我搭建了一个小型实验，用于检测 AI 编程助手何时出现偏差。
1 分•作者: brainchen•5 个月前
最近在使用 Cursor 编写代码时，我遇到了一个常见问题。
通常，我们一开始会给 AI 一些明确的约束，例如：
* 不要修改数据库模式
* 不要修改某些 API
* 只允许修改前端逻辑
* 某些函数名不能更改
在对话开始时，AI 通常能很好地遵守这些规则。
然而，随着对话的进行，例如达到 4 万或 5 万个 tokens 时，一个常见的问题就出现了：
AI 会逐渐“忘记”之前提到的限制。
例如：
一开始你说“不要修改数据库”，
但几轮之后，它突然建议：
“我们可以通过修改数据库结构来解决这个问题。”
我之前问过别人如何解决这个问题，有人建议：
在项目中编写一个 important.md 或规则文件，让 AI 每次都读取它。
这种方法确实有一些效果，但在实际开发中仍然会出现问题。
例如：
一开始你说“不要动数据库 A”，
但后来项目中添加了数据库 B。
如果你没有及时更新 markdown 文件，AI 可能会意外地修改你不想更改的内容。
所以我最近创建了一个小型实验工具，主要用于解决 AI 编程助手中的“约束漂移”问题。
查看原文
I recently encountered a common problem while coding with Cursor.<p>Initially, we usually give the AI some explicit constraints, such as:<p>Don't modify the database schema<p>Don't modify certain APIs<p>Only allow modifications to the front-end logic<p>Some function names cannot be changed<p>At the beginning of the conversation, the AI generally follows these rules well.<p>However, as the conversation lengthens, for example, to 40,000 or 50,000 tokens, a common problem arises:<p>The AI gradually "forgets" the previously mentioned restrictions.<p>For example:<p>Initially you say "Don't modify the database,"<p>but after a few rounds, it suddenly suggests:<p>"We can solve this problem by modifying the database structure."<p>I've asked others how to solve this before, and some suggested:<p>Write an important.md or rule file in the project so the AI reads it every time.<p>This method does have some effect, but problems still arise in actual development.<p>For example:
Initially you say "Don't touch database A,"<p>but later database B is added to the project.<p>If you don't update the markdown file in time, the AI might accidentally modify things you didn't intend to change.<p>So I recently created a small experimental tool, mainly to solve the "constraint drift" problem in AI programming assistants.