← 返回 Skills 市场
127
总下载
0
收藏
0
当前安装
1
版本数
在 OpenClaw 中安装
/install evr-xiaobai
功能描述
Execute-Verify-Report 三步法工具 — 强制AI遵循"执行→验证→报告"流程,杜绝"说做了其实没做"、"完成了但不验证"等问题。基于STANDING-ORDERS核心原则设计。
使用说明 (SKILL.md)
Execute-Verify-Report (EVR) - 三步法工具 ✅
执行不是确认,验证不是可选,报告不是敷衍
问题背景
AI Agent常见错误:
- ❌ "我会做的" ≠ 做了
- ❌ "完成了"但没有验证
- ❌ 无限重试同一失败操作
- ❌ 静默失败不报告
Execute-Verify-Report 三步法从根源解决。
核心原则
| 步骤 | 含义 | 反例 | 正例 |
|---|---|---|---|
| Execute | 实际执行 | "我会去做的" | "已执行:date 输出 13:17:23" |
| Verify | 验证结果 | "应该成功了" | "验证:文件存在,大小2.3KB" |
| Report | 完整汇报 | "搞定了" | "执行→验证→结果→下一步" |
使用方法
方式1:汇报模板(强制)
**任务**: [任务描述]
🔧 **Execute**
```bash
[实际执行的命令]
输出: [粘贴实际输出]
✅ Verify
- 验证方法: [如何验证]
- 验证结果: ✅通过 / ❌失败
- 失败原因: [如果失败]
📋 Report
- 结果: [一句话总结]
- 数据: [具体数字/状态]
- 下一步: [下一步行动]
### 方式2:脚本辅助
```bash
# 开始任务
evr start "修复安全漏洞"
# 执行步骤
evr step "修改配置文件权限"
# 自动提示: 请运行验证命令
# 验证步骤
evr verify "chmod 600 ~/.openclaw/openclaw.json"
# 自动检查: 文件权限是否为600?
# 结束任务
evr done
# 生成完整EVR报告
失败处理
重试规则
// 最多重试3次
for (let attempt = 1; attempt \x3C= 3; attempt++) {
const result = execute(command);
if (verify(result)) {
report("✅ 成功", result);
break;
}
if (attempt === 3) {
report("❌ 连续3次失败,停止重试", result);
escalate("需要人工干预");
}
}
静默失败检测
# 错误示例(静默失败)
rm important-file.txt # 如果文件不存在,命令返回0但无事发生
# EVR正确做法
if [ -f important-file.txt ]; then
rm important-file.txt
# 验证: 文件是否确实被删除
if [ ! -f important-file.txt ]; then
echo "✅ 验证通过: 文件已删除"
else
echo "❌ 验证失败: 文件仍存在"
fi
else
echo "⚠️ 文件不存在,无需删除"
fi
案例分析
案例1:文件删除
❌ 错误方式:
用户: 删除文件
AI: 好的,我会删除的。 [然后没做]
✅ EVR正确方式:
🔧 Execute
$ rm /path/to/file.txt
[无输出,命令成功]
✅ Verify
$ ls -la /path/to/file.txt
ls: cannot access '/path/to/file.txt': No such file or directory
验证结果: ✅ 文件确实已删除
📋 Report
- 结果: 文件删除成功
- 数据: 删除了1个文件
- 验证: 文件不存在确认
- 下一步: 进行下一个任务
案例2:Git提交
❌ 错误方式:
用户: 提交更改
AI: 好的,已提交。 [实际有未提交文件]
✅ EVR正确方式:
🔧 Execute
$ git add -A
$ git commit -m "fix: 修复安全问题"
[master a1b2c3d] fix: 修复安全问题
2 files changed, 15 insertions(+)
✅ Verify
$ git status
On branch master
nothing to commit, working tree clean
$ git log --oneline -1
a1b2c3d fix: 修复安全问题
验证结果: ✅ 提交成功,工作区干净
📋 Report
- 结果: Git提交成功
- 数据: 2个文件,15行改动
- 提交: a1b2c3d
- 验证: working tree clean
- 下一步: 推送到远程
集成到AGENTS.md
## Execute-Verify-Report (EVR) 铁律
**所有任务必须遵循:**
1. **Execute** — 实际执行,不只是确认
- 必须看到命令输出
- 必须等待命令完成
- 必须捕获返回码
2. **Verify** — 验证结果正确
- 文件操作: 检查文件存在/内容/权限
- 网络操作: 检查响应状态码
- 系统操作: 检查进程/服务状态
- 数据操作: 检查数据准确性
3. **Report** — 汇报做了什么和验证结果
- 不要只说"完成了"
- 要说"完成了什么,如何验证的"
- 包含具体数据
**禁止:**
- ❌ "我会做的" ≠ 做了
- ❌ "完成了"但没有验证
- ❌ 最多重试3次,之后必须停止
- ❌ 静默失败
**违反EVR = 严重错误,立即纠正**
最佳实践
验证方法速查
| 操作类型 | 验证命令 | 通过标准 |
|---|---|---|
| 文件创建 | ls -la file |
文件存在 |
| 文件删除 | ls file 2>&1 |
文件不存在 |
| 文件修改 | cat file / md5sum |
内容/哈希匹配 |
| 权限修改 | stat -c "%a" file |
权限值正确 |
| Git提交 | git status / git log |
干净/提交存在 |
| 网络请求 | curl -I url |
HTTP 200 |
| 服务启动 | systemctl status svc |
active (running) |
| 进程运行 | pgrep process |
返回PID |
报告模板
✅ [任务名] 完成
**执行**: [命令]
**验证**: [方法] → [结果]
**数据**: [具体数字]
**下一步**: [行动]
License
MIT
Created by 小白 🤍
基于 STANDING-ORDERS 核心原则
"执行不是确认,验证不是可选,报告不是敷衍"
安全使用建议
This skill is primarily a set of runbooks and templates to force Execute→Verify→Report and is coherent with that purpose. Before using it: (1) Understand that following its templates will cause the agent to run shell commands that can read, change, or delete local files — do not run with root or elevated privileges unless you trust the agent and the specific commands. (2) The README mentions cloning an external GitHub repo and an 'evr' CLI helper, but no install or binaries are provided in the skill itself — if you choose to clone/run that repository, review its code first. (3) If you will allow autonomous execution, limit what the agent is permitted to run (sandboxing, non-root account, explicit allowlist of commands/paths). (4) If you need stronger assurance, ask the author for the 'evr' tool source or an install spec, or keep this skill as a manual guidance template only.
