For art assets, I would not be able to express the exact details of what it should look like up front to the LLM; I'd either need to implement it manually, or I'd need to let the LLM decide the details of the actual user-facing experience rather than just the implementation of the experience I've defined myself (and manually tested to ensure it's identical). That seems like a fundamental difference in terms of code and art assets; the actual user-facing experience of the output of code is further downstream of the way it's produced than the user-facing experience of the output of art assets.