Understanding Specification-Driven Code Generation with LLMs: An Empirical Study Design