Skip to content- Sources: primary
- Summary: Hallie says CLAUDE.md should keep only facts Claude cannot see in the code, and a doctor prompt audit command checks CLAUDE.md and skills against the current model. She cites an Anthropic study where about 14% of people caught a clearly dangerous command while auto mode blocked 89%, so guarantees belong in deny rules, which are checked before the classifier, and in the OS sandbox, which also binds child processes such as npm test scripts. She recommends hooks that run tests at task and session end and exit with code 2, CI autofix, cloud sessions with network secrets, and Claude Code projects, a beta where a Fable orchestrator starts Opus 5.5 thread sessions.
- Why it matters: It gives a concrete checklist for running long unattended Claude Code sessions safely, including the exit code detail that makes hooks actually block completion.
- Background: Auto mode is a Claude Code permission mode in which a separate classifier approves or blocks tool calls that go beyond reading and editing project files. Hallie says Anthropic went from about 2,500 changes a day in May to more than 30,000 a day in September as engineers moved from approving each step to managing dozens of parallel sessions.