Destructive Command Safety Guard (Careful)

Core Concepts / How It Works

List of Monitored Commands

Careful mode intercepts commands in the following categories before they execute.

Category	Example Commands	Risk
File system	rm -rf, rmdir /s, truncate	Permanent deletion of files/directories
Database	DROP TABLE, DELETE FROM (no WHERE), TRUNCATE	Permanent data deletion
Git	git push --force, git reset --hard, git clean -f	Permanent removal of commits/changes
Containers/Clusters	kubectl delete, docker system prune	Service shutdown, image deletion
Processes	kill -9 PID, pkill -f	Forced process termination

When to Use

• When accessing a production server or production database
• When debugging a live system directly or deploying a hotfix
• When working in a shared environment (shared staging, shared branch)
• Before a large data migration or schema change operation
• When you want a safety net before delegating to Claude in a situation where "a mistake would be really bad"

Warning → User Choice Flow

Claude: This command cannot be undone:
        rm -rf ./uploads/
        Do you want to continue? [y/N/explain reason]

User: y  → Execute
User: N  → Cancel
User: "I actually only want to delete the temp folder" → Suggest command modification

User Override Allowed

A warning is a confirmation request, not a "forced stop." If the user knows they are intentionally performing a dangerous operation, they can individually override (y) each warning. The skill trusts the user but acts as a safety net to reduce the chance of mistakes.

How to Activate Careful Mode

# Activate Careful mode for this session
> be careful  # or
> careful mode  # or
> safety mode  # or
> prod mode (explicitly stating it's a production environment)

One-Line Summary

A safety guard skill that displays a warning and requires user confirmation before executing commands that are difficult or impossible to undo, such as rm -rf, DROP TABLE, git reset --hard, force push, and kubectl delete.

Getting Started

/careful

SKILL.md location: ~/.claude/skills/careful/SKILL.md

To customize, copy and modify the SKILL.md content.

Practical Example

Scenario: You deployed a "Student Club Notice Board" project built with Next.js 15 + TypeScript to Supabase. Test data got mixed into the production DB, and you want to clean it up before delegating to Claude, so you turned on Careful mode.

Scenario 1: Deleting Test Data from Production DB

> Turn on careful mode. I want to delete test notice data from the production Supabase DB.

Claude's behavior:

Warning: Careful mode activated. I will confirm before any destructive operation.

About to execute the following query:
DELETE FROM notices WHERE title LIKE '%test%';

Important:
- This operation cannot be undone.
- Affected rows: approximately 12 (from SELECT COUNT(*) result)
- Targeting the production DB.

It is recommended to back up before proceeding.
Do you want to continue? [y/N]

Scenario 2: Accidentally Dangerous Git Command

> Revert the latest commit on the main branch.

Claude's behavior:

Warning: Destructive Git operation detected:
git reset --hard HEAD~1

This command permanently removes the last commit from the main branch.
reset --hard on a shared branch (main) affects team members.

Safe alternative:
  git revert HEAD  (adds a new reverting commit, preserves history)

Do you want to use the alternative? [y(original command)/r(use revert)/N(cancel)]

Scenario 3: Deleting the Entire Uploads Folder

// Code Claude was about to execute
import { execSync } from 'child_process'
execSync('rm -rf ./public/uploads/test-*')

When Careful mode is active:

Warning: Destructive file system operation:
rm -rf ./public/uploads/test-*

List of files to be deleted:
  - test-notice-001.jpg (2.3MB)
  - test-banner.png (1.1MB)
  - ... and 8 more files

11 files will be permanently deleted.
Do you want to continue? [y/N]

Learning Points / Common Pitfalls

• The trap of "I don't make mistakes": Everyone makes mistakes when working in production while tired or losing focus. Careful mode provides a systematic safety net instead of relying on personal concentration.
• Avoid the habit of ignoring warnings: If warnings appear too frequently, they get ignored. It is better to turn off Careful mode in the development environment and only turn it on when accessing production/staging environments.
• Utilize the safe alternative suggestion feature: The core value of the Careful skill is not simply "don't do it" but suggesting safe alternatives like git revert. Reading the alternative is itself a learning opportunity.
• Common mistake: git push --force may be acceptable on a personal feature branch, but must never be used on main. Careful mode weighs risk based on the branch name.
• Next.js 15 perspective: Clearing the build cache (rm -rf .next/) is safe in development but can cause temporary service downtime when run on a production server. Careful mode catches this kind of context.

Related Resources

• guard — Activate Careful + Freeze simultaneously (maximum safety mode)
• freeze — Restrict edit scope (block file modifications outside a directory)
• cso — Security audit (OWASP, STRIDE threat modeling)

---

Field	Value
Source URL	https://docs.anthropic.com/en/docs/claude-code/skills
Author / Source	Anthropic
License	Commentary MIT, original for reference only
Translation Date	2026-04-13