ChatGPT Business: organization-wide instructions for every chat and Codex session

The strongest version of this is probably not one giant hidden system prompt. It should be an auditable instruction stack with scoped precedence. I would separate four layers: non-negotiable organization safety constraints, product-specific defaults, repo/project instructions, and user preferences. The important part is not just applying the policy, but showing the effective stack when something behaves differently than expected. For Codex specifically, an “effective instructions” view would help a lot: org policy version, AGENTS.md sources, skill instructions, MCP/tool policy, and user-level preferences. That makes conflicts debuggable and makes security teams more comfortable with rollout.