Prototyping an LLM-powered application