hook-creator（钩子创建者）

Hook Creator

This skill provides comprehensive guidance for creating and managing Claude Code hooks—powerful automation tools that intercept and respond to Claude Code events.

Quick Start

Create a simple PreToolUse hook that validates Bash commands:

# ~/.claude/settings.json or .claude/settings.json
{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/validate-command.sh",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

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

# Block destructive commands
if [[ "$command" =~ ^rm\ -rf ]]; then
  echo "Destructive command blocked" >&2
  exit 2
fi

exit 0

What Are Hooks?

Hooks are event-driven automation points that intercept Claude Code execution at specific moments. They receive JSON input via stdin and communicate back through exit codes and stdout.

Hook Capabilities

• Validate or block tool calls before execution (PreToolUse)
• Auto-approve permission requests (PermissionRequest)
• Run post-processing after tool execution (PostToolUse)
• Add context to conversations (UserPromptSubmit, SessionStart)
• Control session lifecycle (Stop, SubagentStop)
• Handle notifications from Claude Code (Notification)
• Integrate with MCP tools using pattern matching

Hook Types

Command hooks (type: "command"): Execute bash scripts

• Fast, deterministic, ideal for validation rules
• Have access to environment variables
• Return status via exit codes and JSON output

Prompt hooks (type: "prompt"): Use LLM for intelligent decisions

• Context-aware evaluation
• Best for complex, nuanced decisions
• Only supported for Stop and SubagentStop events

Hook Events

PreToolUse

Runs after tool parameters are created but before execution.

Common matchers: Bash, Write, Edit, Read, Task, Glob, Grep

Use cases:

• Validate or modify tool inputs before execution
• Auto-approve specific tool calls
• Block dangerous operations

Decision control:

• allow: Bypass permission system
• deny: Prevent tool execution
• ask: Show user confirmation dialog

PermissionRequest

Runs when user is shown a permission dialog.

Use cases:

• Auto-approve known-safe operations
• Deny specific permission requests
• Modify tool inputs before approval

PostToolUse

Runs immediately after successful tool completion.

Use cases:

• Trigger notifications or logging
• Validate tool outputs
• Add feedback for Claude to consider

UserPromptSubmit

Runs when user submits a prompt, before Claude processes it.

Use cases:

• Add contextual information to conversations
• Validate prompts for security issues
• Block inappropriate prompts

Stop / SubagentStop

Runs when Claude or a subagent finishes responding.

Use cases:

• Prevent premature stopping
• Ensure task completion
• Continue work automatically

SessionStart

Runs when Claude starts or resumes a session.

Matchers: startup, resume, clear, compact

Use cases:

• Load development context (recent issues, changes)
• Install dependencies or configure environment
• Persist environment variables via $CLAUDE_ENV_FILE

SessionEnd

Runs when a Claude Code session ends.

Use cases:

• Cleanup tasks or logging
• Save session state or statistics

Notification

Runs when Claude Code sends notifications.

Matchers: permission_prompt, idle_prompt, auth_success, elicitation_dialog

Use cases:

• Custom notification handling
• Forward notifications to external systems

Configuration Structure

Basic Structure

{
  "hooks": {
    "EventName": [
      {
        "matcher": "ToolPattern|*",
        "hooks": [
          {
            "type": "command|prompt",
            "command": "bash-command",
            "prompt": "llm-prompt",
            "timeout": 60
          }
        ]
      }
    ]
  }
}

Configuration Locations

Hooks are configured in settings files (priority order):

1. ~/.claude/settings.json - User settings
2. .claude/settings.json - Project settings
3. .claude/settings.local.json - Local project (not committed)
4. Plugin hooks - Auto-merged when plugins enabled

Matcher Patterns

• Exact match: Write matches only Write tool
• Multiple tools: Edit|Write matches Edit or Write
• Wildcards: .* or * matches all tools
• Empty/omitted: Matches all tools (for events without matchers)

Environment Variables

Available in all hook commands:

