Image Style

A library of concrete styles for converting existing images. Each style keeps a reusable conversion prompt, traits observed in the user's references, and one example conversion.

Styles

Convert an image

1. Get the source image as a readable local path. If the user gave none, ask for it.
2. Match the request to a style in the table. If several fit or none clearly fits, show the options and let the user choose. Don't invent a style and present it as saved.
3. Read that style's STYLE.md, and look at its example.
4. Build the prompt from the style's prompt block. Add the user's specifics without dropping style rules, and add observed traits that matter for this image.
5. Convert with the image tool the user asked for or has available, using the source image as the edit target. With gpt-image, use --mode edit --edit-target <source>.
6. Compare the result with the source, the observed traits, and the example. If it drifts, name the difference and tighten the prompt for that trait.

Keep the source image's aspect ratio in the output. Only styles built on their own fixed layout, such as photo-fragment-collage, are exempt.

When the user's request conflicts with a saved style, follow the user for this image and leave the saved style unchanged.

Copyright caution

Some styles are named after artists, such as pablo-picasso and keith-haring, whose works may still be under copyright and whose names may be protected as trademarks. When a user converts an image with an artist-named style, mention once that commercial use of the result may need a rights check. Never copy a specific existing artwork, character, logo, or signature.

Add a style

1. Pick a short, descriptive kebab-case name based on the look. Don't name it after a subject that only appears in the references.
2. Create styles/<name>/STYLE.md from the structure below. Don't save the user's reference images in this skill.
3. Write the prompt in English, keeping every rule the user gave. Don't add rules to the prompt block on your own, except that the prompt must keep the source's aspect ratio unless the style is a fixed layout. If a reference shows a before/after pair, ask which part the user wants as output.
4. Look at the references, and record traits the prompt doesn't state under "Observed in references".
5. Generate a neutral source photo with no real names or logos, and convert it with the prompt using gpt-image. Save only the converted image in styles/<name>/examples/ as WebP at quality 80 (for example, cwebp -q 80 -m 6 in.png -o out.webp), and note any gap from the references.
6. Add a row to the table above.

# <Style name>

<One sentence on the look.>

## Prompt

<The reusable conversion prompt, in a code block.>

## Observed in references

<Visible traits the prompt doesn't state.>

## Examples

<The converted image, its source, and any gap from the references.>