Approval Interrupts -- Automate Right Up to the Line

The useful boundary is not human versus automated. It is reversible versus consequential.

Approval Interrupts keeps an agent working through research, drafting, local preparation, and validation, then pauses immediately before an action can affect money, people, production, public state, access, or recoverability.

What you get

• A workflow map split into observe, prepare, commit, and verify actions
• Approval checkpoints placed at the actual consequence boundary
• Approval cards naming the exact action, target, payload summary, consequence, cost, and recovery path
• Rules for when approval expires because a recipient, price, environment, payload, or destructive class changed
• One-shot execution followed by external-state verification
• Safe handling for uncertain outcomes so charges, sends, publishes, and deletes are not repeated blindly
• Clear distinctions among drafted, approved, attempted, sent, delivered, published, deployed, and verified

This skill avoids two expensive failures: agents that act on vague enthusiasm, and agents that stop to ask permission before every harmless read. It keeps preparation moving while preserving human control over consequential actions.

Use it to harden messaging, marketplace publishing, code delivery, deployments, purchasing, credential management, and destructive maintenance workflows. It does not grant authority by itself; it makes authority specific, visible, and auditable.