Safety Model¶

Safety should be layered. No single prompt, hook, policy, or review step is enough.

Prompt Guidance¶

AGENTS.md and skills describe expected behavior. They are useful for shaping work, but they are not enforcement.

Policy¶

Policy files make approval, sandboxing, verification, and git-risk expectations explicit. They should be conservative for important repositories.

Hooks¶

Hooks are deterministic scripts that can block likely secrets, dangerous commands, or unverified stop events. The examples here are starting points, not hardened production systems.

Sandboxing¶

Sandboxing limits the blast radius of mistakes. Prefer restricted write access and explicit approval for network, destructive, or credential-sensitive operations.

Review¶

Human and automated review should check behavior, security, migrations, and documentation. Review is especially important for changes that touch authentication, permissions, data deletion, billing, or deployment.

Git Isolation¶

Use branches and worktrees to isolate changes. Avoid force pushes and hard resets unless there is an explicit, reviewed reason.