Improving Knowledge Extraction from LLMs for Task Learning through Agent Analysis