The prompts for this are pretty sparse. This could 100% be accomplished with better prompting. Even with the current prompts, it's likely I could complete the task with a follow up request specifying what it did correctly and incorrectly. In fact, this could probably be entirely automated with multiple agents checking each other.