Maggie Appleton 写设计、人类学和智能体界面时，习惯先画清楚隐喻，再谈工具。把「Language Model Sketchbook, or Why I Hate Chatbots」整理成可阅读的中文笔记：问题在哪、界面默认会把人带去哪、落地时该改什么。原站导航、评论回响和广告已去掉。

Daemons

We don’t quite know what to do with language models yet. But we have some hunches. To me they seem obviously useful as epistemic rubber ducks 🐥 – as things we can query for fuzzy answers, bounce ideas off, and think through problems with. They can help strengthen our own critical thinking and reasoning abilities in the same way a good debate partner does.

This isn’t the only way they’re useful, but it’s the one I’m most curious about. So I’ve been playing with it.

Branches

The primary interface everyone and their mother jumps to at this point is the chatbot. We are irreversibly anchored to this text-heavy, turn-based interface paradigm. And sure, it’s a great solution in a lot of cases! It’s flexible, familiar, and easy to implement.

But it’s also the lazy solution. It’s only the obvious tip of the iceberg when it comes to exploring how we might interact with these strange new language model agents we’ve grown inside a neural net.

Epi

I won’t turn this into an anti-chatbot-tirade, but I am certainly brewing one. A few of my friends have eloquently written about this so just go read their work in the meantime.

Back to the point: I have a large, sprawling Figma file full of tiny interface sketches trying to explore what kinds of non-chatbot, epistemic-rubber-ducky-interfaces could be helpful for us. Here’s a couple of them.

落地时建议先做的 5 件事

1. 先写清这个工具在强化思考还是在替人思考。
2. 默认交互不要只会谄媚：该追问、该给反例、该要求证据。
3. 智能体界面要露出推理步骤，而不是只给一个光滑答案。
4. 知识发布用可生长的笔记，而不是一次性营销长文。
5. 改产品前先画隐喻：用户以为自己在做什么，系统实际在做什么。

和智能体产品怎么接

龙虾PRO做 OpenClaw 落地时，最该从这类笔记里拿走的是「别把聊天框当唯一界面」。数字员工要能追问、能验证、能把过程摊开，而不是只负责说好话。

本文侧重全链路风控方法论。落地时请用自身业务单据做回放验证，不要把示例阈值直接当生产策略。 相关：风控体检 · 方案资源