ChatGPT can generate and edit images from a web chat. Attach a reference when you want to guide an edit; availability and usage limits depend on your plan and workspace settings. OpenAI's image-generation documentation
Write the brief around the space the image must fill
Suppose you are making a square announcement for a fictional community repair workshop. You need an illustration for the lower part of the design and a quiet area above it for editable event text.
This original prompt states that arrangement without inventing an event date or asking the image to carry essential logistics:
Create a square illustration for a community repair workshop announcement.
Show a workbench with a small table lamp, a screwdriver, and a pair of
hands repairing the lamp. Use an original flat illustration style,
warm cream background, dark blue outlines, and muted orange accents.
Place the workbench in the lower half. Leave the upper third mostly
empty so I can add the workshop title in my design tool.
Show an unplugged lamp. Do not include words, numbers, logos, or
brand names. Keep the objects easy to distinguish at phone size.
The upper space serves a real layout need. The unplugged lamp is a visible detail to inspect. The color request gives a direction without borrowing a named artist's treatment.
This is a prompt example, not a claim that a generated image has already met the brief.
Inspect the first result against the intended use
Look at the entire composition before asking for more detail. Is there enough space for the title? Can you tell what the hands are doing? Does the image contain lettering you asked it to omit?
Then view it at the size readers will see. Small tools that are clear in an expanded image can disappear in a phone preview. If the repair action becomes hard to recognize, ask for fewer objects and a larger lamp.
For a real event, check any factual detail that the illustration communicates. Do not use an illustration of unsafe equipment handling as the visual instruction for a safety lesson. Choose a decorative scene or have the relevant person review it when accuracy affects the reader's actions.
Ask for one visible correction at a time
OpenAI recommends targeted revisions and specifying what should stay fixed. Keep your follow-up tied to the problem you can actually see. Refine a generated image
If the lamp is too small, this revision is more useful than “make it better”:
Edit this image so the lamp is about twice its current visual size.
Keep the workbench in the lower half and preserve the empty space
above it. Keep the cream background and blue-and-orange palette.
Do not add any lettering or extra tools.
Compare the new result with the previous one. Check the title space again, because a larger subject may fill the area you intended to reserve. An instruction to preserve a detail gives you a review criterion; inspect whether the edit followed it.
Put essential event details in editable text
For the repair-workshop example, add the final title, date, venue, and registration information in the tool where you assemble the announcement. You can then correct a room number without generating a replacement illustration.
If you do request lettering inside an image, provide the exact text and check every word in the result. Dense copy or typography that must be exact may need finishing in a design tool. OpenAI gives the same caution for production-critical text. Text and dense layouts in generated images
On a web page, keep the essential event information as readable page text too. Images of text and accessibility explains why the image alone should not carry the answer.
Finish with the announcement's real destination
Place the reviewed illustration into the actual announcement and check the crop with its title and details. A square image may still be cropped differently in the channel you use. Read the date, location, and sign-up instructions in the final preview before you send it.
Keep the approved illustration beside the editable announcement file so the next correction starts from the version you actually reviewed.