CodeArt2D has the agent write game art as text: PixelSpec JSON for pixel sprites, portable SVG for flat vector props, UI and icons, rig files for characters with planted feet, FX specs, seam-checked autotiles, playable top-down layouts and stylized parallax. Python scripts render, check and package the result.
It is explicit-only. Image generation comes first for characters, props, maps and FX, and this skill steps in when you ask for code-drawn art or when no image or video route is available. Every result is disclosed as code-drawn.
What You Get
- Pixel sprites with palette variants rendered from a PixelSpec file
- Flat vector characters, props, UI and icons from portable SVG with palette classes
- Rigged characters that walk, idle and attack with planted feet, plus FX such as slashes, sparks and impact rings
- A codeart-meta.json that records the art source as code, and preview and review images to inspect
Common Use Cases
- Draw a small pixel gem or coin sprite with several palette variants
- Make flat vector icons and UI pieces for a game prototype
- Produce seam-proven autotiles or a playable layout when no image model is available
- Create pose and layout guides that help an image model
What Makes It Unique
It is built around clarity about its limits: a style envelope document says what code art does well, the skill refuses painterly or likeness art and licensed characters, and every output is labelled code-drawn. Tools publish only complete results, and strict QC publishes nothing on failure.
Under The Hood
The agent runs Python scripts from the skill folder through the shell: a media route check, an SVG render doctor, a PixelSpec renderer, an SVG renderer, a rig animator, an FX builder and a quality checker, with node used to verify FX runtimes. It needs numpy and Pillow, and for SVG one rasterizer such as resvg-py, which you install with pip. It reads your spec files and writes new output folders inside your project, refusing to overwrite an existing one. Drawing needs no image model, though the automated scan found network code in one of the scripts and invisible characters in another, so read the scripts before running them on a sensitive machine.