Prompt Engineering Is Dead, Long Live Context Engineering

Two years ago, finding the magic phrasing that unlocked a model's best behaviour was a real skill worth a blog post of its own. Modern frontier models are far less sensitive to that kind of incantation — and far more sensitive to what is actually sitting in the context window when they generate. The skill shifted from wordsmithing the instruction to engineering everything around it.

What context engineering actually means

• Retrieval quality — the three most relevant passages beat the twelve semi-relevant ones; irrelevant context measurably degrades answer quality, it does not just sit there neutrally.
• Ordering — models weight information near the start and end of the context window more heavily than the middle; put the answer-critical passage there, not buried on page four of retrieved chunks.
• Structure — consistent, labelled formatting (headers, delimiters, explicit "SOURCE:" tags) helps the model parse what it is looking at instead of treating everything as an undifferentiated wall of text.
• Conversation memory — deciding what from prior turns actually needs to persist into the next call, instead of replaying the entire history and paying for it in cost and confusion.

None of this is about finding cleverer words for the system prompt. It is closer to database query optimization than to writing — the question is not "how do I phrase this" but "what is the minimum, highest-signal set of information this call actually needs to succeed."

The instruction still matters — just less than you think

This does not mean prompts are irrelevant. A clear task instruction, explicit output format, and a couple of well-chosen examples still measurably improve results. It means that effort spent tuning adjectives in a system prompt has a far lower ceiling than effort spent on retrieval quality, and teams that keep optimizing the wrong layer plateau early.