Context engineering (chrisloy.dev)

As LLMs move from chatbots to embedded decision-makers, the old art of prompt engineering is giving way to "context engineering" — a deliberate, systems-level practice that treats every token in the model’s context window as an engineering artifact. Rather than hunting for the perfect wording, practitioners now design the full inference context: system messages, curated documents, tool and function-call interfaces, memory summaries, and multimodal inputs. This reframes LLMs from mystical oracles into briefed analysts whose outputs depend on precise, timely, and relevant tokens; the shift matters because it reduces hallucination, improves up-to-date accuracy, and makes LLM-driven components composable and auditable within larger software systems. Technically, context engineering responds to hard constraints (finite context windows) and new capabilities (in‑context learning, chat framing, RAG, function calls, and multimodal tokens). The article illustrates how supplying the current date, relevant statistics, and a defined calculation function lets an LLM coordinate retrieval and an external computation to produce an accurate weekly box‑office figure — instead of regurgitating outdated training data. Practically, it promotes design-pattern thinking (composition over inheritance), treating retrieval, transformation, and tool invocation as interchangeable components so complex agentic systems remain testable, maintainable, and safer as they scale.