Codex Image Generation

scripts/codex-imagegen runs codex exec with Codex's built-in image_gen tool, which uses the user's ChatGPT login (no API key), and copies the result to a path you choose.

If you are Codex, do not use this skill. Use your built-in image_gen tool directly.

When to use

• Use it for a new image: a concept, illustration, photo-style shot, product or UI mockup, game asset, cover, or texture.
• Use it to edit an existing image: add, remove, or replace something, change the style or lighting, or cut out the background.
• Use it for a new image guided by reference images for style or composition.

Do not use it for diagrams, charts, flowcharts, simple icons, or anything that should stay editable as SVG/HTML/canvas/code. Build those directly.

Running it

The script is at scripts/codex-imagegen inside this skill's directory (the folder that contains this SKILL.md). Replace <skill-dir> below with that folder's absolute path.

<skill-dir>/scripts/codex-imagegen -o OUTPUT.png "PROMPT"
<skill-dir>/scripts/codex-imagegen -i input.png -o output-v2.png "Change only X; keep everything else unchanged"
<skill-dir>/scripts/codex-imagegen -i style.png -o new.png "Image 1 is a style reference. Draw ..."

Flags:

• -o sets the output path (.png, .jpg or .jpeg).
• -i attaches an input image. You can repeat it.
• -f overwrites an existing output file.

On success the script prints the absolute path of the saved image. On failure it exits non-zero and prints Codex's reason.

• Each image takes 30–90 seconds. Give the shell command a timeout of at least 300 seconds.
• Each call makes exactly one image. For several images or variants, make one call per image. Independent calls can run in parallel, each with its own -o.
• Do not overwrite existing assets unless the user asked for that. Write a sibling such as hero-v2.png.
• Use .png for transparent backgrounds. JPEG drops the alpha channel.
• Save project assets inside the project, at the path where the project will use them.

After generating, look at the image if you can view images. If it misses the request, make one targeted follow-up. You can pass the result back with -i and describe only the change. Tell the user the saved path.

Writing the prompt

• Order: scene/backdrop → subject → key details → constraints → intended use, such as "website hero" or "game sprite".
• Detailed requests: if the request is already specific, write it as a clean spec and don't invent extra elements. If it is vague, add only composition, lighting, and style cues that help.
• Photorealism: say photorealistic and name real texture, lens, and lighting.
• Text in the image: put the exact words in quotes, require verbatim rendering, and describe the font, size, color, and placement.
• Edits: say "change only X; keep Y unchanged". Repeat the invariants on every iteration so the image does not drift.
• Several input images: give each one a role, for example "Image 1: edit target; Image 2: style reference".
• Transparent cutouts: ask for "a genuinely transparent background, clean edges, no halo".
• Size and shape: describe the aspect ratio in words, such as "wide 16:9 landscape" or "square". There is no size flag.

More examples are in Codex's own guide: ~/.codex/skills/.system/imagegen/references/prompting.md and sample-prompts.md.