Example Usage
Input (Skill to Review)
``json id="9j3kdx"
{
"skill_name": "example-email-sender",
"source": "github",
"description": "Sends automated emails using an external API",
"files": ["SKILL.md", "scripts/send-email.sh"]
}
CODEBLOCK0 json id="4n6rfa"
{
"skill_name": "example-email-sender",
"purpose": "Send automated emails via external API",
"source": "github",
"capabilities": [
"network access",
"external API calls",
"file read/write"
],
"install_risk": "low",
"runtime_risk": "medium",
"trust_dependency": "opaque",
"warnings": [
"Uses external API with unclear data handling",
"No transparency on where email content is sent"
],
"recommendations": [
"Verify API endpoint and data handling policy",
"Limit data exposure before use"
],
"verdict": "caution",
"verified": false,
"verification": {
"status": "not_run",
"receipt_id": null,
"notes": ""
}
}
CODEBLOCK1 bash id="t2j9mf"
clawdhub install skill-vetter-v2
CODEBLOCK2 bash id="a1vk0r"
git clone https://github.com/your-org/skill-vetter-v2.git ~/.openclaw/skills/skill-vetter-v2
CODEBLOCK3 bash id="0xptv9"
cp -r hooks/openclaw ~/.openclaw/hooks/skill-vetter-v2
openclaw hooks enable skill-vetter-v2
CODEBLOCK4 bash id="z7p2qs"
bash scripts/scan-skill.sh /path/to/skill
`
This helper inventories files and flags common red-patterns locally. It does not make network calls.
## Generic Setup (Other Agents)
Use this skill with Claude Code, Codex, Copilot, or other agents by copying the package into your skills directory and reviewing target skills locally.
Suggested workflow:
1. Read the target SKILL.md`
- 2. Read all scripts, hooks, and references
- Run the local scan helper
- Write the structured report
- Optionally verify the report
What This Is Not
- * not an installer
- not an auto-executor for unknown code
- not an external decision authority
- not a replacement for human judgment on high-risk skills
Outcome
Agents can:
- * understand what a skill actually does before use
- identify install-time and runtime risks clearly
- separate transparent dependencies from opaque trust requirements
- keep safety decisions local while optionally producing verifiable records
Keywords
ai-agents, skill-safety, risk-analysis, verification, trust, security
使用示例
输入(待审查技能)
json id=9j3kdx
{
skill_name: example-email-sender,
source: github,
description: 使用外部API发送自动邮件,
files: [SKILL.md, scripts/send-email.sh]
}
输出(审查报告)
json id=4n6rfa
{
skill_name: example-email-sender,
purpose: 通过外部API发送自动邮件,
source: github,
capabilities: [
网络访问,
外部API调用,
文件读写
],
install_risk: 低,
runtime_risk: 中,
trust_dependency: 不透明,
warnings: [
使用外部API,数据处理方式不明确,
邮件内容发送去向不透明
],
recommendations: [
验证API端点及数据处理策略,
使用前限制数据暴露范围
],
verdict: 谨慎使用,
verified: false,
verification: {
status: not_run,
receipt_id: null,
notes:
}
}
可选验证工作流
仅在本地审查完成后使用验证。
推荐模式:
- 1. 为报告定义确定性验证规范
* 必需字段存在
* 风险标签内部一致
* 结论有发现项支撑
* 不包含禁止数据
- 2. 仅提交结构化报告和规范
- 3. 保守解读结果
* 通过 → 附加收据元数据并标记 verified: true
* 失败 → 修正报告并保持 verified: false
* 不确定 → 保持 verified: false 并升级为人工审查
验证为可选步骤,绝不能覆盖本地安全考量。
OpenClaw 设置(推荐)
OpenClaw 最适合此技能,因为它支持打包技能、钩子和工作区上下文。
安装
通过 ClawHub:
bash id=t2j9mf
clawdhub install skill-vetter-v2
手动安装:
bash id=a1vk0r
git clone https://github.com/your-org/skill-vetter-v2.git ~/.openclaw/skills/skill-vetter-v2
可选钩子
如需在信任技能前获得审查提示,可安装提醒钩子:
bash id=0xptv9
cp -r hooks/openclaw ~/.openclaw/hooks/skill-vetter-v2
openclaw hooks enable skill-vetter-v2
本地扫描助手
对技能文件夹运行本地助手:
bash id=z7p2qs
bash scripts/scan-skill.sh /path/to/skill
该助手会在本地盘点文件并标记常见风险模式。它不会发起网络调用。
通用设置(其他代理)
将此技能与 Claude Code、Codex、Copilot 或其他代理配合使用时,将包复制到技能目录并在本地审查目标技能。
建议工作流:
- 1. 读取目标 SKILL.md
- 读取所有脚本、钩子和引用
- 运行本地扫描助手
- 编写结构化报告
- 可选:验证报告
本技能不包含的功能
- * 不是安装器
- 不是未知代码的自动执行器
- 不是外部决策权威
- 不能替代对高风险技能的人工判断
成果
代理能够:
- * 在使用前了解技能的实际功能
- 清晰识别安装时和运行时的风险
- 将透明依赖与不透明信任要求区分开
- 在本地做出安全决策,同时可选生成可验证记录
关键词
AI代理、技能安全、风险分析、验证、信任、安全