A RAG system is built in three engineering layers stacked on a single LLM call. Prompt engineering is the call itself, comprising the system message, instructions, and schema that fixes the output shape. Context engineering fills the model's window through retrieval, compression, and deciding what to keep out. Loop engineering governs what happens around the call, including when the next one fires, when the loop stops, and how the system recovers when a check fails. The article argues that most arguments about RAG are really about which layer one is standing on. It presents this three-layer framing as the dominant 2026 narrative and questions whether the evolution from one layer to the next is a true sequence or a retrospective story. Prompt engineering shipped in 2022-2023 with GPT-3.5 and ChatGPT. Context engineering ships when the prompt alone is no longer enough, with four canonical strategies: write, select, compress, isolate. The term context engineering installs in 2025, used publicly by Karpathy and Tobi Lütke. The article is part of a series on Enterprise Document Intelligence, with companion notebooks on GitHub at doc-intel/notebooks-vol1, each running a brick end to end on a real PDF.