A method for turning a history of resolved work items (bugs, incidents, support tickets, change requests) into agent skills that do first-level analysis of new items, and for measuring and tuning those skills the way you would a machine-learning model: train, validation and test splits, blind inputs, metrics, error analysis, and a test set touched once.
The skill is the "model". The knowledge base of past cases is the "training data". Rules in the skill text are the "weights" you tune. A scorer compares skill output to how items were really resolved.