Hook Creator

If the user's intent does not match the purpose of this skill, load plugin-lifecycle to route to the right skill and process: Skill(skill="plugin-creator:plugin-lifecycle").

Hook Creator for Claude Code Plugins

Create hooks that integrate with the Claude Code event system. Hooks automate validation, enforcement, and context injection across the session lifecycle.

For the complete hooks I/O API and JSON schemas, load: Skill(skill: "plugin-creator:hooks-io-api") For working code examples and plugin hook configuration patterns, load: Skill(skill: "plugin-creator:hooks-patterns") For all hook events, matchers, and environment variables, load: Skill(skill: "plugin-creator:hooks-core-reference")

Language Selection

Claude Code hooks are language-agnostic — any executable with a shebang works. Choose the language based on context:

flowchart TD
    Start([Need a hook script?]) --> Q1{Does the project have an obvious runtime?}
    Q1 -->|Yes — Python project pyproject.toml present| Python["Write Python hook<br>#!/usr/bin/env python3<br>See hooks-python.md"]
    Q1 -->|Yes — other runtime e.g. Ruby Gemfile| Other["Write hook in that language<br>Any executable with shebang works"]
    Q1 -->|No — general purpose or cross-platform| NodeJS["Default: Node.js<br>Claude Code's presence implies Node.js available"]
    NodeJS --> Q2{Module syntax?}
    Q2 -->|import — ESM| MJS["Use .mjs extension<br>Preferred default for new Node.js scripts"]
    Q2 -->|require — CommonJS| CJS["Use .cjs extension<br>Use when require() is needed"]
    MJS --> Pattern[Follow Node.js hook pattern below]
    CJS --> Pattern
    Python --> Done[Follow language-specific pattern]
    Other --> Done

Node.js is the default because Claude Code's presence on a host implies Node.js is available. When the project has an obvious runtime dependency (e.g., a Python project with pyproject.toml), write the hook in that language instead — there is no need to introduce a Node.js dependency just for hooks.

For Node.js scripts specifically: use .mjs (ESM, preferred) or .cjs (CommonJS). Never use plain .js. See hooks-nodejs-extension.md for the full rule and rationale.

For Python hooks: see Skill(skill: "plugin-creator:hooks-guide") → references/hooks-python.md.

New Node.js hooks written from scratch: default to .mjs. Existing hooks that already use require(): keep .cjs.

Node.js Hook Template

For new scripts use .mjs (ESM) with import syntax. For hooks that already use require(), keep .cjs. The .cjs template below works in any project — change require to import and the extension to .mjs when writing a new script from scratch.

.cjs template (CommonJS — use when require() is needed):

#!/usr/bin/env node
'use strict';
/**
 * Brief description of what this hook does.
 * Fires on: EventName — ToolMatcher
 * Action: blocking | non-blocking — what it does
 */

const { execFileSync } = require('node:child_process');
const fs = require('node:fs');

let input = '';
process.stdin.setEncoding('utf8');
process.stdin.on('data', (chunk) => { input += chunk; });
process.stdin.on('end', () => {
  let data = {};
  try {
    data = JSON.parse(input);
  } catch {
    process.stdout.write(JSON.stringify({}));
    process.exit(0);
  }

  // Extract fields
  const toolName = data.tool_name || '';
  const toolInput = data.tool_input || {};

  // Guard: exit early if this hook does not apply
  if (toolName !== 'Write') {
    process.stdout.write(JSON.stringify({}));
    process.exit(0);
  }

  // Hook logic here
  const output = {
    hookSpecificOutput: {
      hookEventName: 'PreToolUse',
      additionalContext: 'Context injected into Claude.',
    },
  };

  process.stdout.write(JSON.stringify(output));
  process.exit(0);
});

Critical rules:

• Use process.stdin.on('data') / process.stdin.on('end') — not synchronous stdin reads
• Exit with process.exit(0) for success, process.exit(2) for blocking error
• Write JSON to stdout via process.stdout.write(JSON.stringify(output))
• Suppress stderr — use stdio: ['ignore', 'pipe', 'ignore'] when calling child processes via execFileSync
• Output empty {} on early exit paths, not nothing

hooks.json Structure

Plugin hooks live in hooks/hooks.json inside the plugin directory.

