the real solution is to simply provide hints in responses so that the model may self-correct, e.g., recommended next actions, describe commands to get schema definitions, etc.