The Auto Approve Discipline Kit is a governance framework designed for developers running Claude Code and other terminal-based AI agents in unattended or long-running modes. The core problem it addresses is the tradeoff between auto mode, where agents can perform dangerous operations, and manual mode, where developers must approve every tool call individually. This skill provides configuration patterns and mental models that sit between those extremes.
The framework implements a strict safety layer specifically targeting high-risk operations: accidental file deletions, forced git pushes, and secret leakage. By applying hardened deny-rules, developers can dispatch autonomous agent chains lasting up to 90 minutes without manually approving each of the 50 or more tool calls such a session might generate.
A key component is the 6-layer rollback ladder, which provides structured recovery paths if an agent operation goes wrong. This turns Claude Code from an interactive pair-programmer requiring constant confirmation into an autonomous operator that can be left to run complex, multi-step tasks safely.
This is a skill — it delivers configuration patterns and governance models rather than a running server or callable API. There are no environment variables to configure and no transport layer involved. It is intended for developers who already work with Claude Code or similar terminal-based AI agents and want to harden their automation without sacrificing the ability to run long, uninterrupted workflows.