Prompt engineering is about how you word the request. Context engineering is about what goes in beside it: which chunk of a document, which earlier message, which tool description, and in what order. In a production AI system this is the bigger half of the work, because the model only ever sees what is inside the window.
Why does it matter? Because models do not use their context evenly. In a measurement across 18 models, where the difficulty of the task was held constant and only the length of the input varied, in every model family the focused 300-token prompt won against the 113-thousand-token version that contained everything — same question, same answer present in both.
The most common misconception is that the large context window made this work unnecessary. It did not: the advertised window size is not the usable window size, and the decline does not start right at the end. Less context, but more precise, measurably gives a better answer.
What that means in practice for a system of your own is set out by Context Engineering: What Replaced Prompt Engineering at length. Directly related terms: context window, prompt engineering and RAG.