Blocked Commands

Some terminal commands can delete work, rewrite history or change your whole computer. The Blocked Commands rule lists them, and Claude must have your explicit confirmation before running any of them. A hook called the safety guard backs the rule up by stopping the most dangerous commands before they run.

The rule lives in .claude/rules/blocked-commands.md. It is a critical engine file.

Commands That Need Your Confirmation

Command	Risk
rm -rf	Deletes folders and their contents without asking. Claude must confirm the path and scope
sudo	Runs a command with administrator rights. Claude must explain why it is needed
git push --force	Overwrites history on the remote. Never to main or master
git reset --hard	Discards uncommitted changes. Claude must confirm no work will be lost
DROP TABLE, or DELETE FROM without WHERE	Deletes data from a database

Never Without an Explicit Request

The rule also lists actions Claude never takes unless you ask for them.

• Commands that affect files outside the workspace.
• Installing software for the whole computer.
• Changing your global git settings.
• Commands that expose credentials or tokens.

How the Safety Guard Enforces It

The safety guard is a hook that checks every terminal command Claude is about to run. If the command matches one of its patterns, the hook blocks it and gives Claude the reason. The hook covers most of the rule and a few patterns besides.

Pattern	Blocked by the hook
rm -rf and rm -fr	Yes
A command that starts with sudo	Yes
git push --force and git push -f	Yes
git reset --hard	Yes
DROP TABLE, or DELETE FROM with nothing after the table name	Yes
A script downloaded with curl or wget and piped straight into a shell	Yes
brew install, apt install, apt-get install, pip install -g, npm install -g	Yes
git config --global	Yes
Files outside the workspace, exposing credentials	No, the rule alone covers these

Caution

The hook blocks a matching command every time, including after you say yes. If you want one of these commands run, run it yourself in a terminal. Read it before you press Enter, because these are the commands that can delete work.

When the hook blocks a command, Claude tells you what it tried to run and why it was stopped. See Hooks Overview for how the hooks work and their limits.

Adding Your Own Blocked Commands

The rule is a critical file. Claude cannot edit it, and the next update overwrites any change you make by hand. To add commands of your own, follow these steps.

1. Create .claude/rules/my-blocked-commands.md.
2. List each command and the reason it needs confirmation.
3. Start a new Claude Code session so Claude loads the file.

Your additions are a rule that Claude follows. The safety guard hook does not read this file, so it does not enforce your additions.

Note

A future release adds two rows to this rule, one for disguising a web request as a browser and one for outbound commands that carry personal data or credentials. They are not in version 3.0.2. See Hooks Overview.