Why context engineering still needs human skill

In this month’s feature, part of our mini-series on moving from content to context, we want to turn to the human value in context management.  

Gartner describes context engineering, the discipline of deliberately structuring what an AI system knows, as the capability that supersedes prompt engineering. Anthropics published guide for treating context as a finite resource to be engineered rather than a bucket to be filled. Nielsen Norman Group argues that the output of design work is moving from documents written for people to curated context that guides the AI, and OpenAI's latest memory architecture now holds a living, synthesised picture of each user rather than a saved list of facts.  

The naming matters less than what sits beneath it. The value is no longer in having more context; it is in how deliberately that context is engineered; and the skill of the human behind it. Organising what you know is a genuinely different discipline from letting a system search your files and surface whatever it finds, and it is the one that keeps an experience coherent. 

Organise what you know before you automate it 

The instinct will be to point an agent at your existing content and let it run. We'd resist that. The more useful first step is structuring what you actually know; the facts, how they relate, and where the gaps and contradictions sit. Retrieval-first approaches pull cold blocks of text and inject whatever turns up, which comes apart the moment two sources disagree. Structure the knowledge first, and everything built on top has something coherent to draw on. 

Expression is part of the context 

Tone, point of view and authority are not decoration added at the end; they travel inside the information itself, and carry a human signature. We've seen this in testing: a set of writing guidelines scattered with em dashes quietly taught the system to scatter its own output with them, overriding the very rule those guidelines were written to enforce. How something is said is carried in the context, often across several layers, so it has to be built in deliberately, or the output drifts towards a generic average. 

Build context, don't just retrieve it 

There is a real difference between a system that reasons over organised knowledge and one that retrieves whatever a search returns. The technical lift is genuine: structured knowledge, and clear rules about what can be drawn on, from where, and when. But get those conditions in place and the experience layer above has room to evolve as the models themselves do. 

This is the second in our short series on what the move from content to context asks of a business. Next month, we turn to the process that runs on top of it.