Did you ask an AI for “a cat in a garden” and get a bland, soulless image that looks nothing like what you had in mind? That’s normal. An image generator can’t read your mind: it interprets your words. And the vaguer they are, the more it fills in the blanks its own way. Knowing how to write a good AI image prompt is really about learning to describe what you see in your head, clearly and completely.
Good news: it doesn’t take any technical skills. In this guide, I give you a simple nine-ingredient structure, a step-by-step method and, above all, plenty of before/after examples you can copy, paste and adapt. These tips work with most tools: ChatGPT, Gemini, Adobe Firefly, Canva or Midjourney. If you’re a complete beginner, start with my guide on how to create an image with AI.
Key takeaways
- A good image prompt spells out the subject, action, setting, style, framing, lighting, colors, format and what to leave out.
- Write in natural language, in plain English, as if you were briefing an illustrator.
- To practice for free: ChatGPT, Gemini, Adobe Firefly and Canva all offer a free version with limits.
- Midjourney is paid, starting at $10 a month.
- Improve the image in small edits, one or two at a time.
Information checked on September 29, 2026.
Contents
- Why the Prompt Changes Everything
- The Structure of a Good Image Prompt
- The Step-by-Step Method
- Before/After Examples to Copy and Paste
- Choosing a Style: A Few Useful Phrases
- Improving an Image: Editing Prompts
- Which Tools Can You Practice With for Free?
- The Most Common Mistakes
- Limits to Keep in Mind
- My Checklist Before Generating
- Frequently Asked Questions
- The Bottom Line
Why the Prompt Changes Everything
The prompt is the description you give the AI. When it’s missing information, the AI decides for you: an “average” style, generic framing, neutral lighting. The result: images that all look the same.
Let’s take a very simple example.
Before:
A cat in a garden.
After:
Realistic photo of an orange tabby cat curled up asleep on an old wooden bench, in a country garden full of lavender. Close-up, seen slightly from the side. Golden late-afternoon light, blurred background. Warm tones, greens and purples. Horizontal format, no text.
The second version leaves almost nothing to chance. In my experience, this kind of description is what takes you from an “OK” image to one you actually want to use.
The Structure of a Good Image Prompt
Here’s the structure I recommend. You don’t have to use all nine ingredients every time, but the more you include, the closer the result gets to your vision.
| Ingredient | Question to ask yourself | Example |
|---|---|---|
| Specific subject | Who or what, exactly? | A baker in her sixties |
| Action | What are they doing? | Pulling loaves of bread out of the oven |
| Setting | Where does the scene take place? | In a small-town bakery |
| Style | Photo, illustration, watercolor? | Realistic photo, documentary style |
| Composition and framing | Wide shot, close-up, angle? | Medium shot, front view |
| Lighting | What kind of light? | Warm early-morning light |
| Colors | What palette? | Golden and brown tones |
| Format | Square, vertical, horizontal? | Vertical format |
| Elements to exclude | What don’t you want? | No text, no logo |
Put together, that gives you:
Realistic documentary-style photo of a baker in her sixties pulling loaves of bread out of the oven, in a small-town bakery. Medium shot, front view. Warm early-morning light. Golden and brown tones. Vertical format, no text or logo.
What the Tool Makers Recommend
The official advice points the same way. For Gemini, Google recommends starting with an action verb (draw, generate, create), specifying the visual style, then describing the subject, action and background in detail. For ChatGPT, OpenAI emphasizes natural language and detail, with the option to include text in the image, ask for a transparent background or start from an existing image to edit it.
The Step-by-Step Method
- Picture the image before you write: close your eyes for two seconds and jot down what you see.
- Start with a verb and a style: “Create a watercolor illustration of…”, “Generate a realistic photo of…”.
- Describe the subject and what it’s doing with concrete details (age, clothing, expression, material).
- Set the scene: place, season, time of day.
- Specify framing and format: close-up, wide shot, top-down view, vertical or horizontal.
- Add lighting and colors to set the mood.
- Say what you don’t want: text, logo, people in the background.
- Generate, look, then adjust in small edits.
Before/After Examples to Copy and Paste
Here are some real-world cases. Each time, the “before” version is what people tend to type on the fly, and the “after” version is what I recommend.
An Illustration for a Blog Post
Before:
An image about working from home.
After:
Create a clean, modern 2D illustration of a woman working on her laptop at a kitchen table, with a sleeping cat next to her and a cup of tea. Three-quarter view, medium shot. Natural morning light through the window. Pastel colors, soft blue and yellow. Horizontal format, no text.
A Flyer for an Event
Before:
School fair flyer.
After:
Create a cheerful retro poster-style illustration for a school carnival: kids playing a can toss game under strings of pennant flags, in a sunny schoolyard. Wide shot, seen from slightly above. Bright colors, red, yellow and blue. Vertical format, with a large empty space at the top to add a title. No text.
For this kind of use, I recommend adding the text afterward in a layout tool: I explain how in my Canva AI guide.
A Product Photo
Before:
Photo of a candle.
After:
Realistic catalog-style photo of a scented candle in an amber glass jar, on a white marble table with a few sprigs of eucalyptus. Close-up, front view angled slightly down. Soft, diffused light, light shadows. Cream, sage green and amber tones. Square format, no label or logo.
A Birthday Card
Before:
Birthday card for my mom.
After:
Watercolor illustration of a bouquet of pink peonies and white roses in a ceramic vase, on a wooden table with a small decorated cake. Centered composition, close shot. Soft, warm light. Dusty pink, white and soft green palette. Vertical format, no text.
