt.
- A negative prompt where there is none. Some recent models have dropped it; one vendor writes plainly that its model does not support it and asks you to describe what you want, not what you do not.
- The "with no X in it" phrasing. If you write "no cars in it", the model is still thinking about cars. Write what is there instead: "an empty street".
- Stacked quality words. "8k, ultra-detailed, award-winning, masterpiece" summoned the style of the older models. Today the concrete description does that job.
- Brand names and the names of living artists. The tools' terms push this onto you: you warrant that you have the right. There is a separate post on the legal side: who owns an AI image, which goes through it properly.
Text in the image
If the picture needs a caption as well, one rule is the same everywhere: put it in quotation marks. Every vendor documents this, and most of them also recommend describing the character of the typeface and where the caption sits. That is a topic of its own, though, because anything beyond short English wording is a separate difficulty: the post on text inside images opens that up.
A template you can overwrite
Swap out the square brackets. The order is deliberate — whatever comes first, the model takes more seriously.
[MAIN SUBJECT: who or what, and what they are doing — one sentence]
[STYLE: photo / illustration / painting / 3D, and what kind within that]
[LIGHT: where it comes from, how hard, what colour]
[COMPOSITION: where we look from, what sits where, where the empty space stays]
[IF TEXT IS NEEDED: the caption "in quotes", the character of the type, its position]
[ASPECT RATIO: 16:9 / 1:1 / 4:5]
What to take away
A good image prompt is not longer, it is more layered: subject, style, light, composition. Write a proper sentence instead of a keyword list, put the most important thing up front, and if something is missing from the picture, move it earlier in the sentence before you ask more loudly. Camera words summon an appearance, they do not simulate optics — one of each is plenty. And drop the quality words and the negative prompt unless your tool's guide explicitly asks for them. The same holds for text prompts, incidentally, only with different layers: the anatomy of a working prompt works that one out in full.