Context Engineering Fundamentals

Ask a candidate to "improve the prompt" and most will edit a string. Ask them to "engineer the context" and the honest answer requires naming everything that lands in the model's window on a given turn: the system prompt, yes, but also retrieved documents, tool schemas, tool outputs, conversation history, and whatever scratch state an agent loop carries forward. Somewhere around 2024, practitioners started calling this second thing by its own name — context engineering — not as a rebrand for its own sake, but because the failure modes of a production LLM system stopped living in the prompt string and started living in what got selected, in what order, and what got left out. A system prompt can be perfect and an application can still fail because retrieval returned eleven mediocre chunks instead of three good ones, or because turn 40 of a conversation pushed the one fact that mattered into the dead zone in the middle of the window.

The mechanical reason this split matters: a modern LLM call is a single forward pass over a fixed budget of tokens, and every token in that budget is either doing useful work or crowding out something that would. Prompt engineering asks "what should the instructions say." Context engineering asks a broader question with an engineering answer: given a token budget, a live retrieval system, a growing conversation, and a set of tools that can each return arbitrary amounts of text, what exactly earns a place in this specific request's context, in what order, and what happens when the accumulated candidate material no longer fits? That is not a prompt-authoring problem — it's closer to a resource-allocation and cache-design problem that happens to be expressed in natural language.

Interviewers ask about context engineering because it's where "I got a demo working" and "I run this in production" diverge. A demo has one static system prompt and a handful of turns; a production system has a retrieval pipeline whose output size varies per query, a conversation that can run to hundreds of turns, tool calls that can return megabytes of JSON, and a hard token budget that all of that has to fit inside without silently degrading quality. This subject builds the vocabulary and the worked arithmetic for that problem: prompt-caching-and-context-cost-optimization picks up the cost and caching side in depth, and context-engineering-for-agents extends this into multi-step agent loops where the context is rewritten and pruned turn over turn — both assume the model built here.