【转译】Context Engineering vs Prompt Engineering，理解Context Engineering

password

Created time

Jul 6, 2025 09:41 AM

type

status

date

slug

summary

category

icon

💡

Context Engineering最近因为ak变的很火，那他是什么，与prompt什么关系呢？有一个视频做了很好的解释，最核心的一点是“ Prompt Engineering 是Context Engineering的一部分，如果 Prompt Engineering 正在编写一个出色的指令......上下文工程决定该指令之前和之后会发生什么——记住什么，从内存或工具中提取什么，整个事情是如何构建的。”，具体如下：

原文作者：Mehul Gupta

YouTubeContext Engineering vs Prompt Engineering

Context Engineering vs Prompt Engineering

Andrej Karpathy has recently coined a new term Context Engineering after Vibe Coding #ai #chatgpt #promptengineering #vibecoding

Context Engineering 包含了prompt的设计，它是面向的大模型对话交互的设计框架考量。

Prompt Engineering focuses on what to say to the model at a moment in time.

Context Engineering focuses on what the model knows when you say it — and why it should care. If Prompt Engineering is writing a brilliant instruction…Context Engineering is deciding what happens before and after that instruction — what’s remembered, what’s pulled from memory or tools, how the whole thing’s framed.

如果 Prompt Engineering 正在编写一个出色的指令......上下文工程决定该指令之前和之后会发生什么——记住什么，从内存或工具中提取什么，整个事情是如何构建的。

Bad Prompt Engineering: 糟糕的提示工程：

Output tone is wrong 输出错误

Instructions get ignored 指令被忽略

Model acts like it’s drunk 模特表现得像傻子一样

You spend hours tweaking commas and synonyms 您花费数小时调整逗号和同义词

Bad Context Engineering: 不良环境工程：

Model forgets why it’s even in the conversation 模型甚至忘记了为什么它出现在对话中

Prompt gets drowned in noise 提示词被噪音淹没

Output is generic, unhinged, or misleading 输出是通用的、不合理的或具有误导性的

RAG breaks, memory leaks, tool chaining fails RAG 中断、内存泄漏、工具链接失败

Protects your prompt. You can write the best instruction ever, but if it’s lost at token 12,000 behind three FAQs and a JSON blob, it won’t matter.

保护您的提示。 您可以编写有史以来最好的指令，但如果它在三个常见问题解答和一个 JSON blob 后面丢失了 12,000 个令牌，那也没关系。

Structures everything around the prompt. Memory, retrieval, system prompt — all exist to support the prompt’s clarity and priority.

围绕提示符构建所有内容。 内存、检索、系统提示符 — 所有这些都是为了支持提示的清晰度和优先级。

Handles scale. You don’t need to keep engineering prompts for every variation. You inject structured context that adapts to different users/tasks.

处理缩放。 您无需为每个变体保留工程提示。您注入适应不同用户 / 任务的结构化上下文。

Manages constraints. Token limits? Latency? Costs? Context Engineering decides what gets dropped and what stays.

管理约束。 令牌限制？延迟？成本？上下文工程决定丢弃哪些内容，保留哪些内容。

Mindset: Prompt Engineering is about crafting clear instructions; Context Engineering is about designing the entire flow and architecture of a model’s thought process.

心态：Prompt Engineering 是关于制定明确的指令;上下文工程是关于设计模型思维过程的整个流程和架构。

Scope: Prompt Engineering operates within a single input-output pair; Context Engineering handles everything the model sees — memory, history, tools, system prompts.

范围：Prompt Engineering 在单个输入-输出对中运行;上下文工程处理模型看到的所有内容 — 内存、历史记录、工具、系统提示。

Repeatability: Prompt Engineering can be hit-or-miss and often needs manual tweaks; Context Engineering is designed for consistency and reuse across many users and tasks.

重复性：Prompt Engineering 可能是偶然的，通常需要手动调整;上下文工程旨在实现许多用户和任务之间的一致性和重用性。

Scalability: Prompt Engineering starts to fall apart when scaled — more users = more edge cases; Context Engineering is built with scale in mind from the beginning.

可扩展性：Prompt Engineering 在扩展时开始分崩离析 — 更多的用户 = 更多的边缘情况;环境工程从一开始就考虑到了规模。

Precision: Prompt Engineering relies heavily on wordsmithing to get things “just right”; Context Engineering focuses on delivering the right inputs at the right time, reducing the burden on the prompt itself.

精度：Prompt Engineering 在很大程度上依赖于文字加工来使事情变得“恰到好处”;上下文工程专注于在正确的时间提供正确的输入 ，从而减轻提示本身的负担。

Debugging: Prompt Engineering debugging is mostly rewording and guessing what went wrong; Context Engineering debugging involves inspecting the full context window, memory slots, and token flow.

调试：Prompt Engineering 调试主要是改写和猜测出了什么问题;上下文工程调试涉及检查完整的上下文窗口、内存插槽和令牌流。

Tools Involved: Prompt Engineering can be done with nothing but ChatGPT or a prompt box; Context Engineering needs memory modules, RAG systems, API chaining, and more backend coordination.

涉及的工具： 提示工程只能用 ChatGPT 或提示框来完成;上下文工程需要内存模块、RAG 系统、API 链接和更多的后端协调。

Risk of Failure: When Prompt Engineering fails, the output is weird or off-topic; when Context Engineering fails, the entire system might behave unpredictably — including forgetting goals or misusing tools.

故障风险： 当 Prompt Engineering 失败时，输出很奇怪或偏离主题;当情境工程失败时，整个系统的行为可能会不可预测，包括忘记目标或滥用工具。

Longevity: Prompt Engineering is great for short tasks or bursts of creativity; Context Engineering supports long-running workflows and conversations with complex state.

长寿：Prompt Engineering 非常适合短期任务或创造力的爆发;上下文工程支持长时间运行的工作流和具有复杂状态的对话。

Effort Type: Prompt Engineering is like creative writing or copy-tweaking; Context Engineering is closer to systems design or software architecture for LLMs.

努力类型：Prompt Engineering 就像创意写作或文案调整;上下文工程更接近于 LLM 的系统设计或软件架构。