优化 AGENTS.md：Claude Code 与 Codex 上下文管理技巧

Highly recommend slimming down your AGENTS.md. Claude Code's docs say to keep each one 𝘂𝗻𝗱𝗲𝗿 𝟮𝟬𝟬 𝗹𝗶𝗻𝗲𝘀, since the longer it gets, the less Claude follows it. Codex reads 𝟯𝟮 𝗞𝗶𝗕 𝗮𝘁 𝗺𝗼𝘀𝘁 by default and skips anything past that.

Simply put, every time you start a session, Codex and Claude Code read your AGENTS.md into context, and it comes along on every turn after that. The longer it is, the more context it eats. After slimming, it looks like the image: only the main file gets read every time, and the rest only when the agent gets to that folder or needs it.

The official docs have a few more things I didn't know: → Codex reads them once, when the session starts: the global one, plus every one from the project root down to the folder you launched in. It won't automatically reread them when you move to another folder → Claude Code works differently: it loads a subfolder's AGENTS.md only when it reads a file in that folder. So rules that only matter for one folder belong in that folder's own AGENTS.md → In Claude Code, @ imports only make the file look tidier. Imported files still load in full at launch, so they 𝘀𝗮𝘃𝗲 𝘇𝗲𝗿𝗼 𝘁𝗼𝗸𝗲𝗻𝘀 → In Claude Code, every installed skill's name and description takes up context on every turn, whether you use it or not. Run /skill-doctor to see what each one costs and how often it gets used, and turn off the ones you don't need → In Claude Code, run /doctor prompt-audit (v2.1.283 or later). It finds instructions written for older models, references to files that don't exist, and files that contradict each other, and changes nothing until you say yes

Send this prompt and the links below to Opus 5.5 👇

"Read these two docs, then read my AGENTS.md (and CLAUDE.md, if the project has one) and the files it imports, and give me three ways to restructure them: 1. Delete instructions that do nothing: things Opus already does, and things it can tell from the code. Merge duplicates, and list any contradictions so I can choose. 2. Keep only the rules needed every time in the root AGENTS.md, under 200 lines. Move rules that only matter for one folder into that folder's own AGENTS.md, and other details into separate files. For both, say in the main file where they are and when to read them. Don't use @ imports for detail files, and keep the line in CLAUDE.md that imports AGENTS.md. Have three subagents each write one version, each more aggressive than the last. This step is proposals only, so don't change any files. For each version, list what was deleted, where things moved, and the main file's line count and size in KiB before and after. Don't lose a single useful rule. Show me the three versions side by side first, I'll pick one, and then open a PR." https://code.claude.com/docs/en/memory https://learn.chatgpt.com/docs/agent-configuration/agents-md