This paper introduces and evaluates a novel Rhetorical Structure Theory (RST)-based in-context learning method to improve the quality of AI-generated educational diagrams, finding that while it reduces hallucinations, LLMs still struggle with complex inputs and require careful application in educational contexts.