Context engineering
The discipline of curating what information goes into an LLM's context window — selecting, ordering, and formatting the system prompt, examples, retrieved documents, and conversation history for maximum effectiveness.
Context engineering is the 2026 evolution of prompt engineering. Where prompt engineering focused on a single user-facing instruction, context engineering treats the entire context window as the design surface: system prompt, tool definitions, few-shot examples, retrieved documents, conversation history, agent state — every token matters.
The core principles: structure the context with stable content at the top (cached, reused), dynamic content at the bottom (changes per request). Use clear delimiters between sections. Order matters — recency bias is real, primacy bias is real. Trim aggressively: long contexts dilute attention and degrade reasoning even within the window limit.
For agent builders, context engineering is where most quality wins happen. A frontier model with poor context performs worse than a mid-tier model with engineered context. The discipline is part craft, part empirical eval — measure the impact of every context decision on your actual eval set.
Frequently asked
How is context engineering different from prompt engineering?+
Prompt engineering focused on the user-facing instruction. Context engineering covers the full context window — system prompt, tools, examples, retrieved docs, conversation history. It is the broader discipline that prompt engineering grew into.
What is the most common context engineering mistake?+
Stuffing too much into the context. Long contexts dilute attention and slow inference. The best operators trim aggressively, retrieve precisely, and put the most important information closest to the model's response position.