Prompt engineering — finding the magic wording that unlocked a good answer — defined an early era of working with language models. As models got better at following plain instructions, that craft mattered less. What replaced it is broader and more durable: context engineering, the discipline of assembling everything a model sees.
Beyond the prompt
The prompt is just one input. The full context includes retrieved documents, tool outputs, conversation history, memory, system instructions, and how all of it is ordered and formatted. Context engineering is the practice of deciding what goes into that window, in what shape, and in what order — because for a capable model, the right context matters far more than clever phrasing.
You used to tune the question. Now you tune everything around it.
Why the shift happened
Modern models follow instructions well, so squeezing quality out of wording hit diminishing returns. Meanwhile, the failure modes that remain are about context: irrelevant material distracting the model, key information buried where it gets overlooked, missing context the model needed, stale history it should have dropped. These are engineering problems, not phrasing problems.
The practical discipline
Good context engineering means retrieving the right information (not just similar-looking text), placing it where the model will actually use it, pruning what's stale, and keeping the signal-to-noise ratio high. It's the connective tissue between retrieval, memory, and generation — and it's where a lot of the real quality of an AI system is now decided.
The name is new; the lesson is old. Give a capable system the right information, well-organized, and it will do well. The work moved from wording the question to curating the world the model reasons over.
