How to Write Better AI Image Prompts
You typed "a cat in a garden" and got a cat in a garden. Technically correct, and nothing like the picture in your head. Image generators fill every gap you leave with the most average choice. The fix is not a longer prompt. It is a prompt that answers the four questions the generator would otherwise guess.
The four parts every image prompt needs
Subject — what is in the picture, and what it is doing. "An old fisherman mending a net" beats "a man". One clear subject usually works better than three competing ones.
Style — what it should look like. A photo, a watercolour, a 3D render, a flat vector illustration, an oil painting. Without a style, most generators default to a glossy, over-detailed look.
Mood — how it should feel. Lighting and colour carry most of it: "soft morning light, muted blues" reads very differently from "harsh neon, deep shadows".
Shape — the aspect ratio. A square, a tall phone wallpaper, a wide banner. Setting it up front saves cropping away half the image later.
Put together: "An old fisherman mending a net on a wooden dock, documentary photo, soft morning light with muted blues, wide 16:9." Four short answers, and the generator has far less to guess.
Words that actually change the result
Lighting words do more than almost anything else: golden hour, overcast, rim light, studio softbox, candlelight. If a picture looks flat, the lighting is usually unspecified. PromptItIn keeps a free list of lighting prompt keywords with what each one does.
Camera words control framing: close-up, wide shot, low angle, overhead, shallow depth of field. They work even for illustrations. The camera angle reference shows each angle side by side.
Style words are where most prompts go vague. "Artistic" means nothing to a model; "risograph print", "ukiyo-e woodblock" or "1970s film photo" mean a lot. Browse art style keywords when you want a look you cannot name.
Telling the generator what to leave out
Stable Diffusion and SDXL accept a separate negative prompt, and Midjourney has the --no parameter. Use them for the things that keep sneaking in: extra fingers, text, watermarks, blur.
ChatGPT, Gemini and Flux work differently. They respond better to describing what you want ("clean empty background") than to a list of what you do not. The negative prompt generator writes both forms for the model you use.
Iterate one change at a time
When a result is close but not right, change one part of the prompt and run it again. Change the subject, style and lighting together and you cannot tell which change helped.
Keep the prompts that worked. A prompt that gave you a good product shot will give you another with a new subject swapped in. Saving them somewhere searchable beats scrolling back through old chats.
Build your image prompt in a few taps
Pick subject, style, mood and shape. PromptItIn writes the full prompt for your generator.
Open the prompt builderGood AI images come from prompts that leave little to guess: a clear subject, a named style, a mood carried by light and colour, and the right shape from the start. Once the image is made, ImageSmith can crop, resize or compress it for wherever it is going — all in your browser.