AI Image Prompts: How to Write Them for Midjourney, OpenAI Image Models and Stable Diffusion
How to structure AI image prompts, what each major generator documents about prompts and parameters, known limitations, and illustrative prompt and expected-result pairs you can test.
Image generators respond to the same basic idea as text models: say clearly what you want. But they differ a lot in syntax, settings and failure modes, and those details change between releases. This guide covers the prompt structure that works across tools, what each vendor's own documentation says (checked 2026-11-05), and a set of illustrative prompts with descriptions of what a good result should contain.
A note on the examples: the prompt and expected-result pairs below were written for this article. They describe what to look for in a result; they are not screenshots or records of actual generations, and results vary by model, settings and run.
A Prompt Structure That Works Across Tools
Most good image prompts answer the same questions, in roughly this order:
| Element | Question it answers | Example |
|---|---|---|
| Subject | What is the main thing in the image? | A ceramic coffee mug |
| Details | What makes this subject specific? | Matte sage-green glaze, small chip on the rim |
| Setting | Where is it? | On a worn oak table by a rainy window |
| Composition | How is it framed? | Close-up, mug in the left third, shallow depth of field |
| Lighting | What is the light like? | Soft, overcast daylight from the window |
| Medium / style | What kind of image is it? | Documentary-style photograph; or flat vector illustration |
| Mood | What should it feel like? | Quiet, unhurried morning |
| Exclusions | What must not appear? | No text, no logos, no people |
You rarely need every row. Start with subject, setting, composition and style; add the rest when a result misses.
Describe instead of stacking "quality" tokens
Older prompt habits relied on strings like "masterpiece, best quality, 8k, ultra detailed". Current tools generally respond better to concrete description of what you want to see. If a result looks generic, add specifics (materials, light direction, lens-like framing, era) rather than more superlatives.
Describe styles rather than naming living artists
Naming a living artist to imitate raises ethical concerns, and some platforms restrict it. You usually get more control by describing the qualities you want: "loose ink linework with watercolor washes and lots of white space" tells the model more than a name does.
What the Vendors Document
Capabilities, model names and parameters change frequently. The points below come from each vendor's official documentation as checked on 2026-11-05. Check the linked pages before relying on a specific setting.
OpenAI image generation (API)
OpenAI's image generation guide (checked 2026-11-05) documents:
- Generating images from text and editing existing images, including edits guided by a mask. The guide notes that masking is prompt-based and that the model "may not follow its exact shape with complete precision".
- Options for output size (including square, landscape and portrait sizes), quality level, and background (including transparent backgrounds).
- A limitations section that lists text rendering ("can still struggle with precise text placement and clarity"), consistency for "recurring characters or brand elements", and composition control for "structured or layout-sensitive compositions".
- Content filtering: "All prompts and generated images are filtered in accordance with our content policy."
Practical implication: write prompts as natural-language descriptions, put any exact text in quotation marks, and check text, logos and layouts carefully because those are the documented weak spots.
Stability AI (Stable Image API)
The Stable Image Core API reference (checked 2026-11-05) documents:
- A required
promptand optionalnegative_prompt,aspect_ratio,seed,style_presetandoutput_format. - Aspect ratios including
16:9,1:1,9:16,3:2and2:3(with a default of1:1). - Style presets including
photographic,digital-art,anime,line-art,cinematicand others. - Word weighting in the prompt using
(word:weight), where the weight is a value between 0 and 1. The docs' example isThe sky was a crisp (blue:0.3) and (green:0.8). - A
seedvalue; omitting it or passing0uses a random seed.
Local Stable Diffusion interfaces (community tools you run yourself) have their own weighting syntax and conventions, which can differ from the API. Check the documentation for the interface you use.
Midjourney
Midjourney's documentation has pages for parameters including aspect ratio (--ar), exclusions (--no), seeds (--seed) and stylize (--stylize). We confirmed these pages exist via search on 2026-11-05, but Midjourney's documentation site blocked automated fetching, so we could not verify current default values, ranges or version numbers. Check Midjourney's Parameter List before using a specific value.
Illustrative Prompt and Expected-Result Pairs
Each pair below has a prompt, a description of what a successful result should contain, and what to check. Use the checks to decide whether to keep a result or revise the prompt.
1. Product photo on a plain background
Prompt:
Studio product photograph of a matte black stainless-steel water bottle
standing upright on a seamless light-grey background. Soft key light from
the upper left, gentle shadow to the lower right. Bottle centered, filling
about two-thirds of the frame height. No text, no logos, no props.
What a good result contains: one upright bottle, centered, on an even grey background; a soft shadow consistent with light from the upper left; a matte (not glossy) surface.
What to check: invented logos or text on the bottle; a second bottle or props appearing; reflections that suggest a glossy finish; the bottle's cap and proportions looking physically plausible.
2. Editorial illustration for a blog header
Prompt:
Flat vector illustration for a blog header about remote team meetings.
Three people on separate laptop screens arranged side by side, each in a
different home setting (kitchen, bookshelf, plant-filled corner).
Limited palette: navy, warm orange and off-white. Wide landscape framing
with empty space on the right for a headline. No text in the image.
