概述
It goes further than prompt wording. It matters because output quality depends heavily on what is in the window, and in agents and long tasks that context is assembled by code on every call, not typed once by a person.
深入探討
Prompt engineering focuses on wording a single request. Context engineering takes a wider view. The model receives one sequence of tokens, and everything in it competes for space and for the model's attention. That sequence usually includes a system prompt, tool definitions, retrieved passages, prior turns, tool outputs and sometimes long-term memory. The term became common in 2025, when teams building agents found that many failures came from the model having the wrong information, too much of it or conflicting information, rather than from poor phrasing. Several principles follow. First, relevance beats volume. Some models accept hundreds of thousands of tokens, but they do not use every token equally well. The 'lost in the middle' study (Liu et al., 2023) found that models often use information at the start and end of a long input better than information buried in the middle. Second, context is a budget. Every token costs money and adds latency, and irrelevant material can distract the model. Third, structure helps. Clear sections, labeled documents and consistent formats make it easier for the model to find what matters. Common techniques include: - retrieval-augmented generation, which fetches only the relevant passages - just-in-time loading, where an agent holds lightweight references such as file names or IDs and uses tools to pull full content when needed - structured note-taking that keeps state outside the window - compaction of long histories - sub-agents that explore in their own context and return condensed findings One misconception is that context engineering is just a new name for prompting. Wording is one part of it, but most of the work is systems design: deciding what gets retrieved, summarized or dropped, and in what order it appears. Another misconception is that a bigger window removes the need for this. A bigger window lowers the pressure, but cost and attention limits remain.
戰略影響
速度與規模
語言工作流程可以在不犧牲一致性的情況下更快地移動。
交通與覆蓋範圍
它擴展了跨語言和溝通方式的訪問。
更明確的決策
團隊可以花更多時間進行判斷，而自動化則可以處理重複。
The Future of Context Engineering
Context windows have grown quickly, and models are getting better at using long inputs. Cost, latency and attention limits still make it important to choose what goes in. More of this work is moving into frameworks and platforms, and several providers already offer built-in memory tools, automatic compaction and managed retrieval. The skill is likely to shift from assembling context by hand toward designing policies for what an agent should remember, fetch or forget, and then testing those policies. It is an open question whether future model designs will need less careful curation. For now, deliberately designed context is one of the most reliable ways to improve model output.
現實世界的實施
A customer support bot retrieves only the three most relevant help-center articles plus the customer's plan tier. It does not paste the whole knowledge base into every request.
A coding agent starts with a list of file paths and function signatures, then uses a read-file tool to load full files only when it needs them.
A meeting assistant saves decisions and action items to a structured notes file and reloads that file next session, instead of replaying every past transcript.
A team changes a database tool so it returns a filtered summary with a row count instead of 5,000 rows of raw JSON. This stops the agent's window from filling with data it will never use.
風險與防護欄
幻覺的事實可以悄悄地進入報告、支持流程或研究成果。
及時的敏感性可能會在類似的請求中產生不一致的結果。
如果存取控制薄弱，敏感文字資料可能會暴露。
實施路線圖
在推出之前定義輸出格式、語氣和品質標準。
當準確性很重要時，請使用可信任來源進行地面回應。
為高風險輸出保留人工審查檢查點。
追蹤故障模式並定期重新訓練提示或工作流程。
不斷探索
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Context Engineering quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
常見問題
What is Context Engineering?
Context engineering is the practice of deciding what a language model sees at each step (instructions, retrieved documents, tool definitions and results, conversation history and memory) and fitting it into a limited context window. It goes further than prompt wording. It matters because output quality depends heavily on what is in the window, and in agents and long tasks that context is assembled by code on every call, not typed once by a person.
What best distinguishes context engineering from prompt engineering?
Context engineering treats the whole token sequence the model receives as something to design: instructions, retrieved passages, tool definitions and outputs, history and memory. Prompt wording is only one part of it.
What did the 'lost in the middle' research suggest about long inputs?
Liu et al. (2023) found that performance often drops when the relevant information sits in the middle of a long context. That is one reason relevance and placement matter more than sheer volume.
Why does the guide describe context as a budget?
Tokens in the window have real costs in price and response time. Irrelevant tokens also compete with relevant ones for the model's attention, so they should be spent deliberately.
What is just-in-time loading?
Instead of loading everything up front, the agent holds pointers and pulls in full content only when a step requires it. This keeps the window lean.
Why is it useful to put stable content, such as system instructions and tool definitions, at the start in a fixed order?
Providers that support prompt caching can reuse a prefix that stays byte-identical across calls. Putting stable content first and the changing conversation last makes that possible.
繼續學習
相關指南
為此主題精選的更多指南