Sometimes Claude wants to over-engineer stuff that just needs iterative feedback. I mean, a use case of LLMs is to build it, and iterate faster with things that work, not try and spec the thing up from the ground. If it gets it wrong, it’s fine if it just has to start over.