Add a wide aspect ratio using your tool's setting (for example, 16:9 where supported).
What a good result contains: three distinct screens and settings; only the three named colors plus neutrals; clear empty space on the right.
What to check: extra or merged faces; garbled pseudo-text on screens or books; the palette drifting toward other colors; whether the empty space is actually usable for a headline.
3. Poster with exact text
Prompt:
Minimal event poster. Large headline text reading exactly "MAKERS MARKET"
in a bold condensed sans-serif at the top. Below it, a simple line
drawing of a canvas tote bag. Bottom line of smaller text reading exactly
"Saturday 10am-4pm". Cream background, black and tomato-red only.
What a good result contains: both text strings spelled exactly as written, in the stated positions; a simple line drawing; the three-color palette.
What to check: spelling letter by letter (the documented weak spot for text rendering); extra invented text; hyphen and time formatting. For anything published, it's usually more reliable to generate the image without text and add typography in a design tool.
4. Character reused across images
Prompt (first image):
Children's book illustration in soft colored pencil. A small red fox
wearing a mustard-yellow scarf and round green glasses, sitting on a
tree stump reading a book. Autumn forest background, warm afternoon light.
Follow-up prompt (second image):
Same fox character: small red fox, mustard-yellow scarf, round green
glasses, soft colored-pencil style. Now walking along a riverbank
carrying a wicker basket. Same warm afternoon light.
What a good result contains: a recognizably similar character across both images, in the same style.
What to check: scarf or glasses color drifting; glasses disappearing; the style shifting. Consistency for recurring characters is a documented limitation for OpenAI's models, and is a common challenge across tools. Repeat the full character description in every prompt, and where your tool supports reference images or seeds, use them.
5. Using a negative prompt (where supported)
Prompt:
Wide landscape photograph of a coastal lighthouse on a rocky headland at
dusk, calm sea, a few low clouds lit pink from below.
Negative prompt (Stable Image API negative_prompt, or your tool's equivalent):
people, boats, birds, text, watermark
What a good result contains: a single lighthouse on rocks, calm water, dusk lighting, and none of the excluded elements.
What to check: excluded items appearing anyway (exclusions reduce the chance, they don't guarantee it); the sky color matching "dusk".
A Simple Iteration Loop
- Generate several variations from one prompt before changing anything. Variation between runs is normal.
- Write down what's wrong in specific terms ("bottle is glossy", "text misspelled") rather than "looks off".
- Change one thing at a time in the prompt or settings, so you know what fixed it.
- Fix the seed if your tool supports it while you test wording, so differences come from the prompt rather than randomness.
- Save the prompt and settings that worked, including model and aspect ratio, so you can reproduce them.
Common Mistakes
- Conflicting styles. "Photorealistic cartoon" forces the model to guess. Pick one, or describe the blend precisely.
- Too many subjects. Each extra subject competes for the composition. Start with one clear subject and add elements gradually.
- Relying on generated text. Check every letter, or add text afterwards.
- Assuming one tool's syntax works in another. Midjourney-style
--parameters, API fields and local-UI weighting syntax are not interchangeable. - Skipping rights and policy checks. See the FAQ below.
Frequently Asked Questions
Which AI image generator is best?
It depends on the job, and models change often enough that any ranking goes stale quickly. Test the same two or three prompts from your real use case across the tools you're considering, and judge against the checks above.
Can I use AI-generated images commercially?
That depends on each provider's current terms, your plan, and the law where you operate, and all three can change. Read the provider's terms for your plan before commercial use, and be cautious with images that resemble real people, trademarks or a specific artist's work.
How do I get consistent characters?
Repeat a full, identical character description in every prompt, keep style wording identical, and use reference-image or seed features where your tool offers them. Expect some drift; vendors list consistency as a limitation.
Should I use negative prompts?
Where the tool supports them (for example, the Stable Image API's negative_prompt), they're useful for removing recurring unwanted elements. Keep them short and specific. In tools without a separate field, describe the scene you want positively.
Templates
To start from a structured template rather than a blank prompt, try the DALL-E 3 prompt creator, Stable Diffusion prompt master, logo design prompt generator or Midjourney prompt generator. Check any parameter they include against your tool's current documentation.
Related Articles
AGENTS.md vs CLAUDE.md: Which AI Coding Tools Read Which Instruction File
A cross-tool guide to AGENTS.md: how Cursor, Claude Code, Gemini CLI, Kiro, Devin Desktop (Windsurf), GitHub Copilot and Antigravity read it, how it relates to CLAUDE.md, GEMINI.md and tool-specific rules, and a shared-file pattern with caveats.
How to Create Claude Code Skills: A SKILL.md Guide with a Worked Example
Create an Agent Skill for Claude Code: directory layout, SKILL.md frontmatter rules, supporting files and scripts, where Claude Code discovers skills, how they are invoked, failure handling, and how skills differ from MCP servers.
How to Evaluate AI Prompts: A Repeatable Regression Workflow
A step-by-step method for evaluating AI prompts: build a test set, write a rubric, combine deterministic and model-graded checks, record model settings, read failures and iterate without regressions.