What they did
The author, enoch3712 on GitHub, had a web game whose hero was drawn as a flipbook: 10 clips and 403 frames taken from generated videos, with weapon and armour baked into every frame. They rebuilt it from a single side-view painting, cut into parts on a Spine 4.2 skeleton played with the PixiJS runtime. An orchestrating model wrote one brief per move and started a separate run for each, so moves were built in parallel. For each move a coding model wrote keyframes and IK on the rig, built it, captured frames and looked at its own result. A reference clip from an image-to-video model in Google Flow gave timing only. The Codex CLI image tool painted parts the painting never showed, such as a far arm and new weapons. A plain Node script then measured every capture for sliding feet, cracks and moves that do not return to rest, and a second model checked 2x close-ups against a list of defects a human had caught before. The models named are Claude Opus 5.5 and Sonnet 5.5, with Fable 5.2 for planning.
How it went
The project ended with six moves on one skeleton, including a shield bash, a seismic slam and a leap. Because the gear sits on its own bones, a hammer, axe or mace and different shields can be swapped and every move still plays, where the old design needed every clip redone. The author packaged the method as an agent skill for Claude Code, Codex and Cursor. The human still has the final say on each move, and each flaw they catch becomes a new gate check or defect entry.
Worth knowing
The rig building and frame capture scripts from the original game are not in the repo, so you must bring your own way to cut the painting into a rig and to write the metrics file the gate reads. The gate limits were calibrated on one game. The README names the models but not the harness that ran the Claude roles. Some plugin install paths for Codex and Cursor are marked untested.