What a prompt is and how to write one
A prompt is a text description of what the neural network should do. Not a command and not a spell: the model does not “understand” you, it completes the picture from the words you gave. The more specific the words, the more predictable the result.
The rule that everything else follows from
Whatever you did not describe, the model will make up on its own, and differently every time. So a good prompt differs from a bad one not in length and not in “magic words”, but in how many decisions you made instead of the model.
What a description is made of
| Part | What it controls |
|---|---|
| Object | what exactly is in the frame |
| Details | material, color, condition |
| Light | mood and volume |
| Background | where it happens; without it the model makes up its own |
| Shot | close-up, medium, wide |
A matte white ceramic mug on a wooden table.
Close-up, the object takes up two thirds of the frame.
Soft diffused light from the left, a light shadow on the right.
Background: a blurred kitchen in light tones.
Format 4:5, vertical.Words that do nothing
Not like this: “beautiful mug, super quality, 8k, masterpiece”
“Beautiful” and “masterpiece” carry no visual information, the model will read them its own way, and differently on every run. “8k” is useless here too: resolution is set by a separate setting, not by a word in the text. In the end you handed all the decisions to the model and got a random result.
How to fix it when it came out wrong
Change one element at a time. If you rewrite the whole prompt, you will not know what exactly worked and will not be able to go back to the good version.
- The wrong object: move it to the very beginning.
- The wrong mood: edit the light, it has the strongest effect.
- Extra details in the frame: describe the background explicitly, otherwise the model will make it up.
- The wrong angle: add the shot type: close-up, medium, wide.
What a prompt cannot do
- Set the resolution or duration. These are settings next to the input field, words do not change them.
- Guarantee a repeat. The same text gives similar but different results, most models have no way to lock it.
- Describe a specific face. Words cannot achieve that, you need a reference, that is, an uploaded image.