What problem does it solve? AI coding agents can silently run destructive git commands like git reset --hard or git push --force, permanently destroying uncommitted work or rewriting shared history. Prompt instructions alone fail the one time it matters; this Skill installs a hook that enforces the block at the tool level. ## Core Features & Use Cases - PreToolUse Bash hook: Intercepts every Bash command, matches it against a blocked-pattern list, and exits with code 2 to stop execution and feed the reason back to the agent. - Two-band blocked list: Band A covers irreversible operations (hard reset, clean -f, force-push, branch -D, filter-branch, reflog expire, rm -rf .git); Band B covers external publishes (any git push) requiring confirmation. - Global or project install: Writes the hook script to ~/.claude/hooks/ and merges into settings.json without overwriting existing hooks, with a mandatory verification step to catch fail-open misconfigurations like missing jq. - Use Case: After an agent nukes uncommitted work with a hard reset, install this guard globally in about five minutes so every future destructive command is blocked and the agent proposes a reversible alternative like git stash or a WIP commit instead. ## Quick Start Install a global Claude Code PreToolUse hook that blocks dangerous git commands like hard reset, force-push, and clean -f, then verify it blocks a test command.
