Free · 7 min read
How to write prompts for AI image generation
How to write a prompt that gets you the image you need: prompt structure, precise words, text in images, references and iterations. With examples for designers.
- The parts of a prompt that work in any image generator
- Why "no cake" gets you a cake, and how to avoid it
- How to put text in an image without getting gibberish
- Free
- 1,504 members
- Join in 1 click
5 steps to a prompt that gets you the right image
- 1
Step 1
Describe the shot, not your wishes
- 2
Step 2
Choose precise words and numbers
- 3
Step 3
Say what you want, not what you don't
- 4
Step 4
Add text and format
- 5
Step 5
Work in iterations and with references
- 1,504
- community members
- 8,915
- messages with questions since July 2024
- 9
- design games since September 2024
- 759
- people wrote in the chat
This guide is for you if you are
Just starting with AI
You type "a nice picture for a website" and get something that fits no project at all.
Need visuals for layouts
You're looking for illustrations, backgrounds and mockups for learning case studies and want control over the result.
Trying different generators
You use Midjourney, Firefly, Gemini or ChatGPT and don't get why the same prompt works differently.
In short
- A good prompt describes a shot: subject, medium (photo, watercolor, 3D), environment, lighting, color, mood and composition.
- Precise words and numbers beat vague ones: "three cats" instead of "cats", "enormous" instead of "big".
- Describe what you want to see: mentioning "no cake" often adds a cake.
- Put text that should appear in the image in quotes and keep it short.
- Every generator has its habits: Midjourney likes short prompts, Firefly likes descriptive ones, so read the tips for your tool.
5 steps to a prompt that gets you the right image
- 1
Step 1
Describe the shot, not your wishes
The most common mistake is asking for a verdict instead of a description: "beautiful", "creative", "modern" tell the model nothing. Midjourney's documentation suggests thinking through several elements and naming the ones that matter to you:
- subject: who or what: a person, animal, object, place;
- medium: photo, illustration, watercolor, 3D render, sculpture;
- environment: where: in a studio, outdoors, underwater, in a city;
- lighting: soft, studio, neon, overcast;
- color and mood: pastel, monochrome; calm, energetic;
- composition: portrait, close-up, bird's-eye view.
💡 Example: instead of "a picture for a coffee shop landing page", try "photo of a cappuccino on a wooden table by the window, soft morning light, warm brown tones, cozy mood, close-up, lots of empty space on the left for a headline". The last detail is pure design thinking: you plan where the text will go right away.
- 2
Step 2
Choose precise words and numbers
Models are literal. Midjourney recommends replacing vague words with precise synonyms ("enormous", "gigantic" instead of "big") and stating quantities: "three cats" instead of "cats", "a flock of birds" instead of "birds". Adobe recommends at least three words for Firefly, in simple language, without "generate" or "create", adding feeling, style and lighting. See how a prompt changes:
- "icon" → "flat house icon, 2 colors, thick lines, white background";
- "people in an office" → "two people discussing a layout on a laptop, daylight, side view";
- "abstract background" → "soft gradient waves from purple to light blue, grain, minimalist".
💡 Short or long depends on the tool. Midjourney says short, simple prompts work best and long lists of instructions confuse it. Firefly, by contrast, recommends being descriptive. Start short and add detail only where the result misses.
Show your prompt together with the result
People in the community break down why a prompt worked or didn't and share wording that works. Come in, it's free.
Share your promptFree. 1,504 designers are already in the community.
- 3
Step 3
Say what you want, not what you don't
Ask for "a party with no cake" and a cake may well appear: the model notices the word "cake", not the negation. Midjourney explicitly recommends describing what you do want to see, and it has a separate parameter for exclusions. Rephrase bans as positive descriptions:
- "no people" → "an empty street at dawn";
- "not bright" → "muted pastel colors";
- "no clutter" → "a single object in the center, plain background".
💡 The same rule applies to video generators: Runway, for example, says negative phrasing isn't supported and may produce the opposite result. More in AI video generation: where to start.
- 4
Step 4
Add text and format
Modern generators can write in images: ChatGPT adds text on request, and Google says Gemini's image model renders text more accurately in all supported languages. Adobe's Firefly help recommends putting the words in double quotes. To keep lettering from turning into gibberish:
- put the text in quotes: a sign that says "Coffee to go";
- keep it short, one to three words, and add longer text in Figma later;
- name the format: "vertical story banner", "square", or pick the aspect ratio in settings; ChatGPT has a separate picker for this;
- if you need a transparent background for a layout, say so: ChatGPT can make the background transparent.