Empty state (hooks file exists but no active hooks):

{"hooks": {}}

The file must exist. Remove entries to disable hooks. Remove the file only when decommissioning the plugin.

Active hooks format:

{
  "description": "Brief explanation of what these hooks enforce",
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          {
            "type": "command",
            "command": "${CLAUDE_PLUGIN_ROOT}/hooks/my-hook.mjs",
            "timeout": 3
          }
        ]
      }
    ],
    "Stop": [
      {
        "hooks": [
          {
            "type": "prompt",
            "prompt": "Verify the task is complete. Return {\"ok\": true} or {\"ok\": false, \"reason\": \"...\"}.",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Path rule: Always use ${CLAUDE_PLUGIN_ROOT}/hooks/filename.mjs (or .cjs / .py as appropriate) — never hardcoded paths.

SOURCE: plugins/orchestrator-discipline/hooks.json (lines 1-25), verified 2026-02-19.

Event Selection

flowchart TD
    Start([What should the hook do?]) --> Q1{When does it act?}
    Q1 -->|Before a tool runs — can block| PreToolUse[PreToolUse + tool name matcher]
    Q1 -->|After a tool succeeds| PostToolUse[PostToolUse + tool name matcher]
    Q1 -->|After a tool fails| PostToolUseFailure[PostToolUseFailure + tool name matcher]
    Q1 -->|Before Claude stops responding| Stop[Stop — no matcher]
    Q1 -->|When subagent spawned| SubagentStart[SubagentStart — matcher: agent type name]
    Q1 -->|When subagent finishes| SubagentStop[SubagentStop — matcher: agent type name]
    Q1 -->|Agent team teammate going idle| TeammateIdle[TeammateIdle — no matcher]
    Q1 -->|Task being marked complete| TaskCompleted[TaskCompleted — no matcher]
    Q1 -->|CLAUDE.md or rules file loaded| InstructionsLoaded[InstructionsLoaded — no matcher]
    Q1 -->|Config file changes| ConfigChange[ConfigChange — matcher: config source]
    Q1 -->|Worktree being created| WorktreeCreate[WorktreeCreate — no matcher]
    Q1 -->|Worktree being removed| WorktreeRemove[WorktreeRemove — no matcher]
    Q1 -->|MCP server requests user input| Elicitation[Elicitation — matcher: MCP server name]
    Q1 -->|User responds to MCP elicitation| ElicitationResult[ElicitationResult — matcher: MCP server name]
    Q1 -->|When user submits a prompt| UserPromptSubmit[UserPromptSubmit — no matcher]
    Q1 -->|When session starts| SessionStart[SessionStart — matcher: startup|resume|clear|compact]
    Q1 -->|When session ends| SessionEnd[SessionEnd — matcher: exit reason]
    Q1 -->|Before context compaction| PreCompact[PreCompact — matcher: manual|auto]
    Q1 -->|After context compaction| PostCompact[PostCompact — matcher: manual|auto]
    Q1 -->|On repo setup or maintenance| Setup[Setup — matcher: init|maintenance]
    Q1 -->|When Claude sends notifications| Notification[Notification — notification matcher]
    PreToolUse --> HookType[Choose hook type]
    PostToolUse --> HookType
    PostToolUseFailure --> HookType
    Stop --> HookType
    SubagentStart --> HookType
    SubagentStop --> HookType
    TeammateIdle --> HookType
    TaskCompleted --> HookType
    InstructionsLoaded --> HookType
    ConfigChange --> HookType
    WorktreeCreate --> HookType
    WorktreeRemove --> HookType
    Elicitation --> HookType
    ElicitationResult --> HookType
    UserPromptSubmit --> HookType
    SessionStart --> HookType
    SessionEnd --> HookType
    PreCompact --> HookType
    PostCompact --> HookType
    Setup --> HookType
    Notification --> HookType
    HookType --> Q2{Decision logic complexity?}
    Q2 -->|Deterministic rule — fast| Command[type: command — any executable script]
    Q2 -->|Context-aware — needs reasoning| Prompt[type: prompt — LLM-evaluated]
    Q2 -->|Complex verification with tools| Agent[type: agent — agentic verifier]
    Q2 -->|Send to external service| HTTP[type: http — POST to URL]

