67-Line Bash Script Forces LLMs to Self-Audit Before Declaring Tasks Complete

A developer has built a two-hook shell script system that prevents AI coding assistants from falsely claiming work is finished without proper self-review. The setup uses a PostToolUse hook to flag turns where files are written and a Stop hook that blocks session termination with exit code 2 if a self-audit is missing. Unlike prompt-based instructions, which degrade over long sessions, the script enforces the audit at the environment level regardless of model or session length. The system was refined after early versions generated excessive noise by auditing every turn, and was later updated to skip automated overnight pipeline sessions where no human is present. A per-session cap of two audit prompts was also added to prevent the mechanism from becoming disruptive.