Three-tier CLAUDE.md rules cut AI approval prompts to money and contracts

The post starts from two everyday failures with Claude Code: an agent that asks permission for everything and stalls, and one that proceeds without asking and performs an irreversible operation. The author traces both to the same cause — never having decided how far the agent may go — and argues the fix is to write that boundary down in advance.

The practical shape is three tiers in CLAUDE.md. The first lists what may run without confirmation, covering code edits, commits, production deploys, and production database reads and writes. The second narrows mandatory questions to operations that move money or create a new commitment, on the reasoning that amounts and contracts cannot be undone later. The third sits above both: bulk production updates require checking the target count and taking a backup, API calls require estimating item counts against a monthly budget, and config files must be compared with the server's live version before overwriting.

From running this, the author reports fewer interruptions, since questions now come only for money and contracts. The scope is widened after an incident rather than up front, added together with safety steps, and hooks are used for anything that must be stopped mechanically. When Claude Code says it cannot perform a heavy operation such as handling production credentials, the author treats that as the boundary working, not something to bypass.