I'm sure there are many ways to implement this in some more automated form. Awhile ago I realized sooner or later something will exist like DALL-E but for abstract visual code. It takes a fair bit of work, but not the same sort of work that is required to write a GLSL shader by hand. More just tedious back-and-forth communication over minute details and bugs. Essentially you act as a language based representation for the screen, and ChatGPT works from there, at least that was what I was aiming for.