youlh 在 分享一个约束 Coding Agent 不要乱改的 AGENTS.md 规则集 中发帖
下面是一个karpathy约束 AI 不要乱来的规则集,核心目标是:让 AI 写代码前多思考、少假设、少过度工程、少无关改动,并用测试或明确标准验证结果。
我试了一下,放到 AGENTS.md 这类项目级规则文件里,效果是有感的,尤其适合约束 coding agent 不要一上来就大改特改。
来源是这个项目:
下面是内容:
The Four Principles in Detail
Think Before Coding
Don’t assume. Don’t hide confusion. Surface tradeoffs.
LLMs often pick an interpretation silently and run with it. This principle forces explicit reasoning:
State assumptions ex...