How to write prompts that actually work

When a generation comes back wrong, the instinct is to switch models. Usually the prompt was underspecified and the model filled the gap with an average of everything it has seen.

1. Name the subject before you decorate it

Models weight the beginning of a prompt more heavily. Put the subject first and the atmosphere after.

Weak: Dramatic golden hour lighting, cinematic, moody, a lighthouse.
Better: A stone lighthouse on a cliff edge, shot at golden hour, long shadows, moody.

2. Say what the camera is doing

Photographic vocabulary carries a lot of meaning in very few words. "Close-up portrait, 85mm, shallow depth of field" fixes framing, compression and background blur in one clause. "Wide establishing shot, low angle" does the same for something else entirely.

3. Describe light, not just mood

"Moody" is an outcome. "Single hard light from the left, deep shadows, dark background" is an instruction that produces it. Models respond much better to the second because it maps onto things they have actually seen labelled.

4. Negatives are weaker than you think

"No text" often produces text. Mentioning a thing at all raises its salience. Where you can, describe the positive: instead of "no cluttered background", write "plain seamless grey background".

5. Change one thing at a time

The temptation after a bad result is to rewrite the whole prompt. Then you have no idea which change helped. Adjust one clause, re-run, compare. This is slower for one image and much faster over ten.

6. Lock the seed once you are close

A seed makes the randomness repeatable. With the seed fixed, a prompt edit shows you the effect of the edit rather than the effect of a different roll of the dice. Set it in the composer's seed control, iterate, then clear it when you want variations of the version you settled on.

Length is not quality

Very long prompts start contradicting themselves, and models resolve contradictions by ignoring parts of them. Two clear sentences with specific nouns generally beat a paragraph of stacked adjectives.