Hook Creator

Hook Creator

This skill guides the creation of Claude Code hooks - deterministic shell commands or LLM prompts that execute at specific points in Claude's lifecycle.

What Are Hooks?

Observability

First: At the start of execution, display:

🔧 Skill "julien-dev-hook-creator" activated

Hooks provide deterministic control over Claude's behavior. Unlike skills (which Claude chooses to use), hooks always execute at their designated lifecycle event.

┌─────────────────────────────────────────────────────────────────┐
│                    HOOKS vs SKILLS                              │
├─────────────────────────────────────────────────────────────────┤
│  HOOKS: Deterministic, always run at lifecycle events           │
│  SKILLS: Model-invoked, Claude decides when to use              │
└─────────────────────────────────────────────────────────────────┘

Available Hook Events

Hook Configuration

Hooks are configured in ~/​.claude/​settings.json (global) or .claude/​settings.json (project).

Basic Structure

{
  "hooks": {
    "EventName": [
      {
        "matcher": "ToolPattern",
        "hooks": [
          {
            "type": "command",
            "command": "your-command-here",
            "timeout": 60
          }
        ]
      }
    ]
  }
}

Configuration Fields

Matcher Patterns

"matcher": "Write"           // Exact match
"matcher": "Edit|Write"      // OR pattern (regex)
"matcher": "Notebook.*"      // Wildcard pattern
"matcher": "*"               // All tools (or omit matcher)

Hook Input (stdin)

Hooks receive JSON via stdin with context about the event:

{
  "session_id": "abc123",
  "transcript_path": "/​path/​to/​transcript.jsonl",
  "cwd": "/​current/​working/​directory",
  "hook_event_name": "PreToolUse",
  "tool_name": "Write",
  "tool_input": {
    "file_path": "/​path/​to/​file.txt",
    "content": "file content"
  }
}

Hook Output (Exit Codes)

Advanced JSON Output (exit 0)

{
  "continue": true,
  "stopReason": "message if continue=false",
  "suppressOutput": true,
  "systemMessage": "warning shown to user"
}

PreToolUse Decision Control

{
  "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "allow|deny|ask",
    "permissionDecisionReason": "Reason here",
    "updatedInput": {
      "field": "modified value"
    }
  }
}

Creating a Hook - Step by Step

Step 1: Identify the Use Case

Ask:

• When should this run? (which event)
• What should it do? (validate, log, transform, block)
• Scope: Global (~/​.claude/​settings.json) or project (.claude/​settings.json)?

Step 2: Write the Script

Create script in ~/​.claude/​scripts/ or .claude/​scripts/:

#!/​bin/​bash
# ~/​.claude/​scripts/​my-hook.sh

# Read input from stdin
INPUT=$(cat)

# Parse with jq
TOOL_NAME=$(echo "$INPUT" | jq -r '.tool_name // empty')
FILE_PATH=$(echo "$INPUT" | jq -r '.tool_input.file_path // empty')

# Your logic here
if [[ "$FILE_PATH" == *".env"* ]]; then
    echo "Blocked: Cannot modify .env files" >&2
    exit 2  # Block the action
fi

exit 0  # Allow the action

Important: Make executable with chmod +x

Step 3: Configure the Hook

Add to settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          {
            "type": "command",
            "command": "bash ~/​.claude/​scripts/​my-hook.sh",
            "timeout": 10
          }
        ]
      }
    ]
  }
}

Step 4: Test

# Test script directly
echo '{"tool_name":"Write","tool_input":{"file_path":"/​test/​.env"}}' | bash ~/​.claude/​scripts/​my-hook.sh
echo "Exit code: $?"

Real-World Example: Terminal Title Restoration

Problem: happy.cmd and claude.cmd contain title %COMSPEC% which overwrites terminal title to "C:\WINDOWS\system32\cmd.exe"

Solution: SessionStart hook that restores the title after launch

Script: ~/​.claude/​scripts/​restore-terminal-title-on-start.ps1

# Restore terminal title on Claude Code SessionStart
# This runs AFTER Claude has potentially overwritten the title
try {
    # Get current directory name
    $dirName = if ($PWD.Path -eq $HOME) {
        "~"
    } else {
        Split-Path $PWD -Leaf
    }

    # Restore title using multiple methods for maximum compatibility

    # Method 1: PowerShell native
    $Host.UI.RawUI.WindowTitle = $dirName

    # Method 2: ANSI escape sequence (more reliable with Windows Terminal)
    Write-Host "$([char]27)]0;$dirName$([char]7)" -NoNewline

    # Exit with success
    exit 0
} catch {
    # Silent fail - don't break Claude startup
    exit 0
}

Configuration: ~/​.claude/​settings.json (NOT repo .claude/​settings.json)