💡 Which language? Gemini and ChatGPT understand many languages. Google Flow, though, recommends English prompts for the best results and warns that quality in other languages may vary.
AI drawing the wrong thing? Show your prompt to the community and ask what to change → - 5
Step 5
Work in iterations and with references
The first result is almost never final. Change one thing at a time (lighting, angle or style) so you know what made the difference. When words aren't enough, add a reference image:
- Midjourney has Image Prompts and Style References: images that set the content and the style;
- in ChatGPT you can save a photo as a reference and call it in another chat with @name;
- in Gemini you can upload several images and ask to combine them into one;
- ChatGPT edits a specific part via the selection tool, but OpenAI warns selections aren't always precise and edits may spill beyond them.
💡 Finish the image by hand: Adobe itself recommends refining in Photoshop or Illustrator after generating. Upload only images you have the rights to; both Google and Adobe warn about this.
You're not alone with these questions
We counted what people ask about most in our community. AI tools are the fourth most popular topic, after learning, portfolios and Figma.
Messages by topic in Spaceberry Community, July 2024 — November 2026. Keyword search.
Practice with a deadline and live feedback
The best teacher is a task with a deadline that other people see. That's why the community regularly runs design games: the same brief for everyone, a deadline and a vote.
Design game · January 2026
NAOMA website redesign
The academy's home page through the eyes of an applicant, a student and a teacher. Three finalists; the community chose the winner.
AI challenge · May 2025
A landing page made by AI
Logo, copy and visuals made only with AI tools. Four entries, an anonymous vote, and participants shared their prompts.
Workshop · April 2026
Figmosha: AI in Figma
The tool's author showed how to generate interfaces in Figma and analyze a design system.
Free. 1,504 designers are already in the community.
How to join: 1 minute
- 1
Tap the button
Telegram opens with the Spaceberry community.
- 2
Join
It's free: all you need is a Telegram account.
- 3
Introduce yourself
Say where you're starting from and ask your first question.
Questions and answers
Didn't find your answer? Ask it in the community.
Which language should I write prompts in?
Gemini and ChatGPT support many languages, including Ukrainian. Some services explicitly recommend English: Google Flow, for example, says English prompts give the best results and quality in other languages may vary. A practical approach: if the model misses a detail, translate just that description and compare.
Is there a universal prompt that works everywhere?
No, but there's a universal structure: subject, medium, environment, lighting, color, mood, composition. Adapt length and syntax to the tool: Midjourney recommends short prompts with parameters at the end, Firefly descriptive sentences without the word "generate", Gemini starting with "draw" or "create". For specific generators see [Midjourney](/en/guides/midjourney-for-designers) and [Adobe Firefly](/en/guides/adobe-firefly).
Why does AI ignore my "no" and "without"?
The model reacts to the words it sees, so "no cars" may add a car. Midjourney recommends describing what you want to see and using a separate parameter for exclusions. Rephrase the ban: instead of "no people", write "an empty street".
How do I get correct text in an image?
Put the words in quotes, keep the text short and say where it goes: "on a sign", "on a book cover". Even models that write well slip on long phrases, so add headlines and paragraphs in Figma on top of a generated background.
Is the Spaceberry community paid?
No. Spaceberry Community is a free Telegram community supported by Spaceberry Foundation, a non-profit organization. All you need to join is a Telegram account. Inside are 1,504 designers who discuss each other's work, plus regular design games: the same brief for everyone, a deadline and a community vote. Conversations are in Ukrainian.
Write your first precise prompt
Join 1,504 designers who learn, share their work and play design games. Show your prompt and result and see how others phrase theirs. The community talks in Ukrainian.
- Free
- 1,504 members
- Join in 1 click
Updated · Spaceberry Foundation
Sources:
- Midjourney Docs — Prompt Basics
- Adobe Help — Writing effective text prompts (Firefly)
- Adobe Help — Generate Video FAQ (Firefly)
- Gemini Apps Help — Generate & edit images
- OpenAI Help Center — Images in ChatGPT
- Google Flow Help — Where you can use Google Flow
- OpenAI Help Center — How to change your language setting in ChatGPT
- Runway Help — Gen-4 Video Prompting Guide