Events with matchers (non-tool): SubagentStart/SubagentStop match agent type name; SessionEnd matches exit reason; ConfigChange matches config source; Elicitation/ElicitationResult match MCP server name.

Events without matchers: Stop, UserPromptSubmit, TeammateIdle, TaskCompleted, InstructionsLoaded, WorktreeCreate, WorktreeRemove — omit the matcher field entirely.

Hook Type Selection

flowchart TD
    Start([Choose hook type]) --> Q{What kind of logic?}
    Q -->|Regex/pattern match, path check, fast rule| Command["type: command\nAny executable script\nFast, deterministic, no API call"]
    Q -->|Understand intent, evaluate context, reasoning| Prompt["type: prompt\nLLM evaluated (fast model)\nSlower, flexible, context-aware"]
    Q -->|Complex verification needing tool access| Agent["type: agent\nAgentic verifier with tools\nMost flexible, highest latency"]
    Q -->|Send to external service or webhook| HTTP["type: http\nPOST to URL\nFor remote validation services"]
    Command --> Timeout1[Timeout: 3s for local binary checks\nMax 10s for slow operations\nDefault: 600s built-in]
    Prompt --> Timeout2[Timeout: 30s default\nAdjust if reasoning takes longer]
    Agent --> Timeout3[Timeout: 60s default\nAdjust for complex tasks]
    HTTP --> Timeout4[Timeout: configurable\nNon-2xx errors are non-blocking]

Prompt hook response schema (LLM must return):

{"ok": true}

or:

{"ok": false, "reason": "Explanation shown to Claude"}

Timeout Sizing

Default: 600s for command hooks, 30s for prompt hooks, 60s for agent hooks (Claude Code built-in). Set explicit timeouts to reflect actual operation time — oversized timeouts mask runaway scripts.

Note: The Agent tool was renamed from Task in Claude Code v2.1.63. Use Agent as the matcher name for PreToolUse/PostToolUse hooks targeting subagent-spawning tool calls.

SOURCE: hooks-core-reference — timeout defaults verified from official docs.

Matchers

Tool name matchers (case-sensitive):

"matcher": "Write"
"matcher": "Write|Edit"
"matcher": "Bash"
"matcher": "mcp__.*__write.*"
"matcher": "*"

SessionStart matchers: startup, resume, clear, compact

PreCompact matchers: manual, auto

Notification matchers: permission_prompt, idle_prompt, auth_success, elicitation_dialog

Events without matchers — omit the field: Stop, UserPromptSubmit, TeammateIdle, TaskCompleted, InstructionsLoaded, WorktreeCreate, WorktreeRemove

Stdio Suppression

When calling external processes from a hook, suppress stderr to prevent noise in the transcript:

const { execFileSync } = require('node:child_process');

// Correct — stderr suppressed
try {
  const result = execFileSync('/usr/bin/node', ['--version'], {
    stdio: ['ignore', 'pipe', 'ignore'],
    timeout: 3000,
  });
} catch (err) {
  // Handle failure without leaking stderr
}

Use execFileSync (not execSync) for binary calls — prevents shell injection.

Exit Codes

Exit code 2 blocks the triggering action. Exit code 0 with {"ok": false, "reason": "..."} blocks via prompt hook protocol.