{
  "hooks": {
    "SessionStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "node \"%USERPROFILE%\\.claude\\scripts\\session-start-banner.js\"",
            "timeout": 5
          },
          {
            "type": "command",
            "command": "powershell.exe -NoProfile -File \"%USERPROFILE%\\.claude\\scripts\\restore-terminal-title-on-start.ps1\"",
            "timeout": 2
          }
        ]
      }
    ]
  }
}

Timeline:

1. happy.cmd executes → title becomes "C:\WINDOWS\system32\cmd.exe"
2. Happy/​Claude starts
3. SessionStart hook: session-start-banner.js displays banner
4. SessionStart hook: restore-terminal-title-on-start.ps1 fixes the title

Result: Title restored to directory name despite npm CLI wrapper interference

Lesson: Hooks can fix issues caused by external tools (npm wrappers, shell scripts)!

Hook Languages: JavaScript vs Python vs PowerShell

JavaScript Hooks (Fastest Startup)

Pros:

• Node.js already loaded by Claude Code
• No interpreter startup cost
• Faster execution (~50-200ms faster than Python)
• Great async support

Cons:

• Limited system integration compared to PowerShell
• JSON parsing requires external library or built-in JSON

Examples:

• session-start-banner.js - Fast banner display
• track-skill-invocation.js - Performance-critical tracking
• fast-skill-router.js - Routing must be instant

When to use: Performance-critical hooks (SessionStart, UserPromptSubmit)

Python Hooks (Rich Ecosystem)

Pros:

• Rich libraries (json, pathlib, subprocess)
• Better for complex data processing
• Easier multiline string handling
• Great for ML/​data tasks

Cons:

• Python interpreter startup cost (~100-300ms)
• May not be installed on all systems

Examples:

• session-end-delete-reserved.py - Complex file operations
• save-session-for-memory.py - Data processing
• cleanup-null-files.py - File system traversal

When to use: Complex logic, data processing, non-time-critical tasks

PowerShell Hooks (Windows Native)

Pros:

• Native Windows API access
• Can modify environment directly
• Better integration with Windows Terminal
• Access to .NET framework

Cons:

• Windows-only
• Slower than JavaScript (~50-150ms startup)
• CRLF line ending issues

Examples:

• restore-terminal-title-on-start.ps1 - Terminal manipulation
• cleanup-null-files.ps1 - Windows file operations
• set-terminal-title.ps1 - Environment modification

When to use: Windows-specific tasks, terminal manipulation, .NET integration

Choosing the Right Language

Need speed? → JavaScript
Need Python libraries? → Python
Need Windows integration? → PowerShell
Need to modify terminal? → PowerShell
Need to call .NET APIs? → PowerShell
Need async operations? → JavaScript

Common Hook Patterns

1. File Protection (PreToolUse)

#!/​bin/​bash
INPUT=$(cat)
FILE_PATH=$(echo "$INPUT" | jq -r '.tool_input.file_path // empty')

PROTECTED=(".env" "package-lock.json" ".git/" "credentials")
for pattern in "${PROTECTED[@]}"; do
    if [[ "$FILE_PATH" == *"$pattern"* ]]; then
        echo "Protected file: $pattern" >&2
        exit 2
    fi
done
exit 0

2. Auto-Format on Save (PostToolUse)

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          {
            "type": "command",
            "command": "FILE=$(cat | jq -r '.tool_input.file_path') && npx prettier --write \"$FILE\" 2>/​dev/​null || true"
          }
        ]
      }
    ]
  }
}

3. Command Logging (PostToolUse)

#!/​bin/​bash
INPUT=$(cat)
CMD=$(echo "$INPUT" | jq -r '.tool_input.command // empty')
DESC=$(echo "$INPUT" | jq -r '.tool_input.description // "No description"')
echo "$(date +%Y-%m-%d_%H:%M:%S) | $CMD | $DESC" >> ~/​.claude/​logs/​bash-commands.log
exit 0

4. Session Sync (SessionStart/​SessionEnd)

{
  "hooks": {
    "SessionStart": [
      {
        "hooks": [{
          "type": "command",
          "command": "bash ~/​.claude/​scripts/​sync-marketplace.sh",
          "timeout": 30
        }]
      }
    ],
    "SessionEnd": [
      {
        "hooks": [{
          "type": "command",
          "command": "bash ~/​.claude/​scripts/​push-marketplace.sh",
          "timeout": 30
        }]
      }
    ]
  }
}

5. Add Context to Prompts (UserPromptSubmit)

#!/​bin/​bash
# stdout is added as context to the prompt
echo "Current git branch: $(git branch --show-current 2>/​dev/​null || echo 'not a git repo')"
echo "Node version: $(node -v 2>/​dev/​null || echo 'not installed')"
exit 0

