What to know: a code tool runs somewhere. It is pointed at a directory on your computer, and that is the world it can see. Part of knowing your setup is knowing where you are pointed.
There are two versions of Claude Code: the desktop app, and the terminal version. Some things only exist in the terminal: the status line, and starting it with --dangerously-skip-permissions.
Approval modes. The AI asks before it acts, and how much it asks is a setting. Manual: it asks for everything. Plan: it works out what it would do and shows you before touching anything. Auto: it acts without asking, within limits. In the terminal, Shift+Tab cycles between them. Skipping permissions entirely is a separate, deliberate switch for when you know what you're doing.
Author's practice: one working directory for everything, and the AI manages the file system inside it, sectioned by domain. Always execute against one file system. It is simpler.
Reveal on the spot: say which directory you are pointed at, which version this is, and which approval mode is on. If the person has several working directories, say so, and say what one would look like.
Unlocks: knowing what the AI can reach, and how much it will ask before it acts.