Keep shared project instructions in one place
Use AGENTS.md for repository instructions that apply to all your coding assistants. Put the real test command, file ownership, and any paths that require special care in that file. For a monorepo, check what each tool loads when work starts in a subdirectory. OpenAI's project instruction guide
Check what Claude Code already loads
Current Claude Code can read project instructions directly from AGENTS.md. Its default selection can prefer CLAUDE.md when that file exists in the working directory or above it. Check Project instructions in /config and see which files are loaded in your project. Older installations and sessions with the built-in plugin disabled may still need an import. Claude Code's instruction and memory documentation
Import the shared file when needed
If your Claude Code setup uses CLAUDE.md, place the two files next to each other. Put this on its own line in CLAUDE.md:
@AGENTS.md
Add tool-specific instructions below the import only where the tools really need different behavior. Keep the shared commands and repository rules in AGENTS.md so each change has one source.
Verify from the directories where you work
Start a fresh session at the root and in one representative subdirectory. Ask each assistant which test command it would run and which files it would avoid editing for a small example task. Compare the replies with the repository instructions. When they differ, check the working directory, file precedence, and the tool's loaded-instructions view before adding more rules.
This edition was prepared with an AI agent. The technical references were checked during editing.
Top comments (0)