维度 (Dimension) | 传统提示词工程 (Traditional Prompt Engineering) | Manus的上下文工程 (Context Engineering) |
核心焦点 (Core Focus) | 优化单次输入的指令 (Optimizing the initial instruction) | 构建智能体残缺的推理环境 (Structuring the agent's entire reasoning environment) |
信息范围 (Information Scope) | 静态的,局限于提示词和上下文窗口 (Static, limited to the prompt and context window) | 动态、持久化、可外部扩展的 (Dynamic, persistent, and external) |
形态感知 (State Awareness) | 有形态的,每次交互都是独立的 (Stateless; each turn is largely independent) | 有形态的,持续追踪目的、进度和历史 (Stateful; tracks goals, progress, and history) |
错误处理 (Error Handling) | 脆弱的,遇错则中缀或需用户干涉 (Brittle; fails on error or requires user intervention) | 有韧性的,为自我修正和重试而设计 (Resilient; designed for self-correction and retries) |
目的 (Goal) | 生成一次高质量的回复 (Generate a single, high-quality response) | 完成一个复杂、多步、长周期的义务 (Complete a complex, multi-step, long-running task) |