你刚刷到这条消息，本来准备顺手划走，但又怕自己错过了真正会影响下一步判断的那一点。
最容易做错的，是Better Models: Worse Tools；代价往往是如果只盯表面热闹，你很容易在错误方向上花掉时间、预算和注意力。；我先给一个保守判断：模型升级后，仓库说明书比提示词更值钱。
Picture the moment: you see the announcement, your thumb is already moving, and then you stop because you are wondering whether this changes your next move or is just more noise. That is where people make the wrong call: Better Models: Worse Tools. If you only watch the surface hype, you can waste time, budget, and attention in the wrong place.
My conservative take is simple: after model upgrades, repo instruction files are worth more than prompt libraries.
More specifically, the most valuable code file in the repo may now be AGENTS.md. For beginners, that is the file that tells an agent what this project is, which commands to run, what rules not to break, and how to move inside the codebase.
Why I think that: one study of 253 Claude.md files found that these files are not generic notes. They are packed with commands, implementation guidance, and architecture rules. Another signal is adoption: MCP has passed 10,000 servers, and AGENTS.md has reportedly been adopted by 60,000+ projects. At that point, this is not “prompt craft.” It is 工作流程（workflow） infrastructure.
The catch is just as important. Bigger context is not automatically better. Another paper found that repo-level context files can materially change agent behavior, raise reasoning cost by 20%+, and often reduce success when the file is bloated or stale.
真正该讨论的是：Better Models: Worse Tools