behavior-hooks

behavior-hooks

Stop telling Claude the same thing twice.

A Claude Code plugin that detects when you're unhappy with something Claude did, and writes an enforceable hook to make sure it never happens again.

The Problem

You've written your CLAUDE.md. You've been clear about your rules. And yet:

• Claude deploys to production when you said staging only
• Claude adds Co-Authored-By to commits when you said not to
• Claude runs git add . when you told it to stage specific files
• Claude adds comments to clean code that didn't need them
• Claude amends commits when you wanted a new one

You correct it. It apologizes. It does it again next session.

CLAUDE.md is a suggestion. Hooks are the law.

The difference between "please don't do this" in a markdown file and a PreToolUse hook is the difference between a polite sign and a locked door. One relies on Claude reading, understanding, and consistently applying your preferences across every session. The other runs a shell command that blocks the action before it happens.

How It Works

1. You get frustrated with something Claude does
2. The plugin detects your correction (it watches for phrases like "don't do that", "why did you", "stop doing X", "I told you not to")
3. It acknowledges concisely (no groveling)
4. It offers to write a machine-enforced hook that blocks the behavior
5. You say yes, the hook is written, tested, and installed
6. Claude literally cannot do the thing again - the hook denies the tool call before it executes

What Can It Hook?

Behavior	Hook Type	Example
Deploying to the wrong environment	PreToolUse (MCP)	Block fleet_deploy for production app
Pushing to protected branches	PreToolUse (Bash)	Block git push origin main
Running dangerous commands	PreToolUse (Bash)	Block rm -rf, DROP TABLE, etc.
Modifying files it shouldn't	PreToolUse (Write/Edit)	Block writes to config files
Using git add .	PreToolUse (Bash)	Block broad staging commands
Adding unwanted code patterns	PostToolUse (Write/Edit)	Warn on added comments, console.logs
Skipping safety hooks	PreToolUse (Bash)	Block --no-verify

For things that can't be hooked (style preferences, communication tone), it saves a memory instead - but it's honest about the difference.

Install

As a Claude Code plugin

Add to your settings:

{
  "enabledPlugins": {
    "behavior-hooks@wrxck-claude-plugins": true
  }
}

Or if you already have the wrxck-claude-plugins marketplace configured, enable it from the plugin menu.

From source

gh repo clone wrxck/behavior-hooks

Why Not Just Use CLAUDE.md?

	CLAUDE.md	behavior-hooks
Enforcement	Advisory - Claude reads it and tries to follow it	Mandatory - hook runs before the action, blocks if violated
Persistence	Claude may forget across sessions or after context compaction	Hooks persist in settings files, survive every session
Specificity	Natural language, open to interpretation	Exact tool name + pattern match + shell condition
Feedback loop	You notice the violation, correct it, hope it sticks	You correct once, the hook prevents recurrence forever
Scope	Per-project	Per-project or global, your choice

CLAUDE.md is still valuable for guidelines, preferences, and context. But for hard rules - things Claude must never do - hooks are the only reliable mechanism.

Philosophy

• Correct once, enforce forever. Every correction should produce a durable guardrail, not a temporary apology.
• Hooks over promises. If Claude can be blocked from doing something, it should be blocked - not asked nicely.
• Honest about limits. Some things can't be hooked. The plugin tells you when a memory is the best it can do.
• No groveling. One-line acknowledgment, then action. Respect your time.

License

MIT