What is context engineering?

Context engineering is the work of deciding everything a model sees before it answers. The team building the system sets the strategy, and the code around the model applies it each time a question arrives. Four things fill that space: the standing instructions, facts fetched from your documents, the conversation so far, and the tools the model may use. Prompt engineering is how you word the question. Context engineering is what sits in front of the model while it reads that question.

Context engineering is the work of assembling what a model reads before it answers. Which sources get searched, how much of the earlier conversation survives, what gets summarised once it grows too long, which tools the model is even told exist: all of that is decided by somebody. Ask a support bot about your refund policy and it invents one, because nobody arranged for the policy to be there. You could paste the policy in yourself, though nobody does that for every customer, so the system has to assemble it instead. Why does this matter more than a sharper prompt? A prompt shapes one answer, while a system running twenty steps needs the right material at every one of them. The constraint is that the context window is finite, so more is not better and adding everything you own buries the part that mattered. When an AI feature works in testing and fails in production, this assembly is usually where it broke.

A surgeon about to operate. How the request was worded matters very little at this point. What matters is whether the scans on the screen belong to this patient, whether the chart lists the right allergy, and whether the instruments laid out are the ones this operation needs. Hand a brilliant surgeon the wrong file and the operation goes badly, however good the surgeon is.