Writing a prompt that gets what you meant

Your prompt is being matched, not understood

The model learned from images paired with text. Your prompt is pushed toward the region of that space where similar captions lived. So the useful question is never "how do I explain what I want". It is "what would the caption under this picture have said".

That one reframing fixes most beginner problems at once.

A rewrite

Weak: *a beautiful picture of a woman selling fruit, very detailed, 8k, masterpiece*

Nothing here locates an image. "Beautiful" sits under a billion different pictures. "8k" mostly appears in stock-photo spam. "Masterpiece" was genuinely useful in 2022, when some models were tuned on data tagged with aesthetic scores and the word pulled toward the higher-rated pile. In current models it does close to nothing, and people keep typing it out of habit.

Stronger: *A fruit seller in her fifties at a covered market in Kumasi, building a pyramid of oranges on a wooden crate. Morning light coming through a corrugated roof in stripes. Camera just above the fruit, looking slightly up at her. 35mm, shallow depth of field.*

Same idea. But now there is a who, doing what, where, in what light, from what angle, in what medium.

The five things worth naming

1. 1Subject and action. Verbs beat adjectives. "Building a pyramid of oranges" does more than "detailed".
2. 2Setting, specific enough to be a place rather than a category.
3. 3Light. The highest-leverage words you have. "Overcast noon", "a single bare bulb", "late sun through dust" give three completely different pictures of the same scene.
4. 4Camera or medium. 35mm, macro, ink line drawing, gouache, screen-printed poster. Name nothing and the model gives you the average of everything, which is a faintly plastic digital painting.
5. 5Framing. Close-up, wide, from below, over the shoulder.

Everything else is decoration.

Negation does not work the way you expect

"A street with no billboards" often returns billboards. The prompt is scored as a whole, the word "billboards" pulls billboard imagery in, and "no" is a weak signal pushing back. Models with larger text encoders handle this better than the 2022 generation, but better is not reliable.

The fix is always the same: say what should be there instead. "A street lined with plain shuttered storefronts and bare concrete walls." Some tools also give you a negative prompt field, which sets what guidance pushes away from. Use it for qualities — blurry, watermark, harsh flash — rather than for objects you are trying to keep out of a scene.

Counting and spatial relations are weak

"Five people" gives you four or six. "The red cup to the left of the blue one" swaps them. This is not a prompting mistake you can write your way out of; it is a real limit, and the next lesson explains where it comes from. Fix it in editing, or frame the shot so an exact count does not matter.

Iterating like an engineer

Most people change five things at once, get a worse image, then change five more. Instead:

• Lock the seed. Change one clause. Look. That is a controlled experiment, and it is the only way to learn what your words actually do.
• Generate four cheap, low-step images to find a composition, then rerun the one you want at full quality on the same seed.
• Keep a text file of prompts that worked. You will not remember, and prompts do not transfer cleanly between models — one tuned for a model trained on terse alt-text is mediocre in a model trained on long descriptive captions.

The honest ceiling

Prompt craft has a limit, and it is lower than the internet suggests. A prompt gets you a good starting frame. Getting a finished image — the right hands, the right sign, a client's product that actually looks like the product — happens in editing. That is the next lesson but one.