功能分析
Type: OpenClaw Skill
Name: evr-xiaobai
Version: 1.0.0
The bundle defines the 'Execute-Verify-Report' (EVR) framework, which is a set of instructions designed to improve AI agent reliability by mandating that the agent verify the results of its actions (e.g., checking if a file was actually deleted). The provided code snippets in SKILL.md and README.md are illustrative examples of bash commands and logic loops for verification and reporting; they do not contain malicious payloads, data exfiltration attempts, or harmful instructions.
能力评估
Purpose & Capability
The name/description match the content: SKILL.md is an instruction template enforcing Execute→Verify→Report. However, README and SKILL.md reference an external 'evr' CLI and a GitHub repo clone (https://github.com/aptratcn/skill-evr.git) even though the package has no install spec or shipped binaries; this is a minor inconsistency (documentation suggests an optional helper that is not provided).
Instruction Scope
The instructions explicitly tell an agent to run shell commands (rm, chmod, git, curl, systemctl, etc.) and to read/check local paths (example: ~/.openclaw/openclaw.json). That behavior is coherent with the skill's goal (forcing verification of actions) but means the agent will be asked to access and modify local system state and possibly sensitive files; the templates grant broad freedom to run arbitrary commands if followed literally.
Install Mechanism
No install spec and no code files — lowest install risk. README suggests cloning a GitHub repo as an optional step, but the skill itself does not install or fetch code automatically; if a user follows the README and runs the clone/install, that would introduce external code which should be reviewed.
Credentials
The skill declares no required environment variables, credentials, or config paths. Despite that, the examples encourage accessing local config and system services (files under ~/.openclaw, systemctl, deleting files). No secrets are requested, but the runtime behavior implies access to potentially sensitive local data.
Persistence & Privilege
No elevated privileges requested: always is false, user-invocable true, autonomous invocation allowed (platform default). The skill does not request persistent presence or modify other skills or global agent config in the provided files.
如何使用
- 确保已安装 OpenClaw(本地或 Docker 部署)
- 在对话框中输入安装命令:
/install evr-xiaobai - 安装完成后,直接呼叫该 Skill 的名称或使用
/evr-xiaobai触发 - 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.0.0
Initial release to ClawHub
元数据
常见问题
EVR Framework 是什么?
Execute-Verify-Report 三步法工具 — 强制AI遵循"执行→验证→报告"流程,杜绝"说做了其实没做"、"完成了但不验证"等问题。基于STANDING-ORDERS核心原则设计。 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 127 次。
如何安装 EVR Framework?
在 OpenClaw 或 Claude Code 对话框中运行命令「/install evr-xiaobai」即可一键安装,无需额外配置。
EVR Framework 是免费的吗?
是的,EVR Framework 完全免费,采用 MIT-0 许可证,可自由下载、安装和使用。
EVR Framework 支持哪些平台?
EVR Framework 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。
谁开发了 EVR Framework?
由 Erwin(@aptratcn)开发并维护,当前版本 v1.0.0。
推荐 Skills