Context Engineering

The discipline of controlling everything that reaches the model as context: conversation history, retrieved documents (RAG), available tools, structured output, and guardrails. It's the layer that precedes — and often outweighs in impact — prompt engineering.

The common instinct is to optimize the prompt. But most of the gains tend to come from an earlier layer: what the model receives as the basis for reasoning. See Context Engineering: what really improves the performance of LLMs.