# Artwork direction and registration

Follow the user's requested subject and style. Without a reference or a different
style request, use the full-color ukiyo-e composition and colored sumi-e anime
linework described in `SKILL.md`: mineral colors, expressive ink contours, visible
brush texture, and a coherent illustrated environment.

## Holographic layers

Generate raster artwork with the image-generation tool and save it in the user's
project. Use one portrait canvas size for all four PNG files:

- `background.png`: a complete environmental painting. Keep the title and lower
  caption areas quiet; leave visual room for the foreground subject.
- `subject.png`: the isolated subject with genuine transparent alpha. Preserve
  the full silhouette and leave room for movement at the maximum tilt.
- `lineart.png`: dark contours on white, registered to the subject's canvas,
  position, scale, and silhouette. This is a mask, not another full-color painting.
  Ask the image tool to trace the subject without moving or reframing it.
- `text.png`: transparent typography generated by `generate_typography.py` from
  the config. Use an installed font with the required glyphs rather than asking
  the image model to draw exact titles and edition numbers.

Inspect real alpha and alignment. A painted checkerboard is not transparency,
and an image-generation edit is not a guarantee of pixel-perfect registration.
Check outlines in the assembled card, especially around eyes, fingers, and fins.
Keep foil and line glow light enough that the artwork remains readable.

## Lenticular image pair

Generate complete artwork A first, then edit A into B using it as the identity
and composition reference. A useful edit prompt is:

> Keep the same face structure, hair or fur, costume anchors, body proportions,
> camera, crop, subject scale, and major silhouette. Change only the requested
> state, expression, lighting, atmosphere, and state-specific effects. Preserve
> the same canvas dimensions. Do not split the image into left and right halves.

For a benevolent/malevolent pair, make the expressions distinguishable while
keeping the character recognizable. Compare facial anchors and silhouettes by
inspection; similar dimensions or a numerical similarity score alone do not
establish identity. See `lenticular-route.md` for pair preparation.