The original title is "Let Your AI Fix Its Own Broken Automation: Building an Unattended Watchdog"

Short summary

A detailed guide to building an unattended watchdog that lets Claude detect, repair, and verify broken automation scripts without human approval. The system separates health checks from post-repair verification, only deploying when verification passes, and rolling back otherwise. The author emphasizes never trusting AI self-reports and stacking multiple guardrails for safe autonomous operation.

• •Watchdog pattern: detect → Claude repairs → independent verify → deploy only on pass → rollback on fail
• •Critical design: health.sh and verify.sh are separate scripts run independently of Claude's repair
• •Seven stacked guardrails enable safe unattended AI repair of cron/launchd automation scripts

Generated with AI, which can make mistakes.