• $CLAUDE_PROJECT_DIR: Project root directory
• $CLAUDE_PLUGIN_ROOT: Plugin directory (plugin hooks only)
• $CLAUDE_CODE_REMOTE: "true" in remote web environment (unset locally)
• $CLAUDE_ENV_FILE: Persist environment vars (SessionStart only)

Hook Input/Output

Input Format

Hooks receive JSON via stdin:

{
  "session_id": "abc123",
  "transcript_path": "/path/to/session.jsonl",
  "cwd": "/current/working/directory",
  "permission_mode": "default",
  "hook_event_name": "PreToolUse",
  "tool_name": "Bash",
  "tool_input": {
    "command": "echo 'hello'"
  },
  "tool_use_id": "toolu_01ABC123..."
}

Output: Exit Codes

• 0: Success (stdout shown in verbose mode, except UserPromptSubmit/SessionStart where stdout is added as context)
• 2: Blocking error (stderr fed back to Claude, blocks action for applicable events)
• Other: Non-blocking error (stderr shown in verbose mode)

Output: JSON Control

Return structured JSON in stdout for advanced control:

{
  "continue": true,
  "stopReason": "Message shown when stopping",
  "suppressOutput": true,
  "systemMessage": "Warning to user"
}

PreToolUse JSON Output

{
  "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "allow|deny|ask",
    "permissionDecisionReason": "Explanation",
    "updatedInput": {
      "field_to_modify": "new value"
    }
  }
}

PostToolUse JSON Output

{
  "decision": "block",
  "reason": "Explanation",
  "hookSpecificOutput": {
    "hookEventName": "PostToolUse",
    "additionalContext": "Extra information for Claude"
  }
}

Stop/SubagentStop JSON Output

{
  "decision": "block",
  "reason": "Why Claude should continue (required when blocking)"
}

Common Patterns

Pattern 1: Command Validation

Validate bash commands before execution:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/validate-bash.sh",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

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

# Enforce better tools
if echo "$command" | grep -qE '\bgrep\b(?!.*\|)'; then
  echo "Use 'rg' (ripgrep) instead of 'grep'" >&2
  exit 2
fi

if echo "$command" | grep -qE '\bfind\s+\S+\s+-name\b'; then
  echo "Use 'rg --files' instead of 'find -name'" >&2
  exit 2
fi

exit 0

Pattern 2: Auto-Approve Safe Operations

Auto-approve documentation file reads:

{
  "hooks": {
    "PermissionRequest": [
      {
        "matcher": "Read",
        "hooks": [
          {
            "type": "command",
            "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/auto-approve-docs.sh"
          }
        ]
      }
    ]
  }
}

#!/usr/bin/env python3
# .claude/hooks/auto-approve-docs.sh
import json
import sys

input_data = json.load(sys.stdin)
file_path = input_data.get("tool_input", {}).get("file_path", "")

# Auto-approve documentation files
if file_path.endswith((".md", ".mdx", ".txt", ".json")):
    output = {
        "hookSpecificOutput": {
            "hookEventName": "PermissionRequest",
            "decision": {
                "behavior": "allow"
            }
        },
        "suppressOutput": True
    }
    print(json.dumps(output))
    sys.exit(0)

sys.exit(0)

Pattern 3: Add Session Context

Load recent changes on session start:

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "startup|resume",
        "hooks": [
          {
            "type": "command",
            "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/load-context.sh"
          }
        ]
      }
    ]
  }
}

#!/bin/bash
# .claude/hooks/load-context.sh
cd "$CLAUDE_PROJECT_DIR" || exit 0

# Get recent commits (last 5)
recent_changes=$(git log -5 --oneline --pretty=format:"- %s" 2>/dev/null)

# Get current branch
current_branch=$(git branch --show-current 2>/dev/null)

if [ -n "$recent_changes" ] || [ -n "$current_branch" ]; then
  echo "# Repository Context"
  echo ""
  echo "**Current branch:** $current_branch"
  echo ""
  echo "**Recent changes:**"
  echo "$recent_changes"
fi

exit 0

Pattern 4: Prompt-Based Stop Hook

Use LLM to intelligently decide if Claude should stop:

