Stop Tuning Prompts. Write AGENTS.md First.

You already know how to use GPT or Claude. Now you are trying to stitch a few AI tools together so you can stop copying the same background into every chat. The easy mistake is treating Codex like just another chatbot and picking it just because someone says it scores higher. If you do that, you keep hauling context by hand and still buy yourself one more round of rework.

The part worth paying attention to is simpler: in the Codex era, the most valuable prompts live in the repo, meaning the project itself, not the chat box. AI tools are no longer just competing for coding tasks. They are competing for the scraps of time you lose switching between a browser, a chat box, and an editor. A lot of people think they need a stronger model. What they really need is fewer windows.

That is why AGENTS.md, the project instruction file, matters more here than one clever chat prompt. The OpenAI Developers docs say Codex reads AGENTS.md before work starts and merges instructions across global, repo, and subdirectory layers. The GitHub review docs say code review can also look for AGENTS.md and use its review guidelines to decide what gets checked.

So this is not "prompt is dead." It is prompt engineering turned into a project file that can travel with the code. If you want to know whether this is just a new shell or a real workflow change, start with Changelog – Codex | OpenAI Developers, then ask whether your standing rules belong in chat or in AGENTS.md. Share this with the person who keeps pasting the same setup into every new thread.