noun · AI practice
Context engineering
also called context management
Packing the backpack well. It is the job of choosing what the AI gets to see at each step, and what stays out.
The practice of deciding what information goes into a model's context window for each request, and in what form and order.
The harness assembles each request from instructions, history, retrieved documents, tool descriptions and tool results, within a fixed window. Engineering it means deciding what to include, trim, summarize or fetch later at every step.
Real-life example
You build an assistant that answers questions about a big codebase. Pasting everything in does not fit, and pasting too little makes it guess. You end up choosing which files to load, keeping a short summary of the chat, and dropping old tool output. The model is the same, and what it sees is what changed.
Where the analogy breaks
-
It is not the same as writing a good prompt. Wording is one piece, and much of the work is choosing which facts, files and history to include.
-
More context is not better context. Extra material costs tokens and can bury the part that matters.
-
It is never finished once and for all. The right contents change on every step of a task, so it is a running job done by the harness, not a one-time setup.