提问 HN：为什么 AI 编程助手对我没用？
1 分•作者: rich_sasha•6 个月前
我真的在尝试用开放的心态使用它们。我正在编写详细的规范。如果失败，我会调整最初的规范，而不是陷入要求多次调整的螺旋。我正在 Cursor 中使用 Claude Opus 4.5。我的目标也相当低。最近的一次是根据一个相当简单的标准，将一个超大的 Python 文件拆分成几个子模块。它甚至不是失败了，更多的是关于方式。它试图通过编写一些 Python 单行代码来编辑文件，但方式非常笨拙——在很多情况下，它无法编写出语法正确的 Python 代码。
我很纠结，因为我不想成为一个老古板，对着天空大喊“LLM 都是垃圾”，而且很多理性的人似乎用得很好。但我的体验相当糟糕。所以，也许我用错了方法？
公平地说，这不仅仅是失败。我发现它在编写大量良性代码方面相当不错，比如测试、一些简单的工具，我懒得自己写，但能为我节省一些时间。但肯定没有什么了不起的。它也擅长一般性的查询和提设计问题。但它并没有真正做好我作为程序员的工作。
在 Google 上搜索这个话题，大多会得到各种骗子的独家在线课程，这些课程充斥着 AdWord 关键词，教你如何快速致富，或者是一些关于拥有数百个为最新代理超调的存储提示的超优化答案，但我希望在这里得到更高质量的答案。
查看原文
I'm really trying to use them with an open mind. I'm writing detailed specs. On failure, I adjust the initial spec, rather than go down the spiral of asking for many adjustments. I'm using Claude Opus 4.5 inside Cursor. My ambitions are also quite low. The latest was to split a mega Python file into a few submodules according to a pretty simple criterion. It's not even that it failed, it is more about the how. It was trying to action the refactor by writing some Python one-liners to edit the file, in an extremely clumsy way - in many cases failing to write <i>syntactically correct</i> Python.<p>I'm torn, as I don't want to be an old man luddite shouting at the clouds "LLMs are garbage", and plenty of reasonable people seem to do well with them. But my experience is rather poor. So, maybe I'm holding it wrong?<p>It's not only failures, to be fair. I found it fairly good at writing a lot of benign code, like tests, simple tools I wouldn't bother with that save me a few mins here and there. But certainly nothing great. Also good at general queries and asking design questions. But not actually doing my job of being a programmer.<p>Googling the topic mostly yields various grifters' exclusive online courses in no-code get rich quick agents packed with AdWord keywords, or hyper optimised answers about having 100s of stored prompts hypertuned for the latest agent, but hoping for higher quality answers here.