SAFE / SHIP
LANGUAGE English
Safety check

04 / GUIDE

PERMISSIONS / BOUNDARIES

AI permissions: start small

How to define the area an AI agent can touch before Claude Code or Codex starts working.

01

Keep the working directory inside the project you intend to change.

02

Deny reads to .env files, secret folders, private keys, and personal data.

03

Require a human gate for deletion, publishing, sending, purchasing, or permission changes.

Run the five-question safety check before you delegate.

Start the safety check →