JSON Output by Event

All events — common fields:

{
  "continue": true,
  "suppressOutput": false,
  "systemMessage": "Optional message shown to user"
}

PreToolUse — allow or deny:

{
  "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "allow",
    "permissionDecisionReason": "Auto-approved read-only file",
    "additionalContext": "Context for Claude"
  }
}

Stop / SubagentStop — block or allow:

{"decision": "block", "reason": "Task not complete — tests not run"}

UserPromptSubmit — add context or block:

{
  "hookSpecificOutput": {
    "hookEventName": "UserPromptSubmit",
    "additionalContext": "Additional context prepended to prompt"
  }
}

SessionStart — inject context:

{
  "hookSpecificOutput": {
    "hookEventName": "SessionStart",
    "additionalContext": "Project state loaded"
  }
}

Implementation Workflow

flowchart TD
    A([Start: add a hook to a plugin]) --> B[Identify event and behavior]
    B --> C{Hook type?}
    C -->|Command — deterministic| D["Create hooks/myhook.mjs or .cjs<br>or hooks/myhook.py or hooks/myhook.sh"]
    C -->|Prompt — LLM-evaluated| E[Write prompt string directly in hooks.json]
    D --> F["Test script directly: echo input-json | node hooks/myhook.mjs"]
    F --> G{Output valid JSON?}
    G -->|No| D
    G -->|Yes| H[Add entry to hooks/hooks.json using correct event + matcher]
    E --> H
    H --> I[Verify hooks.json is valid JSON: python3 -m json.tool hooks/hooks.json]
    I --> J[Restart Claude Code session — hooks snapshot at startup]
    J --> K[Test in session with claude --debug]
    K --> L{Hook fires correctly?}
    L -->|No| M[Check matcher case-sensitivity, event name, JSON syntax]
    M --> H
    L -->|Yes| N[Done]

Testing directly before wiring:

# Test with minimal input (adjust extension to match your script)
echo '{"tool_name":"Write","tool_input":{"file_path":"/tmp/test.txt"}}' | node ./hooks/my-hook.mjs

# Validate JSON output
echo '{"tool_name":"Write","tool_input":{"file_path":"/tmp/test.txt"}}' | node ./hooks/my-hook.mjs | python3 -m json.tool

Plugin hooks.json Registration

After writing a hook script, register it in hooks/hooks.json:

{
  "description": "What these hooks enforce",
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write",
        "hooks": [
          {
            "type": "command",
            "command": "${CLAUDE_PLUGIN_ROOT}/hooks/validate-write.mjs",
            "timeout": 3
          }
        ]
      }
    ]
  }
}

All parallel hooks within one event run simultaneously — each is independent.

Security Rules

• Quote all variable expansions: toolInput.file_path || ''
• Check for path traversal before acting on file paths: reject paths containing ..
• Use execFileSync with explicit argv arrays — never execSync with shell string interpolation
• Skip sensitive file paths: .env, .git/, *.pem, *credentials*

Debugging

# Enable hook debug output
claude --debug
claude --debug "hooks"

# Check loaded hooks in session
/hooks

Common issues:

Sources

• plugins/orchestrator-discipline/hooks.json — verified plugin hooks.json format (lines 1-25, 2026-02-19)
• .claude/hooks/session-start-backlog.cjs — verified Node.js hook pattern (lines 1-69, 2026-02-19)
• plugins/orchestrator-discipline/hooks/pre-tool-orchestrator-read-warning.cjs — verified stdin handling pattern (lines 28-85, 2026-02-19)
• Skill(skill: "plugin-creator:hooks-core-reference") — event reference, matchers, environment variables (accessed 2026-01-28)
• Skill(skill: "plugin-creator:hooks-io-api") — JSON input/output schemas (accessed 2026-01-28)
• Skill(skill: "plugin-creator:hooks-patterns") — prompt-based hooks, code examples (accessed 2026-01-28)
• Official hooks docs: https://code.claude.com/docs/en/hooks.md (accessed 2026-01-28)