دانلود کتاب مهندسی نرم‌افزار بومی هوش مصنوعی – ساخت با عوامل هوش مصنوعی و توسعه مبتنی بر مشخصات

When you interact with an AI coding agent, everything the model knows about your task, your codebase, your coding standards, and your intentions must arrive in a single place: the context window. The model can’t browse your files on its own initiative, remember what you agreed on last Tuesday, or look up your team’s preferred testing library unless something in its environment delivers that information. It can only work with what it receives. And what it receives, you control. That’s the insight behind context engineering: what matters most isn’t how clever your prompt is. It’s whether the model receives the right information to reason well. Your system prompt, your rules files, the code snippets you reference, the tool definitions the runtime injects, the conversation history that accumulates over the session: all of it lands in the context window and shapes what the agent does next. Context engineering is the discipline of making that context work well. It means deciding what to include and what to leave out, structuring information so the model’s attention lands where it matters, and keeping the context accurate as sessions evolve and codebases change. Consider a straightforward scenario: You ask your coding agent to add a new API endpoint to a Node.js service. The agent writes something that looks reasonable. Then you open the file and notice it uses Moment.js, which your team deprecated two years ago because of bundle size. It skipped the Zod validation your architecture requires for all inputs. The test it wrote is a unit test, but your CI pipeline gates on integration test coverage. Nothing the agent produced was wrong in the abstract–it was just wrong for your system. The temptation is to respond by writing a more detailed prompt next time. But the real issue isn’t how you phrased the request. It’s that the agent had no context about Moment.js being banned, about Zod being required, or about your testing conventions. All three problems would disappear if the relevant information were reliably already in the context before the agent started reasoning. That’s context engineering. Context engineering is different from tool selection and prompt crafting, though they overlap. Choosing whether to use Cursor or Copilot, or deciding whether to use Claude or GPT-x as the underlying model, are tool selection decisions. Figuring out how to phrase a request clearly and precisely is prompt crafting. Context engineering is the layer beneath both: it concerns the totality of what the model sees when it begins reasoning, regardless of which tool or model you’re using.