Let's start with the hype
"Prompt engineering is dead" — people have been writing this regularly since 2025, and you will read it in 2026 too. It is worth knowing what sits behind it, because it is half marketing and half real insight.
What is not true: that phrasing does not matter. A precise instruction still improves the result more than a more expensive model does.
What is true: that phrasing accounts for an ever smaller share of the work. In a serious system the question is not how you ask, but what is in front of the model when you ask. Managing that is what context engineering means.
What is actually in front of the model?
The context window is not a big empty box you type the question into. By the time your question arrives, typically all of this is already in there:
- the system instruction, which sets the role and the boundaries
- the descriptions of the available tools — each with its own documentation
- the entire history of the conversation so far
- whatever came back from the search, if there is a search
- the examples, if there are any
- the uploaded files
- and finally your question
Those seven things share the same space, and they eat from the same attention. Context engineering is the editing of those seven.
Why has this become more important than it used to be?
Two things changed.
One: agents. An AI agent runs in a loop, and in every round whatever the previous step produced goes back into the context window. After twenty steps the window is full of partial results, most of which no longer matter. You are not writing one prompt, you are managing a continuously growing text.
Two: search. When the model works from your documents, the question is not how you ask but which five paragraphs got put in front of it. If those are wrong, the best prompt will not help. Embeddings and Vector Search is about exactly that.
The four operations
In practice context engineering means four things, and all four are about what should not be in there.
Selection. Only what the current step needs should go in. A big window is an option, not an obligation: eighty pages of which three matter will degrade the answer and inflate the bill.
Compression. If the conversation or the workflow is long, you periodically have to summarise and swap the raw history for the essentials. Agent systems do this automatically too — but what they keep is a matter of design.
Arrangement. Where things sit is not irrelevant. Information in the middle of a long text is used less reliably than information at the beginning or the end — which is why the important instruction goes after the material. Marking the structure belongs here too: delimiters and sections.
Offloading. Whatever is not needed at every step should not live in the window but in a file, a database, a note — and should come out only when it is needed.
The mindset behind it
We used to treat the prompt as a well-phrased question. Context engineering treats it as the allocation of a scarce resource.
Every token in there does two things: it consumes the budget and it draws attention away. So the good question is not what else should I put in, but what could I leave out and still get this same answer. Whatever fits that description can go.
This also explains something many people notice: that a short, clean prompt often gives a better result than a long, thorough one. The problem is not the length, it is the noise.
What does this mean if you are not building a system?
If you only use a chat interface, three things remain from all this, and all three are usable immediately:
- A new conversation for a new topic. Don't patch onto the old one. Accumulated history does not help, it interferes.
- Upload only the material you need. Not the whole folder, just those two files.
- The important instruction at the end. If you hand over a long piece of material, your request should come after it, not before.
What has not changed
Phrasing still matters, and the five basic elements of a prompt still hold. Context engineering is not a replacement for prompting but a layer around it: the same thinking, applied not to one sentence but to everything that gets put in front of the model.
Anyone telling you that you no longer need to know how to prompt is probably trying to sell you something else.
What to take away
The context window is attention, not storage. Context engineering means managing it deliberately: a small amount of well-selected, well-arranged material goes in, and the rest is left out or offloaded. Phrasing stayed just as important — it is simply no longer the only thing it depends on.