From prompting to context engineering

Early LLM use was about phrasing a clever instruction. As soon as systems added retrieval, tools, and multi-turn state, the bottleneck moved: quality now depends on everything in the context window, not just the user's sentence.

Two related disciplines

• Prompt engineering — how you phrase instructions and examples for a task.
• Context engineering — deciding what information fills the window: system prompt, examples, retrieved chunks, tool results, memory, history — and in what order and format.

Prompt engineering is a subset of context engineering. For a chatbot the prompt may be most of it; for a RAG or agent system, the assembled context dominates.

Why the shift happened

• The window is a finite, attention-biased budget (lost in the middle) — what you include and where matters more than clever wording.
• Most of what the model "knows" at answer time is what you put there: a model with no relevant context can't be prompted into facts it doesn't have.
• Agents accumulate tool outputs and history that must be curated, summarized, or dropped each turn.

The mental model

Stop asking only "what's the best prompt?" and start asking "what's the smallest, most relevant context that lets the model succeed — and in what order?"

Everything else in this branch — prompt anatomy, assembling context, memory — is a tool in that larger job.

Connects to: context window · RAG · agents