Prompt, Context, Loop: The Three Engineering Layers Every RAG System Is Built On

Every Retrieval-Augmented Generation (RAG) system is constructed from three distinct engineering layers: prompt, context, and loop. Prompt engineering shapes the individual LLM call with instructions and schemas, while context engineering curates the information that fills the model's attention window through retrieval and compression. Loop engineering then orchestrates the entire process, deciding when to make subsequent calls, how to handle failures, and when a task is complete. This apparent evolution of disciplines is actually a story of shifting bottlenecks, as improving models and larger context windows moved the primary production challenges from prompts to context management and finally to agentic loops.