Generate
3D icons

One skill turns a general image model into a 3D icon generator with a fixed style: whatever subject you ask for, it comes back as a few large acrylic blocks on the same grid, from the same camera.

Image models can draw anything, which is exactly the problem for icon work — ask twice and you get two styles. voxel-icon is an agent skill for GPT Image 2 that closes that gap. It carries the style contract, the palette registry, and the calibration references, so your agent generates icons that sit next to each other like one family. This page walks through generating your own; the free set shows where it lands.

Add the skill to any coding agent that supports the skills format.

npx skills add BIAsia/voxel-icon

That single install ships everything the agent needs: the style contract for still icons, a 40-preset palette registry, two internally generated calibration images, and two post-processing scripts (Pillow + NumPy) that normalize every render to an exact #F4F4F4 background. Nothing else to configure — the skill is self-contained and public on GitHub.

Ask for a subject — a toaster, a crown, an hourglass — as a voxel icon. The skill handles the rest.

A prompt as plain as “generate a voxel icon of a toaster” is enough. The skill constrains the model to one isolated subject built from a small number of large rectilinear blocks, shot from a fixed orthographic three-quarter camera, in polished face-lit acrylic with broad whole-face highlights. The subject floats with barely any shadow, and the background comes out #F4F4F4 every time — that consistency is what makes a set of one-off generations read as one icon family.

Native keeps the object's own colors. Directed recolors the same geometry with one preset relationship.

By default every subject keeps its recognizable component colors — a red toolbox stays red. Ask for directed mode and the skill picks one complete relationship from its 40 bundled presets and recolors the approved geometry without touching its shape; every icon on this site ships in both variants. An aggressive mode goes further, mapping preset colors across the object's distinct real layers.

Motion is opt-in: a four-frame rest → forward → rest → back loop with real 3D pose changes.

Ask for a loop and the skill generates four complete keyframes — connections stay fixed, the pose actually moves in 3D — then the bundled scripts assemble and verify a GIF and WebP. No masks, no inpainting. The animated tiles on the wall are exactly these loops.