Stars: Improving Knowledge Extraction From Llms For Task Learning Through Agent Analysis