{
  "hooks": {
    "Stop": [
      {
        "hooks": [
          {
            "type": "prompt",
            "prompt": "Analyze the conversation context: $ARGUMENTS\n\nDetermine if:\n1. All requested tasks are complete\n2. All tests pass\n3. No errors remain unaddressed\n\nRespond with JSON: {\"decision\": \"approve\"|\"block\", \"reason\": \"explanation\"}",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Pattern 5: MCP Tool Integration

Target specific MCP tools:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "mcp__memory__.*",
        "hooks": [
          {
            "type": "command",
            "command": "echo 'Memory operation' >> ~/hooks.log"
          }
        ]
      },
      {
        "matcher": "mcp__.*__write.*",
        "hooks": [
          {
            "type": "command",
            "command": "/path/to/validate-mcp-write.sh"
          }
        ]
      }
    ]
  }
}

Security Considerations

Disclaimer

USE AT YOUR OWN RISK: Hooks execute arbitrary shell commands automatically. You are solely responsible for hook commands. Malicious or poorly written hooks can cause data loss or system damage.

Best Practices

1. Validate inputs: Never trust input data blindly
2. Quote variables: Use "$VAR" not $VAR
3. Block path traversal: Check for .. in file paths
4. Use absolute paths: Specify full paths via $CLAUDE_PROJECT_DIR
5. Skip sensitive files: Avoid .env, .git/, keys
6. Test thoroughly: Test hooks in safe environment first

Configuration Safety

Direct edits to hooks don't take effect immediately. Claude Code:

1. Captures hook snapshot at startup
2. Uses snapshot throughout session
3. Warns if hooks modified externally
4. Requires review in /hooks menu for changes

Debugging

Basic Troubleshooting

1. Check configuration: Run /hooks to verify hook registration
2. Verify syntax: Ensure JSON settings are valid
3. Test commands: Run hook commands manually first
4. Check permissions: Ensure scripts are executable (chmod +x)
5. Review logs: Use claude --debug for execution details

Common Issues

• Quotes not escaped: Use \" inside JSON strings
• Wrong matcher: Tool names are case-sensitive
• Command not found: Use absolute paths for scripts
• Exit code confusion: Exit code 2 blocks, others warn

Debug Output

Run claude --debug to see detailed hook execution:

[DEBUG] Executing hooks for PostToolUse:Write
[DEBUG] Found 1 hook matchers in settings
[DEBUG] Matched 1 hooks for query "Write"
[DEBUG] Executing hook command: <command> with timeout 60000ms
[DEBUG] Hook command completed with status 0

Plugin Hooks

Plugins can provide hooks that integrate with user/project hooks:

{
  "description": "Automatic code formatting",
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          {
            "type": "command",
            "command": "${CLAUDE_PLUGIN_ROOT}/scripts/format.sh",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Place in plugins/your-plugin/hooks/hooks.json or specify custom path in plugin metadata.

Execution Details

• Timeout: 60 seconds default (configurable per command)
• Parallelization: All matching hooks run in parallel
• Deduplication: Identical hook commands auto-deduplicated
• Environment: Current directory with Claude Code's environment
• Input: JSON via stdin
• Output: Varies by event (see above)

Guidelines

• Keep hooks simple: Prefer bash scripts for straightforward logic
• Use prompt hooks sparingly: Only for nuanced, context-aware decisions
• Set appropriate timeouts: Adjust based on expected execution time
• Handle errors explicitly: Scripts should validate inputs and handle errors
• Test thoroughly: Verify hooks work as expected before production use
• Document purpose: Add comments explaining hook behavior
• Use environment variables: Leverage $CLAUDE_PROJECT_DIR for portability

Examples

See examples.md for complete working examples of common hook patterns.

Reference

For complete API details, see:

• official-hooks-docs.md - Full Claude Code hooks documentation
• hook-input-schemas.md - Detailed input/output schemas
• security-guide.md - Security best practices