📄 中文摘要
许多系统提示的设计往往过于简单，例如"你是一个有帮助的助手，帮助用户满足他们的请求"。虽然这样的提示可以使模型产生响应，但输出结果往往不够一致和结构化。当要求模型进行代码审查时，可能会得到风格评论和安全发现的混合，而不是清晰的结构；在诊断事件时，模型可能会生成一大段文本，掩盖可操作的步骤；在设计架构时，模型可能会选择服务而不解释权衡。因此，对于需要产生一致且结构化输出的代理，无论是单代理工作流还是多代理系统，问题在于提示的设计，而非模型本身。
有效的系统提示编写：AI代理的RISEN框架
📄 English Summary
Writing System Prompts That Actually Work: The RISEN Framework for AI Agents
Many system prompts are overly simplistic, such as 'You are a helpful assistant. Help the user with their request.' While such prompts elicit responses from the model, the outputs are often inconsistent and unstructured. For instance, when asked to review code, the model may produce a mix of style comments and security findings without a clear structure. When diagnosing an incident, it might generate a wall of text that obscures actionable steps. In architecture design, it may select services without explaining trade-offs. Therefore, for agents that need to produce consistent and structured outputs, whether in single-agent workflows or multi-agent systems, the issue lies in the prompt design rather than the model itself.