A skill isn't static. Waggle re-optimizes it against real outcomes — and keeps every version, with the score that earned it. You don't tune prompts; the hive does.
How it improves: when a skill runs, Waggle's optimizer (GEPA) proposes refinements and a three-model judge scores them against held-out tasks. A new version only ships if it actually wins — and the old ones stay, so you can always roll back. Nothing about your data leaves the machine.