A Social Media Visual
Before:
Instagram image about fitness.
After:
Realistic, dynamic photo of a man in his forties running on a coastal trail at sunrise, with the ocean in the background. Wide shot, side view, sense of motion. Low, orange light. Warm and blue tones. Square format, no text or brand logos.
A Children’s Illustration
Before:
A cute dragon.
After:
Children's book illustration, colored pencil style, of a little green dragon with big eyes reading a book under a giant mushroom in an enchanted forest. Medium shot, front view. Soft end-of-day light with fireflies. Gentle, warm colors. Horizontal format, no text.
A Wallpaper
Before:
Mountain wallpaper.
After:
Generate a digital illustration of snowy mountains at dusk, with a perfectly still lake reflecting the sky. Very wide shot, horizon on the lower third of the image. Purple-to-orange gradient sky, first stars appearing. Peaceful mood. Vertical format for a phone, no people or text.
Choosing a Style: A Few Useful Phrases
Style is often the ingredient that changes the result the most. Here are phrases that, in my experience, produce clearly distinct looks:
- Realistic photo, documentary style, catalog style, black-and-white photo.
- 2D illustration, clean flat illustration, retro poster style.
- Watercolor, colored pencils, ink drawing, oil painting.
- Soft 3D render, claymation, paper cutout.
- Comic book style, children’s book style.
I recommend naming a style rather than a specific artist: it’s clearer for the AI, and you avoid copying a real person’s work.
Improving an Image: Editing Prompts
The first image is rarely perfect. Instead of rewriting everything, ask for specific adjustments. Here are phrasings that work well:
Keep exactly the same scene and style, but change the sweater color to forest green.
Remove the mug from the table, don't change anything else.
Make the lighting warmer, like late afternoon.
Move the character to the left to leave empty space on the right.
Same image, but in vertical format.
Golden rule: one or two changes at a time, and always say what should stay the same. In both ChatGPT and Gemini, you can also upload an existing image and ask for edits. Heads-up: in the US, you need to be 13 to use Gemini with your own Google account, and image editing is limited to users 18 and older.
Which Tools Can You Practice With for Free?
To practice without paying, four tools offer a free version, with limits:
- ChatGPT: create and edit images in the conversation, in natural language. See my guide to creating images with ChatGPT.
- Gemini: generate images by simply describing the scene. For more generations, Google AI plans start at $9.99 a month.
- Adobe Firefly: a limited number of free generations every day.
- Canva: up to 20 AI uses on the free plan, handy for adding text afterward.
What About Midjourney?
Midjourney is better suited to more advanced users: it’s paid, starting at $10 a month, and available from age 13. Its V8.2 version follows prompts better and offers a new editing tool, the Edit Model. Dedicated parameters are still useful for setting the aspect ratio, stylization or variations. I cover it in detail in my Midjourney guide.
The Most Common Mistakes
- Being too vague: “a landscape” leaves all the work to the AI.
- Dumping everything in a pile: a list of unconnected keywords often leads to incoherent images. Write sentences.
- Contradicting yourself: “minimalist and highly detailed,” “nighttime lighting in broad daylight.”
- Overloading the scene: five characters, three animals and ten objects is a lot. Start simple.
- Forgetting the format: cropping afterward means losing part of the image.
- Asking for long text in the image: a few short words often work, a paragraph much less so. Add it afterward instead.
- Changing everything on every try: you lose track of what improved or worsened the result.
Limits to Keep in Mind
Even with a great prompt, AI can’t do everything. Hands, reflections, text and very fine details can still contain errors: always view the image full size before using it.
As for the rules, avoid depicting real people without their consent, copying brands, logos or protected characters, and reread the tool’s terms before any commercial use. Each service has its own rules, and a prompt doesn’t exempt you from following them.
My Checklist Before Generating
- I have a specific subject, with concrete details.
- I described the action and the setting.
- I picked a clear style.
- I specified the framing and format.
- I specified the lighting and colors.
- I said what I don’t want (text, logo, distracting elements).
- I’m ready to edit in small steps.
Frequently Asked Questions
Do you need special keywords or technical jargon?
No. The main tools understand everyday language very well. In my experience, a clear prompt in plain English gets good results; what matters is precision, not jargon.
How long should a good prompt be?
There’s no strict rule. Two to five sentences are often enough to cover the subject, style, framing, lighting, colors and format.
Can you reuse the same prompt in several tools?
Yes, a natural-language description works in most tools. The results will differ, since each model has its own signature look. Midjourney also offers dedicated parameters.
How do you get readable text in the image?
Put the exact text in quotation marks and keep it short. ChatGPT can put text into an image, but I recommend checking the spelling and, for longer text, adding it afterward in a tool like Canva.
Why doesn’t my image look like what I asked for?
Often because an ingredient is missing or there’s a contradiction. Reread your prompt against the nine-ingredient table, then adjust one thing at a time.
Can you practice for free?
Yes, with ChatGPT, Gemini, Adobe Firefly or Canva, within their free quotas. Midjourney, on the other hand, requires a subscription.
The Bottom Line
To write a good AI image prompt, describe the subject, action, setting, style, framing, lighting, colors and format, then say what you don’t want. Write in full sentences, in plain English, and improve the image in small edits. Practice with the free tools before moving on to Midjourney, if you want to.
To go further, also check out how to write a good ChatGPT prompt.
