For a while, “prompt engineering” carried an air of secret knowledge — the right incantation, the magic phrase that unlocked a better answer. That era is quietly ending, and not because the models got worse at listening. They got better. The bottleneck moved.
The harder problem today is not how you phrase the question. It is what the model can see when it answers. Give a capable model the wrong three documents and it will reason beautifully toward the wrong conclusion. Give it the right three and a plain request, and it rarely needs coaxing. The work has shifted from wording to assembly — which is why I have started calling it context engineering.
Think of it like briefing a brilliant new hire. You would not obsess over the exact sentence you use to assign a task. You would worry about whether they have the account history, the constraints, the one email thread where the decision was actually made. The model is the same. Its output is only as good as the working set you hand it.
Picture an airline’s chat assistant. A passenger types, “My flight got cancelled — fix it.” The clever-prompt approach tries to script the perfect reply. The context-engineering approach asks a better question: before answering, what should the assistant be holding? Your booking, your fare rules, the next three flights with real seats, your loyalty tier, and the airline’s rebooking policy for today’s weather. Hand it those, and a plain request produces a genuinely useful answer. Hand it nothing, and even the most beautifully worded prompt invents a flight that does not exist.
The same logic shows up everywhere. An e-commerce helper that can see your order, the courier’s live status, and the returns policy can actually resolve a complaint instead of apologising in a loop. A payments assistant that knows which transactions are yours, which were already flagged, and the bank’s dispute rules can tell fraud from a forgotten subscription. In each case the magic is not the phrasing — it is the curated bundle of facts placed in front of the model at the right moment.
This reframing changes how you build. Retrieval quality matters more than prompt cleverness. Knowing what to leave out matters as much as what to include, because a bloated context buries the signal. And the discipline starts to look less like copywriting and more like systems design: sources, freshness, relevance, and the ruthless editing of noise.
The phrasing still helps at the margins. But the leverage has moved upstream. The teams shipping reliable AI right now are not the ones with the cleverest prompts. They are the ones who decide, deliberately, what their model is allowed to know.