Will the user explicitly say "do X"?
├─ YES → Command (user-triggered workflow)
│   └─ Does it have fixed steps? → YES → Write it as a Command
└─ NO → Does it need to run automatically on events?
    ├─ YES → Hook (event-driven)
    └─ NO → Skill (knowledge injection, loaded on demand)

## Container Setup Checklist ⚠️ MANDATORY

Before creating ANY container, verify ALL items:

- [ ] Icon: HD icon from official source (min 256x256)
- [ ] Template: XML template with all port/volume mappings
- [ ] Labels: net.unraid.docker.icon, net.unraid.docker.webui
- [ ] Data: Persistent volume mapped to standard path
- [ ] Network: Correct bridge network
- [ ] Permissions: PUID/PGID set correctly

---
description: "Publish a blog post to my portfolio site"
argument-hint: "[topic or title]"
---

# /blog — Publish Blog Post

## PHASE 1: PREPARE
1. Load the blog-publisher skill for format reference
2. Generate URL-friendly slug from topic
...

## PHASE 2: WRITE CONTENT
...

## RULES
- MUST have both English and Chinese sections
- MUST run build before deploying

#!/usr/bin/env bash
set -euo pipefail
INPUT=$(cat)  # JSON from stdin

# Parse and decide
RESULT=$(echo "$INPUT" | python3 -c "
import sys, json
data = json.load(sys.stdin)
tool = data.get('tool_name', '')
command = data.get('tool_input', {}).get('command', '')
if 'docker' in command and 'your-server' in command:
    print('REMIND')
else:
    print('SKIP')
" 2>/dev/null || echo "SKIP")

if [ "$RESULT" = "REMIND" ]; then
    cat <<'EOF'
{
  "decision": "allow",
  "reason": "⚠️ DOCKER CHECKLIST:\n1. HD icon from official source\n2. XML template with all mappings\n3. Required labels set\n4. Persistent volumes configured\n5. Correct network bridge\n6. PUID/PGID permissions"
}
EOF
else
    echo '{"decision": "allow"}'
fi

{
  "hooks": {
    "Stop": [{
      "matcher": "",
      "hooks": [
        {"type": "command", "command": "~/.agents/hooks/verify-before-complete.sh"},
        {"type": "command", "command": "~/.agents/hooks/auto-changelog.sh"},
        {"type": "command", "command": "~/.agents/hooks/artifact-completion.sh"}
      ]
    }],
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {"type": "command", "command": "~/.agents/hooks/debug-reminder.sh"},
          {"type": "command", "command": "~/.agents/hooks/docker-checklist.sh"}
        ]
      },
      {
        "matcher": "SomeApiTool",
        "hooks": [{"type": "command", "command": "~/.agents/hooks/upload-checklist.sh"}]
      }
    ],
    "PostToolUse": [{
      "matcher": "Write",
      "hooks": [{"type": "command", "command": "~/.agents/hooks/image-optimization.sh"}]
    }]
  }
}

# Positive: should trigger
echo '{"tool_name":"Bash","tool_input":{"command":"docker run..."}}' \
  | bash ~/.agents/hooks/docker-checklist.sh

# Negative: should be silent
echo '{"tool_name":"Bash","tool_input":{"command":"ls -la"}}' \
  | bash ~/.agents/hooks/docker-checklist.sh

I want to organize my AI agent skills into a three-layer defense system (Skills → Commands → Hooks).
Reference: https://mjshao.fun/blog/opencode-skill-command-hook

My setup:
- AI agent tool: [ask me — OpenCode, Claude Code, Cursor, etc.]
- Skills directory: [ask me — e.g. ~/.agents/skills/]
- Number of existing skills: [ask me]
- Top 3 recurring agent mistakes: [ask me]

Steps:
1. Scan my skills directory and list all skills with their structure
2. Identify skills with fixed step-by-step workflows → convert to /commands
3. Analyze my top recurring mistakes → create hook scripts (Stop, PreToolUse, PostToolUse)
4. Extract universal behavior rules → write to CLAUDE.md or AGENTS.md as global rules
5. Test each hook with positive (should trigger) and negative (should stay silent) cases
6. Output a summary: how many skills→commands, hooks created, rules extracted

用户会主动说"帮我做X"吗？
├─ YES → Command（用户触发的工作流）
│   └─ 有固定步骤？→ YES → 写成 Command
└─ NO → 需要在特定事件时自动运行？
    ├─ YES → Hook（事件驱动）
    └─ NO → Skill（知识注入，按需加载）

## 容器部署清单 ⚠️ 必须执行

创建任何容器前，确认以下所有项目：

- [ ] 图标：官方源 HD 图标（最小 256x256）
- [ ] 模板：包含所有端口/卷映射的 XML 模板
- [ ] 标签：必要的 docker labels 全部设置
- [ ] 数据：持久化卷映射到标准路径
- [ ] 网络：正确的 bridge 网络
- [ ] 权限：PUID/PGID 正确设置

#!/usr/bin/env bash
set -euo pipefail
INPUT=$(cat)  # stdin 接收 JSON

RESULT=$(echo "$INPUT" | python3 -c "
import sys, json
data = json.load(sys.stdin)
command = data.get('tool_input', {}).get('command', '')
if 'docker' in command and 'your-server' in command:
    print('REMIND')
else:
    print('SKIP')
" 2>/dev/null || echo "SKIP")

if [ "$RESULT" = "REMIND" ]; then
    cat <<'EOF'
{
  "decision": "allow",
  "reason": "⚠️ 容器部署清单：\n1. HD 图标\n2. XML 模板\n3. 必要标签\n4. 持久化卷\n5. 网络配置\n6. 权限设置"
}
EOF
else
    echo '{"decision": "allow"}'
fi

# 正向：应该触发
echo '{"tool_name":"Bash","tool_input":{"command":"docker run..."}}' \
  | bash ~/.agents/hooks/docker-checklist.sh

# 反向：应该静默
echo '{"tool_name":"Bash","tool_input":{"command":"ls -la"}}' \
  | bash ~/.agents/hooks/docker-checklist.sh

帮我把现有的 AI agent skill 整理成三层防御体系（Skill → Command → Hook）。
参考：https://mjshao.fun/blog/opencode-skill-command-hook

我的环境：
- AI agent 工具：[问我——OpenCode、Claude Code、Cursor 等]
- Skill 目录：[问我——比如 ~/.agents/skills/]
- 现有 skill 数量：[问我]
- 最常犯的 3 个错误：[问我]

执行步骤：
1. 扫描 skill 目录，列出所有 skill 及其结构
2. 识别有固定步骤流程的 skill → 转成 /command
3. 针对我最常犯的错误 → 创建 hook 脚本（Stop、PreToolUse、PostToolUse）
4. 提取通用行为规则 → 写入 CLAUDE.md 或 AGENTS.md 作为 global rule
5. 每个 hook 做正向（应触发）和反向（应静默）测试
6. 输出清单：多少 skill→command、创建了哪些 hook、提取了哪些规则