Context Engineering

Disclaimer. Provider-specific notes (context-window sizes, tool-calling APIs, MCP support) reflect the patterns at time of writing. Verify current limits and APIs in the provider's documentation before sizing a system.

LLMs are stateless. Every call starts from zero. What the model "knows" about your task is exactly what's in the context window — system prompt, conversation history, retrieved documents, tool results, the user message. Context engineering is the discipline of getting the right information into that window and nothing else.

The context window is a budget

You are spending tokens on:

[ system prompt ] [ persona / role ] [ few-shot examples ]
[ retrieved knowledge ] [ conversation history ]
[ user query ] [ scratchpad / reasoning ] [ output ]

Every token costs money, latency, and (past a point) quality — long contexts degrade attention to detail and amplify the lost-in-the-middle effect. Treat the window like RAM: relevant data only, evicted when stale.

RAG (Retrieval-Augmented Generation)