simonw/pedalican

Simon Willison stumbled into a feature he apparently missed when it launched back in May: Codex Desktop has "pets," little animated Clippy-style robots that pop up while you work. He accidentally triggered one, discovered you can build your own, and promptly created a pelican on a bicycle that now bounces around his screen giving him updates on his Codex tasks.

What caught his attention wasn't the cute animation itself but the process behind it. Willison told the system he wanted a pelican riding a bicycle, and GPT-5.6 Sol xhigh took it from there, running multiple rounds through gpt-image-2 to generate the sprite assets needed for the animation loops. He had it document every step along the way, and the resulting GitHub repo contains all the generated images, the combined sprite sheets, and GIFs for each animation, including one called waving.gif.

That waving animation, it turns out, was compiled from a single gpt-image-2 output. And that image traces back through a chain of prompts, starting with an initial reference-image prompt describing "a compact adorable baby pelican with a round cream-white body, soft coral-orange bill and feet, riding a tiny sky-blue bicycle," posed against a flat magenta chroma-key background sized for easy animation at 192x208 pixels.

Willison has been on the lookout for ways to use image generation for game-ready sprite work, so he dug into the underlying mechanism rather than just admiring the output. The core pieces are open source: two Apache 2.0 licensed skills, hatch-pet from openai/skills and imagegen from openai/codex, which together handle the sprite generation pipeline.

And one detail he seemed particularly pleased with: GPT-5.6 Sol came up with the name "Pedalican" on its own. Willison's verdict was simple approval, but the more telling part of the post is how openly he traced the entire generation chain, prompt by prompt, image by image, rather than treating the tool as a black box.