This paper investigates how assertion-level features influence the quality of test code summaries generated by Large Language Models (LLMs). It introduces a new benchmark of Java test cases and demonstrates that prompting with assertion semantics significantly improves summary quality while being more token-efficient.