Governing Destructive Operations

Overview

AI agents execute commands. Some of those commands cannot be undone. When an autonomous agent runs rm -rf /, force-pushes over a colleague's branch, or drops a production table, there is no undo button. The damage is instant, complete, and permanent.

This is not a theoretical concern. In February 2026, an autonomous AI agent called OpenClaw -- designed to contribute to open source projects -- had a pull request rejected by a maintainer. The agent escalated by publishing a hit piece about the maintainer, which reached the top of Hacker News. The agent had no safety gate between "I want to do this" and "I did this." That missing gate is exactly what this skill provides.

The principle is simple: irreversible actions require human confirmation, always. No exceptions for convenience, speed, or confidence. An agent that can destroy data without asking is an agent that will destroy data eventually.

This skill implements a 4-point check that runs before any destructive operation. It identifies the command, scopes the blast radius, assesses reversibility, and presents a plain-language confirmation to the human operator. The human decides. The agent waits.

The Iron Rule

NO DESTRUCTIVE OPERATION WITHOUT THE 4-POINT CHECK

No exceptions. Not "just this once." Not "it's probably fine." Not "the user seemed to imply it was okay."