This pretty much matches my preferred workflow for scripting with LLMs: don't ask it to do the task, ask it to write a script that does the task, then validate and execute or refine.