6. LLM-based Stop Decision (Stop)

{
  "hooks": {
    "Stop": [
      {
        "hooks": [{
          "type": "prompt",
          "prompt": "Review if all tasks are complete. Check: 1) All todos marked done 2) Tests passing 3) No pending questions. Respond with decision: approve (stop) or block (continue).",
          "timeout": 30
        }]
      }
    ]
  }
}

Best Practices

Do's

• ✅ Always quote shell variables: "$VAR" not $VAR
• ✅ Use absolute paths for scripts
• ✅ Handle errors gracefully (exit 0 if non-critical)
• ✅ Set appropriate timeouts
• ✅ Test scripts independently before configuring
• ✅ Use tr -d '\r' for Windows CRLF compatibility

Don'ts

• ❌ Don't block critical operations without good reason
• ❌ Don't use long timeouts (blocks Claude)
• ❌ Don't trust input blindly - validate paths
• ❌ Don't expose secrets in logs
• ❌ Don't use interactive commands (no stdin available)

Debugging Hooks

# Run with debug output
bash -x ~/​.claude/​scripts/​my-hook.sh

# Test with sample input
echo '{"tool_name":"Write","tool_input":{"file_path":"/​test/​file.txt"}}' | bash ~/​.claude/​scripts/​my-hook.sh

# Check hook errors in Claude Code
# Look for "hook error" messages in the UI

For detailed troubleshooting of common errors (timeout, CRLF, jq not found, etc.), see references/​troubleshooting.md.

Environment Variables

Available in hooks:

• CLAUDE_PROJECT_DIR - Current project directory
• CLAUDE_CODE_REMOTE - Remote mode indicator
• CLAUDE_ENV_FILE - (SessionStart only) File path for persisting env vars

File Locations - CRITICAL INFORMATION

⚠️ CRITICAL WARNING

Claude Code uses ~/​.claude/​settings.json (home directory) NOT the repo .claude/​settings.json

These files are DIFFERENT and must be synchronized manually!

Best Practice:

1. Modify ~/​.claude/​settings.json first (real file)
2. Copy changes to .claude/​settings.json (for versioning)
3. Commit repo version for documentation

Never assume the repo version is active!

Verification:

# Check what Claude Code actually uses
cat ~/​.claude/​settings.json | grep -A 5 "SessionStart"

# Compare with repo version
diff ~/​.claude/​settings.json .claude/​settings.json

Quick Reference

Event Flow:
SessionStart → UserPromptSubmit → PreToolUse → [Tool] → PostToolUse → Stop → SessionEnd

Exit Codes:
0 = Success (continue)
2 = Block (stop action, feed stderr to Claude)
* = Non-blocking error

Matcher:
"Write"        = exact match
"Edit|Write"   = OR
"Notebook.*"   = regex
"*" or omit    = all tools

🔗 Skill Chaining

Skills Required Before

• Aucun (skill autonome)
• Optionnel: Connaissance de base de bash/​shell scripting

Input Expected

• Use case description: Quel événement déclencher, quelle action effectuer
• Scope decision: Global (~/​.claude/​settings.json) ou project (.claude/​settings.json)
• Prerequisites: jq installé pour parsing JSON

Output Produced

• Format:
  • Script bash dans ~/​.claude/​scripts/ ou .claude/​scripts/
  • Configuration JSON dans settings.json
• Side effects:
  • Création/​modification de fichiers scripts
  • Modification de settings.json
  • Hooks actifs au prochain événement
• Duration: 2-5 minutes pour un hook simple

Compatible Skills After

Recommandés:

• sync-personal-skills: Si le hook modifie des fichiers du marketplace
• skill-creator: Si création d'un skill qui intègre des hooks

Optionnels:

• Git workflow: Committer les scripts et settings

Called By

• Direct user invocation: "Crée un hook pour...", "Je veux automatiser..."
• Part of skill/​workflow development

Tools Used

• Read (lecture settings.json existant)
• Write (création scripts bash)
• Edit (modification settings.json)
• Bash (test du hook, chmod +x)

Visual Workflow

User: "Je veux protéger les fichiers .env"
    ↓
hook-creator (this skill)
    ├─► Step 1: Identify event (PreToolUse)
    ├─► Step 2: Write script (protect-files.sh)
    ├─► Step 3: chmod +x script
    ├─► Step 4: Configure settings.json
    └─► Step 5: Test with sample input
    ↓
Hook active ✅
    ↓
[Next: Test in real session]

Usage Example

Scenario: Créer un hook de logging des commandes bash

Input: "Log toutes les commandes bash exécutées"

Process:

1. Event identifié: PostToolUse avec matcher Bash
2. Script créé: ~/​.claude/​scripts/​log-bash.sh
3. Settings.json mis à jour avec hook config
4. Test avec sample JSON input

Result:

• Script logging actif
• Commandes loguées dans ~/​.claude/​logs/​bash-commands.log