Context Engineering
What you put in the context window decides what you get out. Most people get this wrong.
Grounded in the research on context window (machine learning)
What It Actually Is
Context engineering is the deliberate craft of deciding what information goes into an AI's context window — and what gets left out. Not prompt writing. Not vibes. It's information architecture. The context window is the AI's entire working memory for a given call. Garbage in, garbage out. Strategic signal in, sharp output out. Every token is a choice about what the model gets to reason over.
Where It Came From
The term crystallized around 2024 as transformer-based models scaled to 100k+ token context windows. Andrej Karpathy popularized the framing in mid-2025, explicitly distinguishing context engineering from prompt engineering. The underlying insight is older: retrieval-augmented generation (RAG), introduced by Facebook AI in 2020, was context engineering before anyone called it that — stuff the relevant docs in, skip the irrelevant ones.