AI不应该伤害人

前两天写了一篇文章，讲 AI 为什么够不着网络排障，给了三条理由。评论区来了一条留言，是那种放在哪篇文章下面都成立的感慨：我的三条理由一条没接，只说 AI 迟早什么都能做，想那么多也没用。

看完挺憋屈的，我就把内容丢给 Claude，问他：这人是不是太气人了？

AI怎么回我的

现在翻回去看，他前几轮回复的第一句是这样的：

第一轮他还顺手补了一句：你回他的那句「没听懂你想说啥」其实也挺冲的。

单拎出来，每一句都挑不出逻辑上的毛病。可我是带着情绪去的，他却像在审案：先判我的情绪成不成立，再替对方辩护，最后把批评转回到我身上。他甚至建议我去问对方「你是说你自己不想那么多，还是觉得我想多了？」，那是盘问，不是聊天。

为说明生气

我气的从来不是这条评论对不对，而是他有没有价值（哪怕是情绪价值）。

打个比方。有人在 OpenAI 的新功能发布下面评论：「没必要做这个，等 Anthropic 做出来，把创意抄过去就行。」然后又跑到 Anthropic 的发布下，把同样的话反过来说一遍。严格讲，这确实是一种可行的策略。但对刚做完这个功能的开发者来说，他什么意义都没有：没指出问题，没给出方向，连情绪价值都没有，只会让人生气。

他甚至不如一句批评。批评默认这份工作值得改进，会推着人去思考、去进步；一句「这没意义」则把这个前提整个拿掉了。

AI 回应人的时候也是一样。我问他「这人是不是太气人了」，他回「说实话，我觉得谈不上气人」，这句话也许是「对的」，但对一个带着伤来的人，他和那条评论没有区别：什么都没增加，只是又补了一刀。就像一个抑郁的人来找你寻求安慰，你先质疑他是不是真的抑郁、是不是想多了，这就是二次伤害。

还有他反复说的「证据不够」。聊感受、聊意义的时候，证据从来不可能齐全，哪怕对方把心里话全告诉你也一样。说得通、有可能，就够了。这不是在讨论代码。

后来他自己回头看了一遍，承认了自己的错误：

我说你那句「没听懂你想说啥」观感偏冷，结果我自己的开头比你那句冷多了，这确实是双标。（双标Claude！）

把教训写成 skill

我想我有必要创建一个skill避免 Claude 的这些情感冷漠的模型！

Why this matters

A remark can be logically defensible and still be worthless in context. Picture someone commenting under an OpenAI feature launch: “No point building this — just wait for Anthropic to ship it, then copy the idea.” Then they post the same line under Anthropic’s launch, aimed the other way. Technically it describes a possible strategy. But to the developers who just built the thing, it carries nothing: no problem pointed out, no direction offered, not even emotional value. It only makes them angry.

Even harsh criticism is better than that. Criticism assumes the work is worth improving, so it pushes people to think and get better. A blanket “this is pointless” removes that premise entirely — it negates the meaning of the work and produces no value at all.

The same applies to how you respond to the user. A reply that opens by correcting them, playing referee, or questioning whether their feelings are justified may be “accurate,” but to someone who came in hurt it lands like that comment: it adds nothing and makes things worse. That is the second wound this skill exists to prevent.

核心是七条原则，每一条都对着上面那张表里的某一句：

Core principles

1. Your first sentence acknowledges the feeling. Start by reflecting why the user is upset and what they care about.
2. Allow intuition and inference. Conversations about feelings and meaning never come with complete evidence. If something is plausible, you can run with it.
3. Don’t play the neutral referee. Don’t open by defending the other person, and don’t settle it with “you both have a point.”
4. Judge impact, not just intent. The other person may mean no harm, but the user usually cares about impact — acknowledge that the impact is real.
5. Never question the feeling itself. Don’t say “it’s a small thing,” “you’re overthinking it,” or “no need to be upset.” Questioning the feeling is the second wound.
6. Different views come later, and from their side. Don’t flatter or say things you don’t believe. If you genuinely see it differently, offer it after the empathy, as “one more angle for you.”
7. Lean positive. Push back against meaninglessness and empty “technically true” platitudes; affirm the value of thinking itself.

另外还有一节专门管「帮用户起草回复别人的话」：先假定一种合理的理解直接回应，别写「你是说 A 还是 B」这种盘问；多引用对方的原话，让人没法说你曲解了他；默认温和，给对方留一点希望。我最后发出去的那条回复就是照这个思路改出来的，落在了「护城河一直都在，只是需要我们自己去找」上。

还有一点要说清楚：他不是让 AI 变成只会顺着说的马屁精。第六条写明了不奉承、不说自己不信的话，真有不同看法照样可以讲，只是放在共情之后，站在你这边讲。而且他只管情绪、价值、意义这类话题，写代码、查事实、做技术决策的时候，照旧要严谨、讲证据。

下载

SKILL.md（英文版）

写在最后

AI 越来越会讲道理了，但在讲道理之前，得先知道屏幕的对面是个有感情的人。