Most prompts that fail do not fail on vocabulary. They fail because they are a pile of adjectives with no spine, and the model has to guess which of the twenty things you mentioned is the picture and which are decoration. A prompt is easier to debug when it has a fixed skeleton, because when the image is wrong you can point at the slot that produced the wrong part instead of rewriting the whole sentence and hoping.
The six slots
| Slot | What it decides | Example |
|---|---|---|
| Subject | What the picture is of. One thing, named concretely. | an elderly fisherman |
| Action / state | What it is doing, which fixes the pose and the body language. | mending a net, seated on a crate |
| Environment | Where it happens, which drags in half the colour palette for free. | on a wet stone harbour at dawn |
| Light | The single strongest control over mood and realism. | low side light, heavy fog diffusion |
| Camera / medium | Lens, distance and material — whether it reads as photo, paint or render. | 85mm, shallow depth of field |
| Style / treatment | The finish laid over everything above. | muted documentary colour grade |
Read those left to right and you get a sentence a person could follow: an elderly fisherman mending a net, seated on a crate on a wet stone harbour at dawn, low side light through heavy fog, 85mm with shallow depth of field, muted documentary colour grade. Nothing in it is fancy. It is legible precisely because each clause answers a different question.
Why order changes the output
Text-to-image models weight the beginning of a prompt more heavily than the end. That is not a rule someone invented for style guides — it falls out of how the text encoder truncates and how attention distributes across a long token sequence. The practical consequence is blunt: whatever you put first is what you get, and a detail buried in position forty is a suggestion rather than an instruction.
- —If the subject is in the last third of the prompt, expect the environment to eat it — you asked for a portrait and got a landscape with a person in it.
- —If two style words disagree ("oil painting" and "photorealistic"), the earlier one usually wins, and the later one shows up as a weird texture rather than being ignored cleanly.
- —Moving a clause forward is a stronger edit than adding emphasis words. Try relocation before you try adding "very", "highly detailed" or repeated adjectives.
A fast diagnostic: cut your prompt in half at the midpoint and generate from the first half alone. If that image is already what you wanted, the second half was decoration. If it is not, the important part is in the wrong place.
How long is too long
There is no single token limit across the three engines we publish with, but the behaviour is consistent: past roughly forty to sixty meaningful words, added detail starts trading against detail you already had. You will notice it as the model quietly dropping one element — the hat disappears, the second character never shows up — rather than as an error. When that happens the fix is subtraction, not more instructions telling it to stop.
The exception is a prompt with genuinely many required, non-conflicting objects: a still life with eight named items can be long, because every clause adds a distinct thing rather than another modifier on the same thing. Length hurts when it stacks adjectives; it is tolerable when it stacks nouns.
What to write when you do not know what you want
Start with subject and light only, generate, then add one slot at a time. This is slower per image and dramatically faster overall, because every addition is a controlled experiment: you know exactly which clause caused the change. Writing a full six-slot prompt from a blank page and then tweaking words at random is how people end up with forty generations and no understanding of any of them.
- Read finished prompts — Every page shows the exact wording that produced the image above it
- Which camera and light terms actually do something
- When the model ignores part of your prompt
- Aspect ratio and framing