Many people think they need a better model. What they actually need is fewer windows. If you use GPT or Claude and still bounce between browser, chat, and editor, this is the miss: in Codex, the most valuable prompt lives in your repo [C001].
You search in one tab, paste background into a chat, then jump back to the editor to change three lines. That loop is the real tax. AI tools are not only competing for code output. They are competing for the scraps of time you lose every time you re-explain context.
So the wrong move is treating Codex like the same category as a chat box and comparing only model quality. If you do that, you keep hand-carrying context right where the tool could have absorbed it for you.
AGENTS.md is just a plain repo file that tells the coding agent how to work. Codex docs say it reads that file before starting, then merges rules from global, repo, and subfolder levels. That turns prompting from one-off wording into shared project setup.
GitHub review follows the same pattern. It looks for AGENTS.md and uses its review rules, so the instruction is no longer trapped in one person's chat history. Same repo, same file, same place to update the team's default behavior. Anchor: Changelog – Codex | OpenAI Developers [C001].
The prompt did not die. It moved. Put the repeatable rules in AGENTS.md, keep them in version control, and let the chat box handle the temporary stuff. Boundary: docs-based, not a hands-on benchmark. Share this with the person still re-pasting setup into every new chat.