Abstract: Scientific experiments are human-in-the-loop procedural knowledge work: users must follow protocols, manipulate tools and materials under safety constraints, and continuously align observations with step-level states. Existing laboratory assistants often rely on end-to-end vision-language inference or protocol-text retrieval, making operational skills difficult to reuse, audit, and transfer. SciSkillRAG is a training-free agentic retrieval-augmented generation (RAG) framework that constructs a scientific skill graph G and an event-centered skill-card library K from heterogeneous laboratory annotations, and retrieves, verifies, and composes evidence-grounded assistance from observations, standard operating procedure snippets, questions, and history. From 8897 real annotated records, the system builds 8897 skill cards and a graph with 19243 nodes and 21955 edges. The reconstructed 900-case Standard-900R benchmark comprises three splits and five tasks, excluding test sources and execution groups while retaining skill reuse within known SOP families. SciSkillRAG achieves a macro task score of 0.882, outperforming the strongest external baseline KG2RAG at 0.688, with an issue macro-F1 of 0.911 and an evidence F1 of 0.672. Removing the planner, graph path, or verifier reduces the macro score to 0.837, 0.838, and 0.662, respectively. These results demonstrate the role of structured skill knowledge in evidence-grounded assistance for scientific laboratory operations. Task answers, retrieved evidence, and graph relations are evaluated separately.