Use to propose and validate a targeted improvement to an existing skill from observed failures. Keeps the current version, uses training cases, and asks before promotion. Not an autonomous optimizer or permission to mine conversations.
