A useful image prompt describes what you want to see, how it should be arranged, and which details matter most. For an edit, say both what to change and what to preserve. Then review the result and make focused revisions. OpenAI’s image-prompting guidance recommends this as a starting approach—not a guarantee that every detail will appear exactly as requested.
A prompt template you can adapt
Create [image type and purpose] featuring [main subject]. Compose it with [placement, framing, foreground/background, and negative space]. Use [visual style, lighting, palette, and materials]. Include the exact text “[copy]” at [position], if needed. Use [aspect ratio or dimensions, if supported]. Avoid [specific unwanted elements].
Start with the image’s purpose and composition; add styling after the important content is clear. If the image needs words, provide the exact wording and where it belongs rather than asking for a generic “headline.” For edits, state what must stay unchanged as well as the one thing you want altered. Refine one thing at a time, then inspect the result.
Three copy-and-adapt prompt examples
These are illustrative prompts, not documented test results. Replace details with your own requirements, and check the generated image before using it.
1. Poster with exact text
Create a square poster for a neighborhood spring market. Show a small flower stall in the lower third, with shoppers in the background and clear open space across the top. Use a cheerful screen-print style, warm cream paper, leaf green, and coral red. Put the exact title “SPRING MARKET” at the top and “Saturday · 10 AM–2 PM” below it. Keep both lines centered and easy to read. No other text or logos.
Customize: event details, image shape, palette, and where the copy sits. Check every character and punctuation mark in the output.
2. Product image without required text
Create a portrait product photograph of a handmade blue ceramic mug on a pale oak table. Show the mug at a three-quarter angle, with its handle visible. Add soft morning window light from the left and a simple, softly blurred kitchen in the background. Keep the mug as the clear focal point; no people, labels, lettering, or extra cups.
Customize: the product’s true color, shape, material, viewing angle, and background. If the image represents a real item, compare those details with the actual product before publishing.
3. UI concept
Create a landscape visual concept for a simple online bookstore homepage. Show a top navigation bar, a large featured-book area, three book-cover cards, and a search field. Use a calm navy, ivory, and muted gold palette with a clean editorial feel. Keep the layout balanced and leave the book-cover designs generic. Do not include readable brand names or extra interface panels.
Customize: screen orientation, sections, palette, and level of detail. Treat the result as a visual concept—not working software or a verified interface specification.
How to refine the first result
- Generate a first version. Check whether the main subject, composition, and format are close enough to refine.
- Pick the most important mismatch. For example, the subject may be too small or the title may be in the wrong place.
- Request one focused change. Try: “Move the title to the upper-right corner; keep the subject and background unchanged.”
- Inspect the new image. Confirm the requested change happened, then check that important details—including any exact text—are still correct.
- Repeat if needed. Make another specific change rather than rewriting the entire prompt unless the basic concept is wrong.
A focused edit request is useful guidance, but it does not promise that unmentioned image details will remain untouched.
Using image prompts through an OpenAI API
OpenAI’s API guidance distinguishes between generating or editing a single image from one prompt with the Image API and building a conversational, multi-turn image workflow with the Responses API. Consult the current image-generation guide for the workflow and model it supports.
Production decision: the OpenAI GPT-Image-2 model page identifies gpt-image-2 and gpt-image-2-2026-04-21 as Deprecated. That page does not explain the deprecation, identify a replacement, or provide migration instructions. On that evidence, don’t choose GPT Image 2 for a new production integration. Confirm the currently supported image model and its matching request pattern in OpenAI’s live API documentation before building; the available information here is not enough to name a verified replacement or provide a runnable API example. API model names are also distinct from ChatGPT product names.
Quick checklist
- Did you state the subject and intended image type?
- Did you describe composition and placement—not just mood?
- Did you give exact text and placement if words must appear?
- Did you specify format or dimensions where supported?
- For an edit, did you say both what to change and what to preserve?
- Have you inspected the final image for incorrect text or other important details?
Bottom line: be specific about the image’s content and layout, make one revision at a time, and verify the result. For API work, confirm the currently supported model and matching request pattern in OpenAI’s documentation before integrating.