We Do Not Teach Thinking to AI

我们并不需要教 AI 如何思考

Most of us learned to prompt AI by guiding its thinking. “Think step by step.” “Here’s an example of how to solve this.” “First check A, then compare B, finally conclude with C.” These techniques made sense — because we were working with models that needed a path. Without structure, they’d rush to a conclusion. 我们大多数人学习提示词（Prompting）的方式，是引导 AI 的思考过程。比如：“一步步思考”、“这是一个解决问题的示例”、“先检查 A，再对比 B，最后得出 C 的结论”。这些技巧在当时是有意义的，因为我们所使用的模型需要指引。如果没有结构化的引导，它们往往会匆忙得出结论。

But reasoning models shift this premise. Thinking Comes Before the Answer. General conversational models excel at producing natural answers quickly. For tasks where the direction is clear — brief summaries, simple explanations — this is sufficient. Reasoning models work differently. Rather than pushing problems straight toward conclusions, they’re designed to compare conditions, trace possible paths, and hold problems longer before forming answers. Models like Claude’s extended thinking or OpenAI’s o-series represent this direction — built to spend more computation on internal reasoning. A reasoning model isn’t one that writes longer answers. It’s one built to grapple with harder problems for longer. 但推理模型改变了这一前提。在推理模型中，思考先于答案。通用的对话模型擅长快速生成自然的回答。对于方向明确的任务（如简短的摘要、简单的解释），这已经足够了。但推理模型的工作方式不同。它们不是直接将问题推向结论，而是被设计用来对比条件、追踪可能的路径，并在形成答案前对问题进行更长时间的推敲。Claude 的扩展思考功能或 OpenAI 的 o 系列模型就代表了这一方向——它们旨在将更多的计算资源投入到内部推理中。推理模型并不是指那些写出更长答案的模型，而是指那些能够更长时间地处理复杂问题的模型。

When Your Old Methods Get in the Way. With general models, “think step by step” can be helpful. It forces intermediate steps rather than jumping to conclusions. But with reasoning models, the same approach doesn’t always work. When you strongly specify an arbitrary sequence of thinking to a model already designed to break problems down, you narrow the space for it to find a better path. The same goes for examples. Good examples show the standard for an answer. But overly detailed examples can lock the model into a specific solution method — even when a superior approach exists. This isn’t about Chain of Thought being wrong. It’s about using the same habits when your tool has fundamentally changed. 当旧方法成为阻碍时：对于通用模型，“一步步思考”可能很有帮助，它能强制模型产生中间步骤，而不是直接跳到结论。但对于推理模型，同样的方法并不总是有效。当你向一个本就设计好如何拆解问题的模型强行指定一套随意的思考顺序时，你反而限制了它寻找更优路径的空间。示例也是如此。好的示例能展示答案的标准，但过于详细的示例可能会将模型锁定在特定的解题方法上，即便存在更好的方案。这并不是说“思维链”（Chain of Thought）是错的，而是说当你的工具已经发生本质变化时，你不应再沿用旧的习惯。

Specify the Goal, Then Step Back. With reasoning models, sometimes saying less is better. Rather than mapping out the process in detail, give: A clear goal, The criteria for a good answer, The output format you need. Then leave the middle steps to the model. The model doesn’t need you to design its thinking process. It needs to know what counts as a good answer. 明确目标，然后退后一步：使用推理模型时，有时少说反而更好。与其详细规划过程，不如提供：一个明确的目标、好答案的评判标准、你需要的输出格式。然后，把中间的步骤留给模型。模型不需要你来设计它的思考过程，它只需要知道什么样的答案才算是一个好答案。

This is an excerpt. The full piece — including a side-by-side prompt comparison and when reasoning models are the wrong tool entirely — is at Dechive. Dechive is a quiet library for the AI age — a place to read slowly, think deeply, and ask why. 本文为节选。完整文章（包括提示词对比分析，以及何时不应使用推理模型）请访问 Dechive。Dechive 是 AI 时代的一座静谧图书馆——一个适合慢读、深思并探究本质的地方。