A prompt is written for one request; context is assembled across many, so it cannot be as specific. Context engineering is the discipline of deciding what belongs in that shared layer: what loads on every turn, what loads on demand, what is expressed as a tool interface instead of a sentence, and what is better handed over as a reference such as a test suite or a mockup.
It is also the layer that ages. Guidance written to compensate for a model’s weaknesses depreciates with each release, while guidance describing your own codebase, conventions, and risk tolerance does not. Anthropic removed over 80% of Claude Code’s system prompt for its Claude 5 generation models with no measurable loss on coding evaluations, which is the sharpest illustration so far that context accumulated for older models becomes a cost rather than a safeguard.