Deconstructing Agent Skills: A LangGraph Deep Dive