The thing we used to call prompt engineering was a useful idea for one model generation. The clever phrasing tricks stopped mattering as the models got better. What replaced them is harder and more interesting. It is the work of designing what the model sees before you ask it anything.
Two years ago, the difference between a working AI feature and a broken one was often a sentence in the prompt. Adding "think step by step" or restructuring the instruction format could change a result from useless to good.
Modern models do most of that on their own. The marginal value of clever prompting has collapsed. The marginal value of better context has gone up.
Context engineering is the practice of deciding what information the model has in its window when it answers. Which documents are loaded, which tools are exposed, which past conversations are summarized, which constraints are enforced before the user even types.
A good context is small, relevant, and structured. A bad context is everything the team could think of, dumped in. Most production AI systems fail because someone confused "more context" with "better context."
I think about it as five layers, in roughly this order.
Two failure modes. The first is the cluttered context, where the model has access to so much that it cannot tell what is relevant and starts guessing. Symptoms include hallucinated answers that combine details from unrelated documents.
The second is the missing context, where the model is asked a question it cannot possibly answer correctly because the right document was never retrieved. Symptoms include confident wrong answers about things that are right there in the knowledge base.
I start by writing the worst possible context for the task, then deleting from it. Everything that survives has to justify its presence. Anything I cannot defend gets cut.
I also write a one-line description of what each piece of context is for. If I cannot describe what a chunk of system prompt is doing, neither can the model. That chunk is noise.
The new skill is closer to information architecture than to copywriting. You are designing the world the model wakes up inside. What it can see, what it can touch, what it has been told about itself.
The teams winning with AI right now are not the ones with the cleverest prompts. They are the ones who treat the context window like a product surface, with the same care they would give to an onboarding flow.