How to Write AI Image Prompts That Actually Work (Beginner Guide)
You typed "a beautiful girl in a garden" and got something bland, plastic and forgettable. The tool isn't broken. The prompt is just missing the details that the model needs to make choices for you.
In this guide
Why most prompts fail
An AI image generator turns your words into a picture by predicting what an image matching those words usually looks like. When the prompt is vague, the model falls back on the most average, most common version of the idea. That is why "a dog" gives you a generic stock-photo dog, while "a muddy golden retriever puppy shaking water off on a rainy porch, shot on a 50mm lens, soft overcast light" gives you something with character.
The most common reasons a prompt produces weak results are:
- Abstract words with nothing concrete. "Beautiful", "amazing" and "epic" don't describe anything the model can draw.
- Too many competing ideas. Five subjects and three styles in one sentence confuse the composition.
- No lighting or framing. Light and camera position change an image more than almost any other word.
- Contradictions. "Minimalist, highly detailed, cluttered market" asks for opposites.
The 6-part prompt formula
You don't need secret words. You need a checklist. Write your prompt in this order and you will cover what the model needs:
- Subject – who or what is the focus. Add one or two defining details: age, material, colour, clothing.
- Action and setting – what the subject is doing and where.
- Style or medium – photograph, watercolour, 3D render, anime, oil painting, pencil sketch, flat vector.
- Lighting – golden hour, soft window light, neon, studio softbox, candlelight, overcast.
- Camera and composition – close-up, wide shot, top-down, low angle, shallow depth of field, rule of thirds.
- Mood and colour palette – calm, mysterious, joyful; pastel, muted earth tones, teal and orange.
Here is the formula filled in:
An elderly fisherman mending a net (subject), sitting on a wooden pier at dawn (setting), documentary photograph (style), warm low sunlight with long shadows (lighting), medium close-up, shallow depth of field (camera), calm and nostalgic, muted blues and warm skin tones (mood)You can drop the brackets, of course. The point is that each part does a job.
Before and after examples
Seeing the difference is the fastest way to learn. Try these pairs in any generator.
Weak: a cozy room
Better: a small reading nook with a green velvet armchair, stacked books and a sleeping cat, rainy window behind, warm lamp light, wide shot, soft focus background, cozy autumn mood
Weak: cyberpunk city
Better: a narrow street in a futuristic Asian megacity at night, glowing neon signs reflected in puddles, a lone figure with a transparent umbrella, cinematic wide angle, teal and magenta palette, light rain
Weak: logo for coffee shop
Better: minimal flat vector logo for a coffee shop called "Bean There", a simple coffee cup with steam forming a leaf, two colours, dark brown and cream, white background, centered
Notice that the better versions are not longer for the sake of it. Every added phrase answers a question the model would otherwise answer with a guess.
How different tools read prompts
The formula works everywhere, but each tool has its own habits.
- Midjourney responds well to short, comma-separated phrases and style references. Parameters such as aspect ratio are added at the end (for example
--ar 16:9). - ChatGPT image generation understands full natural sentences very well. You can describe the scene like you are talking to an illustrator, then ask for changes in follow-up messages ("make the lighting warmer", "remove the second person").
- Google Gemini also handles conversational descriptions and edits. Clear, literal descriptions of position ("on the left", "in the background") help.
- Stable Diffusion-based tools often support negative prompts, a separate box where you list things you don't want, such as "blurry, extra fingers, watermark".
Features change often, so check each tool's own help page when something behaves differently from what you expect.
Words that change the result the most
Some words carry more weight than others because they change the whole look of an image:
- Medium words: "35mm film photograph", "watercolor on textured paper", "clay render", "ukiyo-e woodblock print".
- Lighting words: "rim light", "backlit", "volumetric fog", "hard noon sun", "blue hour".
- Lens words: "macro", "85mm portrait", "fisheye", "tilt-shift", "drone shot".
- Composition words: "symmetrical", "negative space", "over-the-shoulder", "bird's-eye view".
If you only add one thing to your prompts this week, add lighting. It is the quickest upgrade.
A simple refining workflow
Great images rarely come from the first attempt. Use this loop:
- Write a prompt with the six parts.
- Generate 2–4 versions.
- Pick the closest one and decide the single biggest problem (wrong angle, too dark, wrong style).
- Change only the part of the prompt that controls that problem.
- Repeat until you're happy, and save the final prompt so you can reuse it.
Changing everything at once makes it impossible to learn which words matter. Changing one part teaches you how your chosen tool thinks.
Common mistakes and fixes
- Faces look plastic: add "natural skin texture, candid photo, soft daylight" and remove "perfect", "flawless".
- Busy, messy images: reduce to one subject, add "simple background" or "negative space".
- Wrong number of people or objects: state the count clearly ("two children, one dog").
- Text comes out garbled: keep text very short, put it in quotes, and expect to fix it in an editor if the tool struggles with lettering.
- Everything looks the same: swap the medium word. Moving from "photograph" to "gouache illustration" changes everything.
Building your own prompt library
Professionals keep a collection of prompts that worked. You can do the same in a notes app, but it gets messy fast. Group prompts by category (portraits, products, landscapes, social media) and keep the exact wording, including the lighting and camera parts.
If you want a head start, the AI Image Prompts Android app has a ready, categorised library of prompts for portraits, anime, product photography, food, logos, cinematic scenes and more. You can search, copy a prompt with one tap, save favourites, and paste it into whichever AI image tool you use. The app itself doesn't generate images; it's a prompt library you take to your generator. Use the prompts as starting points and then apply the refining loop above to make them your own.
Respect copyright and people
Avoid prompts that copy a living artist's name to imitate their work, and don't create realistic images of real people without consent. Most platforms have content rules; read them, especially before using images commercially.
FAQ
How long should an AI image prompt be?
Usually one to three sentences or 20–60 words is enough. Longer prompts are fine if every phrase adds a specific detail, but repetition and contradictions make results worse.
Do I need special keywords like "8K" or "masterpiece"?
Not really. Newer models respond more to concrete descriptions of subject, light and composition than to quality buzzwords. Test with and without them and keep what helps.
Why do I get different images from the same prompt?
Generators add randomness each time. That variety is useful: generate several versions and refine the best one. Some tools let you fix a "seed" to repeat results.
Can I use the same prompt in Midjourney and ChatGPT?
Yes, the core description transfers well. You may need to remove tool-specific parameters such as --ar and rewrite short phrases into full sentences for conversational tools.
Where can I find good prompts to start with?
Prompt libraries, community galleries and apps such as AI Image Prompts give you tested starting points. Always adjust them to your own idea.