Most of what you send is not your prompt
By 2026 the typical production request looks like: 200 tokens of system prompt, 50 tokens of user message, and 30,000 tokens of context (RAG results, conversation history, tool definitions). The instructions are 1% of what the model is reading. "Prompt engineering" is misnamed; most of the work is context engineering.
What this changes
- The leverage moves from "better instructions" to "better context selection."
- The error budget moves from prompt phrasing to chunk relevance.
- The cost budget moves from output to input — output is small, input is huge.
- The eval shape changes — you have to measure retrieval, not just LLM output.
The full-system view
Everything in this track — chunking, retrieval, reranking, citations, caching, compression, ordering — is part of the prompt now. A junior writes the instruction; a senior shapes the context.