Yea, I am now separating them based on

1. Those I force into the system prompt using rules based systems and "context"

2. Those I let the agent lookup or discover

I also limit what gets into message parts, moving some of the larger token consumers to the system prompt so they only show once, most notable read/write_file