Yeah that's fair, I've been there before myself. It doesn't help when it throws "This is the smoking gun!" at you. I've started using subagents more, specifically a subagent that shells out codex. This way I can have Claude throw a problem over to GPT5 and both can come to a consensus. Doesn't completely prevent wild goose chases, but it helps a lot.

I also agree that many more times the LLM is like a blood hound leading me to the right thing (which makes it all the more annoying the few times when it chases a red herring).