feat: /careful skill — destructive command warnings hook

[josephfung]

Idea

Add a /careful skill that activates a PreToolUse hook intercepting destructive shell commands before they execute. Rather than hard-blocking, the hook returns {"permissionDecision":"ask","message":"..."} so Claude pauses and asks for confirmation — giving the user override capability while preventing accidental damage.

Goal

Catch common high-impact mistakes before they happen: force-pushing, dropping tables, recursively deleting files, pruning containers. Particularly valuable in terminal-heavy workflows or when working near production systems.

Commands to intercept

• rm -r / rm -rf (with safe-exceptions for node_modules, dist, .next, coverage, tmp)
• git push --force / git reset --hard / git checkout .
• DROP TABLE / TRUNCATE
• docker rm -f / docker system prune
• kubectl delete

Implementation approach

1. Add skills/careful/SKILL.md — skill prompt that activates the hook when invoked via /careful
2. Add hooks/check-careful.sh — PreToolUse bash script that:
   • Matches the incoming command against destructive patterns
   • Checks a safe-exceptions list (build artifact dirs, etc.)
   • Returns {"permissionDecision":"ask","message":"Destructive pattern detected: <pattern name>. Confirm?"} on match
   • Logs the pattern name (never the full command) to ~/.claude/sysops/usage.jsonl
3. Register the hook in settings/hooks.json scoped to the skill (not always-on like no-chaining)
4. Update install.sh to handle skill-scoped hook registration

Notes

• Warn-and-confirm (not deny) is the right UX here — the user should be able to override
• Log pattern names only, never command content, for privacy
• Complements the existing no-chaining hook (different concern: safety vs. discipline)