If you have ever typed a prompt, waited thirty seconds, and gotten back something that looks like a wax museum version of what you pictured, you are not alone. Most people blame the tool. Nine times out of ten the problem is how the request was written. You can make images from a text prompt in seconds, but getting a usable result comes down to a few habits that separate a lucky accident from a repeatable process.
Describe the light, not just the subject
Beginners write "a woman in a red dress." That tells the model what is in the frame and nothing about how it should feel. Add direction and quality of light and the whole image changes. "A woman in a red dress standing near a tall window, soft afternoon light from the left, gentle shadow across the floor" gives the model a scene to build, not a sticker to place. Light is the single biggest lever you have, and it costs you six extra words.
Name the camera or medium
Models have absorbed millions of photos, paintings, and film stills. When you say "shallow depth of field, 85mm portrait lens," you pull the output toward real photography. When you say "gouache illustration with visible brush texture," you get paint. Leaving this out means the model picks for you, and its default is usually a smooth, plastic-looking middle ground that reads as fake.
Stop stacking adjectives
"Beautiful stunning gorgeous breathtaking masterpiece" adds nothing. These words are so overused in training data that they blur together into generic polish. Swap them for specific nouns and verbs. Instead of "beautiful forest," try "a forest of birch trees with pale bark, morning fog sitting low between the trunks." Concrete details survive the translation from words to pixels; vague praise does not.
Fix composition before you fix detail
If the framing is wrong, no amount of prompt tweaking will save it. Decide where the subject sits in the frame. Do you want a wide establishing shot, a tight crop, or a full-body shot with headroom? Say it. "Wide shot, subject on the right third, empty space on the left for text" is a writing instruction the model can follow. Cropping and reframing after the fact almost always costs you resolution.
Iterate on one variable at a time
The fastest way to learn what a model understands is to change exactly one thing between runs. Same seed if your tool allows it, same prompt, but swap "golden hour" for "overcast." Look at what moved. When you change four things at once and the result improves, you have no idea which change did the work, and you cannot reproduce it tomorrow.
Keep a personal prompt file
When a prompt works, paste it into a plain text file with a one-line note about what the image looked like. After a month you will have your own private library of phrases that produce results you like. This beats any generic "best prompts" list online, because those were tuned for someone else's taste and someone else's model version.
The habits compound
None of this is complicated. It is just slower at the start and much faster later. Describe light, pick a medium, drop the empty praise, lock your framing, change one thing at a time, and write down what worked. Do that for two weeks and you will spend far less time generating and far more time choosing. Head back to the generator and try one prompt rewritten with these rules next to your old attempt. The gap tends to be obvious.