How to write a prompt that gets the image you pictured
The five ingredients, worked examples, and the fastest way to fix a result that missed.
SamImageGenerator ·
A good prompt is a description, not a command and not a list of tags. The model reads it the way it read the captions it learned from, which means a plain sentence outperforms a pile of keywords almost every time.
The five things worth naming
Subject, setting, light, framing, finish. Not all of them every time, but the ones you leave out are the ones the model decides for you. "A cat" gives you the average cat in the average room. "A ginger cat asleep on a windowsill, late afternoon sun through the glass, close-up, shallow depth of field" gives you the picture you had in mind.
| Weak prompt | Why it fails | Better |
|---|---|---|
| a beautiful landscape | no subject, no light | a valley of pine forest under low mist at sunrise, wide shot |
| product photo, professional, 4k | tags, not a description | a ceramic mug on a white seamless backdrop, soft studio light |
| a man, no hat | mentions the hat | a man with short dark hair, plain background |
| minimalist detailed painting | contradiction | a minimalist painting, two flat colours, one shape |
Iterate on one thing at a time
When a result is close but wrong, change one element and regenerate. Change three and you learn nothing about which one did the work. Keep what worked: copy the prompt, edit the single phrase, run it again.
If the result is ninety per cent right, stop prompting and start editing. Upload it, brush the part that is wrong and describe only that — the rest of the frame stays untouched, which is a different and much cheaper operation than rolling the dice again.
The parts that are not the prompt
Aspect ratio and style are separate controls, and both are more reliable than asking for them in words. Ask for "16:9" in the text and you are making a request; pick the ratio and it is a parameter. The full checklist lives in the prompt guide, and the editing path is in editing a photo with AI.