Context engineering

The practice of designing which information reaches an AI model, in what order, and in what format, so the model can do useful work.

Context engineering designs the request-time information environment around a model. It covers source retrieval, state, instructions, examples, tool schemas, memory selection, ordering, compression, provenance, and token limits. Context interacts with model capability and runtime behavior: the same packet can help one configuration and distract another.

Builder example

Wrong or stale context is one important failure source, alongside model limits, tool errors, ambiguous goals, and bad evaluation. A support bot that retrieves an outdated policy can answer incorrectly even with a capable model. Versioned sources, freshness filters, citations, and representative tests make that part of the system observable.

You ask an AI to draft a client email. It produces three paragraphs of corporate filler because it has no examples of how you write to clients.

Show it previous emails you wrote for a similar audience and purpose. Ask it to follow their tone and level of detail, then compare the draft with your examples.

Common confusion: Context engineering does not mean maximizing text or building retrieval for every task. A direct request may need no external source, while a policy answer may need one authoritative passage and its date.