Create images with Codex: from your first request to editing and saving

Codex image work uses the image-generation capabilities currently available. Naming a skill alone does not create a tool or connection. If the feature is absent, check what the environment supports and needs.

Example: ‘Check which image creation and editing capabilities are available in this task. If possible, I would like a guide cover. Tell me if anything needs installation or connection.’

The official guide describes natural-language requests and selecting $imagegen. Button locations and attachments may differ across app, CLI and extension.

Say where the image will appear: a guide cover, vertical poster or small interface icon needs a different composition. Then describe the subject, background, colors and whitespace.

Example: ‘Create a horizontal cover illustration for a beginner file-organization guide. Place document folders and a laptop on a bright desk; leave space on the left for a title. Use blue and cream, and put no text inside the image.’

Include exact pixel dimensions when needed, then check the output file's actual size. Specific object placement gives you a better revision criterion than ‘Make it premium.’

To edit an existing picture, attach it or provide a path readable in the working environment. If ‘the earlier image’ is ambiguous among several candidates, identify the filename or attachment again.

Distinguish references: ‘The first image is the original to edit; use the second only for colors.’ For a product, state which shape, logo and colors must be preserved.

Image analysis and image generation or editing can be separate capabilities. Reading an image does not prove an edited result was saved; inspect the actual output.

Separate what you like from what should change. Begin with an identifiable target and position, such as ‘Reduce only the folder on the right,’ rather than redescribing everything. Check the parts meant to stay unchanged too.

Example: ‘Edit the first attached version. Lower the right folder so more of the laptop screen is visible. Preserve the desk position, blue-and-cream palette and left whitespace. Add no other objects or text.’

If you made several candidates, continue from the selected file. Referring to a different candidate at each edit can change a composition or palette you liked.

Check that you can open or download the generated image. For a local project, specify the destination and filename. If the tool returns another location, verify the real path before asking it to move the file.

Example: ‘Save the selected image to public/images/file-guide-cover.png in this project. If that name exists, do not overwrite it; suggest a new one. Tell me the saved path and actual dimensions.’

A path in a response is not proof that a file exists. Open it and, if placed on a web page, inspect that page too. Write alternative text describing the image.

A small preview may hide awkward text or shapes. Compare Korean text, numbers and product names with the source, and verify that required content is not cropped in the actual interface. Do not describe an illustration as a real product screenshot.

If you requested transparency, verify actual transparent pixels; a painted checkerboard is not transparency. For several covers, compare dimensions, whitespace and colors side by side.

Keeping the final file and request can help with the next image. Provide the finished image as a reference for a consistent mood and describe only the subject that should change.