How to write AI image prompts: a simple formula

Updated 2026-10-13 · 5 minute read · works for Midjourney, Flux, Stable Diffusion, Gemini and ChatGPT

A good AI image prompt is not long or magic. It is a short list of clear decisions, written in a sensible order. This guide gives you a formula you can reuse in any image tool, plus before and after examples. If you want to skip the typing, the free prompt generator on this site builds the final wording for you from short slash commands.

The formula: subject, view, light, style, extras

1. Subject first. Say what the picture is of, in plain words: "a lone lighthouse on a cliff". Models give the most weight to the start of the prompt.
2. View and lens. Say how close or far the camera is and the angle: "wide 24mm field of view", "close-up", "low angle", "top-down".
3. Light and weather. Light changes the mood more than anything else: "golden hour, warm low-angle sunlight", "violent storm, dark clouds", "soft window light".
4. Style or medium. Photo, oil painting, 3D render, matte painting, watercolor.
5. Extras. Colors, detail level, and things to avoid such as text or watermarks.

Before and after

Plain idea:

a lone lighthouse on a cliff

The same idea, with the formula applied (Midjourney format):

a lone lighthouse on a cliff, wide 24mm field of view, environmental context with natural perspective, violent storm, dark roiling clouds, lightning, wind, dramatic lighting, strong contrast between light and shadow --ar 16:9 --no text, watermark, logo, signature, caption, username

The wording after the subject came from four short slash commands: /storm /dramaticlighting /24mm /16:9 plus /clean for the avoid list. You can try this exact request in the generator.

Each tool likes a slightly different format

• Midjourney reads comma-separated phrases. Settings go at the end: --ar 16:9 for the aspect ratio and --no text to leave things out. See the Midjourney prompt generator.
• Flux follows natural sentences well. See the Flux prompt generator.
• Stable Diffusion takes a prompt plus a separate negative prompt and a size. See the Stable Diffusion prompt generator.
• ChatGPT and DALL-E and Gemini understand ordinary sentences. See the ChatGPT and Gemini generators.
• Google Flow and Kling are video-oriented and use 16:9 or 9:16 frames. See the Google Flow and Kling generators.

Common mistakes

• Too many style words. Stacking ten modifiers makes models ignore most of them. Keep about eight at most and drop clashing ones, such as "black and white" with "vibrant colors".
• Naming equipment. Writing "200mm telephoto lens" can make some models draw a lens or camera in the picture. Describe the look instead: "200mm look, strong background compression".
• Hiding the subject. If the subject is buried in the middle, results drift. Keep it first.
• Fighting the format. Pasting Midjourney parameters into a sentence-based tool does nothing useful. Match the tool.
• Expecting an edit. A prompt creates a new picture. To change an existing picture, use an image-to-image or instruction-edit model.

A quick checklist

• Subject in the first few words
• One clear view or lens look
• One clear lighting idea
• One style or medium
• The right aspect ratio for where the image will be used
• An avoid list if you see unwanted text or watermarks