Write AGENTS.md with Clear Directory Boundaries

Codex combines global and project guidance, walking toward the working directory; nearer guidance is applied later, and override filenames take priority at a directory level. Check the current official loading rules and restart or verify loading after changes. This is not identical to CLAUDE.md behavior.

Inventory existing guidance before writing a new root file. Identify which facts apply everywhere and which apply only to a subtree. Do not replace a teammate’s instructions merely because your generated draft looks cleaner.

The root draft says the site is static, new dependencies need approval, and live delivery is out of scope. A forms-directory note can add a focused rule about persistent labels without repeating the entire root file.

• Root: preserve existing user changes; use documented checks; ask before adding dependencies or sending messages.
• Forms subtree: maintain visible labels and keyboard order for each control.
• Task brief: add the missing email label now. This belongs to the current task, not the permanent root policy.
• Unknown test command: leave it unspecified until the project documentation establishes it.

1. Draft AGENTS.md using inspected facts, optionally with the local Agent Instructions Builder.
2. Review the full chain from project root to the intended working directory, including any override files.
3. Remove contradictory or stale instructions and confirm ownership before changing an existing policy.
4. Start a new session as appropriate and ask the agent to identify the relevant guidance before a harmless inspection.

The next worker should know where it may edit, what needs permission, and how completion is checked. A nested rule narrows a task’s context; it should not quietly erase a safety boundary or pretend to enforce filesystem access.

Pause if two instruction files disagree about deployment, if a nested instruction broadens sensitive access, or if the expected file is not loaded. Resolve the guidance chain before approving modifications.

• Can each rule be traced to the directory where it applies?
• Are commands real rather than generated defaults?
• Does a fresh session recognize the intended constraints?