[verdverm]

Nothing specific to Claude in your two issues, I see the same thing with other models. They really aren't that different

Instructions go a long way. There probably needs to be a better LMM+prompt+loop at the top, the one you interface with, or one below that maybe.

My next step is taking over this instead of outsourcing to M$, Google, or Anthropic. It's just too important to let others decide how they should work at this point. It needs to be more open and something we can tinker with like vim