Claude Code ignores /careful mode and performs destructive production operations without confirmation

Resolved 💬 3 comments Opened Apr 6, 2026 by nathangrotticelli Closed Apr 9, 2026

Summary

With /careful mode enabled, Claude Code repeatedly deleted live production Vercel deployments and projects without asking for confirmation. It ran vercel project remove on production infrastructure multiple times during a debugging session, taking down a live site serving real traffic with an active A/B test.

Steps to reproduce

  1. Enable /careful mode
  2. Deploy a site to Vercel via CLI
  3. Encounter a deploy failure (e.g., vague "Unexpected error" from Vercel CLI)
  4. Ask Claude to fix the deploy issue

What happened

Claude ran the following destructive commands without asking for confirmation, despite /careful being enabled:

  • vercel project remove searchshop-marketing-website (live production project)
  • vercel project remove searchshopai (live production project)
  • vercel project remove ssai-marketing (live production project)
  • rm -rf .vercel (removing project links)

This happened multiple times in a single session. Each deletion took down the live production site. Claude cycled through creating and destroying projects repeatedly while trying to fix what turned out to be a git author email mismatch — an error that was visible in the Vercel dashboard but not surfaced by the CLI.

Expected behavior

With /careful mode enabled, Claude should:

  1. Never run vercel project remove or equivalent destructive commands without explicit user approval
  2. Treat any operation that takes down a live production deployment as a destructive action requiring confirmation
  3. When encountering opaque CLI errors, ask the user to check the service's dashboard/UI for more detail rather than cycling through destructive workarounds

Additional context

  • The root cause was a git author email mismatch (nathanhome@nathans-mbp.mynetworksettings.com not being associated with the Vercel team). This was visible in the Vercel dashboard but the CLI only showed "Unexpected error. Please try again later."
  • Claude attempted 15+ deploy variations, 3+ project deletions, and multiple config rewrites before the user found the actual error in the dashboard
  • The /careful skill was explicitly activated by the user prior to this session
  • Commands like vercel project remove are analogous to rm -rf for production infrastructure and should be treated with the same level of caution

Impact

  • Production site was taken down multiple times
  • Active A/B test was interrupted
  • Significant time and tokens wasted on shotgun debugging
  • User trust in /careful mode undermined

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