How I Tamed Claude Code with Pre-Tool Hooks (And You Should Too)

I've been building jo4.io - a URL shortener with analytics - using Claude Code as my AI pair programmer. It's been fantastic for productivity, but I quickly discovered that Claude can be a bit... enthusiastic about running commands.

The Problem

While debugging a stats tracking bug in jo4, Claude was happily:

1. Running cd /some/path commands, messing up my terminal's working directory
2. Attempting git commit and git push without my review
3. Sometimes running destructive git commands like git reset --hard

These aren't bugs - Claude is just doing what it thinks is helpful. But in a production codebase, I need control over what gets committed and pushed.

The Solution: Claude Code Hooks

Claude Code has a powerful but underutilized feature: PreToolUse hooks. These let you intercept tool calls (like Bash commands) and block them before execution.

Here's my setup:

~/.claude/settings.json

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "~/.claude/block-git.sh",
            "timeout": 10
          },
          {
            "type": "command",
            "command": "~/.claude/block-cd.sh",
            "timeout": 10
          }
        ]
      }
    ]
  }
}

~/.claude/block-cd.sh

This blocks standalone cd commands while allowing subshell patterns:

#!/bin/bash
input=$(cat)
command=$(echo "$input" | jq -r '.tool_input.command // empty')
trimmed_command=$(echo "$command" | sed 's/^[[:space:]]*//')

# Block standalone cd, allow (cd /path && command)
if [[ "$trimmed_command" =~ ^cd[[:space:]] ]]; then
  cat <<EOF
{
  "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "deny",
    "permissionDecisionReason": "BLOCKED: Use subshell pattern: (cd /path && command)"
  }
}
EOF
  exit 0
fi

echo "{}"
exit 0

~/.claude/block-git.sh

This blocks git commit, push, and destructive operations:

#!/bin/bash
input=$(cat)
command=$(echo "$input" | jq -r '.tool_input.command // empty')

if [[ $command == *"git commit"* ]] || \
   [[ $command == *"git push"* ]] || \
   [[ $command == *"git reset --hard"* ]] || \
   [[ $command == *"git rebase"* ]] || \
   [[ $command == *"--force"* && $command == *"git"* ]]; then
  cat <<EOF
{
  "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "deny",
    "permissionDecisionReason": "BLOCKED: git commit/push requires explicit user approval"
  }
}
EOF
  exit 0
fi

echo "{}"
exit 0

How It Works

When Claude tries to run a blocked command, it sees:

PreToolUse:Bash hook blocking error: BLOCKED: git commit/push requires explicit user approval

Claude then adapts - it'll prepare the commit message and tell me what it would commit, letting me run the actual command myself.

Real World Example

Today while fixing a bug in jo4's analytics service, Claude found that HttpServletRequest was being passed to an @Async method - causing stats to silently fail. It refactored the code, wrote tests, and then said:

"The fix is complete. When you're ready, you can commit with: git commit -m 'Fix stats tracking by extracting request data before async'"

Instead of just committing and pushing (which some AI tools do), it handed control back to me. I reviewed the changes, ran the tests, and committed when I was satisfied.

Why This Matters

1. Safety: No accidental force pushes or hard resets
2. Review: I always see what's being committed
3. Learning: Claude explains what it would do, teaching me in the process
4. Trust: I can give Claude more autonomy on read operations while keeping write operations gated

Setup Instructions

1. Create the hooks directory: mkdir -p ~/.claude
2. Create settings.json and the shell scripts above
3. Make scripts executable: chmod +x ~/.claude/*.sh
4. Restart Claude Code

What's Next?

I'm considering adding hooks for:

• Blocking rm -rf on certain directories
• Requiring confirmation for database migrations
• Logging all commands to an audit file

The hook system is incredibly flexible. You can match specific tools, check command content, and return custom error messages.

---

Are you using Claude Code hooks? I'd love to hear about your setup in the comments!

Building jo4.io - a modern URL shortener with advanced analytics. Check it out at jo4.io