来自 Anthropic 的 Eric 讨论了如何在生产环境中负责任地使用 AI 进行氛围编程，将其与编译器类比，并强调通过产品级验证建立信任。
Eric from Anthropic discusses how to responsibly use AI for vibe coding in production, drawing parallels to compilers and emphasizing trust through product-level validation.
要点 · TL;DR
氛围编码意味着完全沉浸在氛围中，忘记代码的存在。 Vibe coding means fully giving into the vibes and forgetting the code exists.
AI 任务长度每 7 个月翻倍；很快我们必须信任 AI 编写大型代码。 AI task length doubles every 7 months; soon we must trust AI to write large code.
要负责任地氛围编码，就要充当 AI 的产品经理，而不是程序员。 To vibe code responsibly, act as a product manager for the AI, not a coder.
核心观点 · Key points
Vibe coding 意味着完全投入氛围并忘记代码的存在。 Vibe coding means fully giving into the vibes and forgetting the code exists.
AI 任务时长每 7 个月翻倍；很快我们必须信任 AI 编写大量代码。 AI task length doubles every 7 months; soon we must trust AI to write large code.
要负责任地 vibe code，就要充当 AI 的产品经理，而非编码者。 To vibe code responsibly, act as a product manager for the AI, not a coder.
将 vibe coding 集中在叶节点——代码库中没有依赖关系的部分。 Focus vibe coding on leaf nodes—parts of codebase with no dependencies.
设计具有可验证输入/输出的系统，无需阅读代码即可验证。 Design systems with verifiable inputs/outputs to validate without reading code.
反共识 · Contrarian takes
在生产环境中进行 vibe coding 不适合非技术人员；这很危险。 Vibe coding in production is not for non-technical people; it's dangerous.
技术债务无法在不阅读代码的情况下验证，这与其他领域不同。 Tech debt cannot be validated without reading code, unlike other domains.
管理者已经在处理他们不理解的具体实现；工程师也必须这样做。 Managers already handle implementations they don't understand; engineers must too.
花 15-20 分钟准备提示词比快速聊天更有效。 Spending 15-20 minutes preparing a prompt is more effective than quick chats.
指数增长意味着模型将变得好数百万倍，而不仅仅是两倍。 The exponential means models will get millions of times better, not just twice.
本期章节 · Chapters（共 12）
0. 氛围编码简介与定义Introduction and definition of vibe coding