Leon Lin shares his AI demo workflow and releases Cook, an agent skill for visual builds

Leon Lin has shared the workflow behind his AI-built visual demos, including Athenaion, a Greek-inspired 3D world. His advice starts with strong visual references, continues through repeated comparisons and ends with a new agent skill called Cook.

In his post, Lin says he often starts major demos with 9–12 reference images and a short prompt. For 3D projects he generates reference images, then asks a coding model to build toward them. He recommends placing the result next to the references because weaknesses become easier to see.

Set a target and keep comparing

Lin uses numerical targets to communicate how closely he wants the result to match. He acknowledges that figures such as 99.9% are aspirational, not a measured guarantee. He also gives specific feedback on color, timing and distant details, and asks fresh-context agents to critique the work.

For video, he renders still frames first and reviews them before committing to the final render. He says Athenaion took about 45 hours of work and its final 4K video render about 13 hours of GPU time. His $1,800–$2,300 figure is an estimate at API prices; he says the work ran on his Max plan.

Cook packages the workflow

The Cook repository describes a visual-build skill that uses references, iterative stills and independent critics. The project is available for Claude Code and Codex. Lin also links Earshot, a sound-effects tool he helped build. These are the creator’s methods and examples; the reported timings and results are not independent benchmarks.