What Context Engineering Actually Is: A Working Definition

The term gets used loosely, which breeds confusion. Here is a precise definition of context engineering, why each part of it matters, and how it differs from both prompt engineering and the emerging discipline of harness engineering.

Context engineering is the discipline of designing the dynamic systems that decide what information an AI model sees, in what format, and at what moment, so the model has exactly what it needs to accomplish a task and nothing that would distract it. That definition is deliberately broad because the discipline is broad. But every phrase in it carries weight, and unpacking them explains why this skill became central to building reliable AI.

Dynamic Systems, Not Static Text

The word "dynamic" is doing heavy lifting. A prompt is static text you write once. An engineered context is the output of a system that runs every time the model is called, assembling the relevant pieces programmatically. This is the core distinction: prompt engineering produces a string, context engineering produces a system that produces the right string for each situation.

The Right Information

"Right information" means relevant, not exhaustive. The instinct to give the model everything it might possibly need is, as practitioners repeatedly discover, reliably wrong. More information past a certain point reduces quality because it dilutes attention and introduces distractors. Right information means the smallest set that fully covers the task, which is much harder to assemble than simply including everything.

The Right Format

"Right format" means structured so the model can parse it efficiently. Interestingly, research suggests the specific serialisation format (JSON vs YAML vs Markdown) matters less than people assume. What matters is that the structure makes the information clear: that relationships are explicit, that sections are distinguishable, that the model can find what it needs without wading through ambiguity.

The Right Time

"Right time" means just-in-time, loaded when the current step needs it rather than pre-loaded wholesale. A sophisticated context system maintains lightweight references to information and fetches the actual data only when the immediate task requires it. This keeps the baseline context lean and spends tokens only on what is actually needed.

The Onboarding Analogy

A useful mental model: think of the AI as a brilliant new hire with no knowledge of your company. Prompt engineering is asking them a clear question. Context engineering is building the onboarding, the documentation access, the tools, and the workspace that let them actually do the job. The question matters, but the environment matters far more. A brilliant hire with no access to information cannot perform; a well-supported one thrives.

Where It Sits Among Disciplines

Context engineering is the middle layer of three nested disciplines. Below it is prompt engineering, the message-level work of phrasing. Above it is harness engineering, the emerging system-level discipline of designing the entire agent environment with its rules, validators and feedback loops. Context engineering is where the leverage currently sits: high enough to matter enormously, mature enough to have clear, applicable techniques.