你一会儿在浏览器搜资料，一会儿回聊天框补背景，一会儿再回编辑器改几行代码。
最容易做错的，是把 Codex 当成同一种工具，以为谁分高谁就适合自己。；代价往往是如果把它们都当成同一种工具，你会在最该省事的地方继续手动搬运上下文，最后多一轮返工。；我先给一个保守判断：Codex时代，最值钱的提示词在仓库里。。
The cost is not abstract. It looks like this: you search in the browser, jump back to a chat window to restate context, then return to the editor to change a few lines of code. You are still doing the glue work yourself. AI tools are starting to compete for more than coding tasks. They are competing for the broken minutes lost to constant switching.
My main takeaway from the current Changelog – Codex | OpenAI Developers is simple: the most valuable prompt in the Codex era lives in the repo. [C001]
What the post is really pointing to is not a cosmetic product update. Codex reads instructions before it starts work, and those instructions can come from three layers: global, repo, and subdirectory. That matters because the reusable context is no longer trapped in a chat box. It becomes part of the working environment.
That same repo instruction file affects more than one surface. The GitHub docs say code review looks for AGENTS.md and uses its review guidance to decide what to focus on. The Memories docs make a similar point from another angle: if a team rule should always apply, it belongs in repo docs like AGENTS.md, not in memory as a convenience layer.
So no, this is not “prompting is dead.” It is prompting turned into versioned team infrastructure. For people who are tired of repeating background, re-explaining constraints, and moving context between windows, that is a much more useful shift than another abstract model comparison.
Many people think they need a stronger model. Often they just need fewer windows.
Boundary: this is a restrained read based on the current changelog plus the AGENTS.md, GitHub, and Memories docs, not on a benchmark and not on a full team rollout. So the practical move is also restrained: if you were standardizing Codex this week, start by writing the repo instruction file first, then share this with the person on your team who is still pasting the same context into every tool.
真正该讨论的是：AI 工具真正开始抢的，不只是代码活，而是你来回切换的那些碎时间。