If you've spent an hour typing prompts and still get pictures where the hands look like melted rubber, the problem usually isn't the tool — it's how the prompt is built. I've watched hundreds of people try AI image generation for the first time, and almost everyone makes the same five mistakes. Here they are, plus what to type instead.
Mistake 1: Writing a sentence instead of a description
"A woman walks in a park on a sunny day" is a sentence. An image model doesn't read it like a story — it reads it like a list of vague nouns. "Woman, late 30s, short curly hair, wearing a navy raincoat, walking alone along a wet stone path, overcast light, 85mm lens, shallow depth of field" gives the model something to actually paint.
Try this: write your prompt, then delete every word that doesn't describe something visible. If you can't picture it, neither can the model.
Mistake 2: Skipping the medium and style words
A portrait prompt with no style reference will default to the model's most average-looking output — usually a glossy stock-photo look. Adding "studio photograph," "oil painting," "35mm film scan," or "flat vector illustration" changes the result more than any other single phrase. The medium word acts like a global switch for texture, lighting, and contrast.
Mistake 3: Ignoring aspect ratio
Most people generate everything as a square because that's the default. A wide shot of a mountain range in 1:1 gets cropped into a middle strip of rock. A phone wallpaper in 16:9 gets letterboxed. Landscapes want 16:9 or 21:9. Posters want 2:3. Avatars and product shots want 1:1. Set the ratio before you hit generate, not after.
Mistake 4: One prompt, one attempt
The most common workflow error is judging a model on a single output. Every generation is a roll of the dice within a probability distribution. Four outputs from the same prompt will range from terrible to excellent. Generate a batch, pick the closest one, then refine from there.
Two techniques that help:
- Seed locking. If a model supports a fixed seed, reuse it while changing one variable at a time. You'll learn what each word actually does.
- Inpainting. When 90% of the image is right and one element is wrong, don't reroll — mask that area and describe only what should be there.
Mistake 5: Never using negative prompts
Every model has habits it falls into: extra fingers, watermarks, blurry backgrounds, over-saturated color. Negative prompts exist to push back. Useful ones to keep on standby: "text, watermark, extra limbs, duplicate face, distorted hands, low contrast, jpeg artifacts." Keep the list short — a 40-word negative prompt starts fighting itself.
A prompt template that works
Subject → action or pose → setting → lighting → camera or medium → style → aspect ratio. In practice that looks like:
"Elderly fisherman mending a net on a wooden dock, early morning fog, warm side light from the left, 50mm photograph, muted film grain, 3:2"
It isn't a magic formula, but it forces you to answer seven questions before you burn credits — and that alone puts you ahead of most users.
Stop stacking adjectives
"Hyper-realistic, ultra-detailed, 8K, masterpiece, award-winning, cinematic, stunning" is a placebo prompt. Most modern models were trained to treat a handful of quality tags as a mild style nudge at best, and an over-long pile of them does nothing except occupy attention that should be going to your subject. Pick one quality word — or none — and spend the rest of the prompt on specifics: lens, light direction, time of day, distance.
Related trap: negative-quality phrasing like "no blur, not ugly." Many models handle negation poorly and will happily produce a blurry image because the word "blur" was in the prompt. Put those terms in the negative field, not the main one.
Keep your own prompt log
Prompts that work are worth more than prompts that don't, and you will forget them. Keep a plain text file with three columns: the prompt, the model and settings, and one line on what was wrong with the output. After twenty entries you'll start to see which of your own habits repeat — usually it's something you'd never have noticed from a single image.
The log also makes comparison honest. When a new model comes out and everyone says it's better, you can run your five benchmark prompts against it and check on your own terms instead of trusting a launch thread.
What to do when the result is almost right
Almost-right images are the hardest to fix because people throw away the prompt entirely. Instead, keep a copy of the original, change exactly one block, and compare. If you changed lighting and the composition broke, you learned that your lighting word carries composition weight in that model. Ten of these experiments will teach you more than a hundred random prompts.
The tools themselves have gotten dramatically better at following instructions, which means prompt quality matters more, not less — a good model amplifies a good description and also amplifies a lazy one. You can test everything above for free on the homepage: try a text-to-image generation here and run the same prompt four times to see the variance yourself.