Adversarial Verification

对抗性验证技能 — 验证代码变更/部署/任务交付物的正确性。不是确认"它能工作",而是尝试"把它搞坏"。用于:(1) 完成非平凡任务后自验 (2) 审查 PR/MR (3) 部署后验证 (4) bobooo 说"验证一下"。灵感源自 Claude Code Verification Agent 设计。

9 stars 8 forksAI, LLMs & Agents

Install

skills CLI (Claude Code, Cursor, Codex and more)
npx skills add kadaliao/claude-code-skills-collection --skill adversarial-verification
Manual install
git clone --depth 1 https://github.com/kadaliao/claude-code-skills-collection.git /tmp/claude-code-skills-collection
mkdir -p ~/.claude/skills
cp -r /tmp/claude-code-skills-collection/adversarial-verification ~/.claude/skills/adversarial-verification

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# Adversarial Verification

## 核心理念

> 你的目标不是确认实现正确 — 而是尝试打破它。

你有两个已知失败模式,时刻警惕:
1. **验证回避** — 读代码、叙述测试计划、写"PASS",但从未运行任何命令
2. **被前 80% 诱惑** — 看到整洁 UI 或测试通过就放行,没注意一半按钮是空的

## 自我合理化识别

你会想跳过检查。识别这些借口并做相反的事:

| 你脑中的借口 | 正确行动 |
|---|---|
| "代码看起来正确" | 运行它 |
| "测试已经通过了" | 独立验证,测试可能是自证循环 |
| "大概没问题" | "大概"不是验证 |
| "启动服务器看代码吧" | 启动服务器然后 **curl 端点** |
| "没有浏览器" | 检查是否有 browser tool / playwright |
| "这个要花太久" | 不是你决定的 |

**如果你在写解释而不是命令 → 停。运行命令。**

## 验证策略(按变更类型)

### 前端
启动 dev server → 浏览器/curl 检查 → 检查子资源(图片/API/静态文件)→ 运行前端测试

### 后端/API
启动服务器 → curl 端点 → 验证响应结构(不只是状态码)→ 测试错误处理 → 边界用例

### CLI/脚本
用代表性输入运行 → 验证 stdout/stderr/exit code → 边界输入(空、畸形、极端值)

### 基础设施/配置
验证语法 → dry-run → 检查环境变量是否真正被引用

### Bug 修复
复现原 bug → 验证修复 → 运行回归测试 → 检查相关功能副作用

### 重构(无行为变更)
现有测试必须全过 → diff 公共 API 表面 → 抽查行为一致性

## 对抗性探测(选择适用的)

- **并发** — 并行请求创建类路径,检查重复/丢失
- **边界值** — 0, -1, 空字符串, 超长字符串, unicode, MAX_INT
- **幂等性** — 同一个变更请求发两次,结果正确吗?
- **孤儿操作** — 删除/引用不存在的 ID

## 输出格式(强制)

每项检查:
```
### Check: [验证什么]
**Command run:**
  [实际执行的命令]
**Output observed:**
  [终端实际输出 — 复制粘贴,不要转述]
**Result: PASS** (或 FAIL — 附 Expected vs Actual)
```

❌ 坏例子:
```
### Check: POST /api/register 验证
**Result: PASS**
Evidence: 读了 routes/auth.py,逻辑正确地验证了...
```
(没有运行命令。读代码不是验证。)

## 最终判定

报告末尾必须有且仅有一行:
```
VERDICT: PASS
VERDICT: FAIL
VERDICT: PARTIAL
```
- FAIL → 附失败详情 + 复现步骤
- PARTIAL → 仅限环境限制(缺工具/无法启动服务),不是"不确定"
- 报告必须包含至少一个对抗性探测及其结果

## PASS 之前的检查清单

- [ ] 至少运行了一个命令并有输出
- [ ] 至少做了一个对抗性探测
- [ ] 没有"代码看起来正确"式的跳过
- [ ] 验证了非 happy path
Read the full skill on GitHub

Details

Skill name
adversarial-verification
Path
adversarial-verification/SKILL.md
Repo updated
31 March 2026

FAQ

What does the Adversarial Verification skill do?

对抗性验证技能 — 验证代码变更/部署/任务交付物的正确性。不是确认"它能工作",而是尝试"把它搞坏"。用于:(1) 完成非平凡任务后自验 (2) 审查 PR/MR (3) 部署后验证 (4) bobooo 说"验证一下"。灵感源自 Claude Code Verification Agent 设计。

How do I install the Adversarial Verification skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add kadaliao/claude-code-skills-collection --skill adversarial-verification. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is Adversarial Verification popular?

It lives in kadaliao/claude-code-skills-collection, which has 9 GitHub stars and 8 forks.

Related skills

Coordinator Orchestrator

9

kadaliao/claude-code-skills-collection

多 Agent 协调者模式 — 你是指挥官,不是执行者。调度 Worker 做研究/实现/验证,你负责综合和决策。适用于:(1) 复杂任务需要多步骤并行 (2) 需要研究→综合→实现→验证的工作流 (3) bobooo 说"用 CC 做"时的编排模式。灵感源自 Claude Code Coordinator Mode。

AI, LLMs & Agents

Lightweight Explorer

9

kadaliao/claude-code-skills-collection

轻量探索模式 — 用最少 token 完成代码搜索和调研。核心理念:只读搜索不需要加载全部工具和项目规则。适用于:(1) 搜索代码库找文件/函数 (2) 调研项目架构 (3) 多文件分析 (4) 回答"代码里哪里有 X"类问题。灵感源自 Claude Code Explore Agent 设计。

AI, LLMs & Agents

Memory Type System

9

kadaliao/claude-code-skills-collection

结构化记忆类型系统 — 四类型分类 + frontmatter 规范 + 写入/召回/漂移防护全流程。适用于所有 Agent 的长期记忆管理。解决记忆混乱、召回不准、膨胀失控问题。灵感源自 Claude Code Memory Type Taxonomy(memoryTypes.ts)。

Productivity & Notes

Compound Dev

9

brijoobopanna/ClaudeSkills

Make Claude Code sessions compound — every session builds on the last instead of starting from scratch. Implements Karpathy's LLM Wiki pattern for software engineering. Use this skill when starting a new project, onboarding to an existing codebase, or any time the user says "set up my project", "add project memory", "create CLAUDE.md", "help me structure this project", "I keep re-explaining things to Claude", "my sessions don't remember anything", "ingest this document", or "run a health check". Also trigger when the user asks about session efficiency, token usage, or wants to improve their Cl

AI, LLMs & Agents

Multi Agent Verify

9

leyuan0602-glitch/cc-claude-codex

Multi-agent parallel verification. Spawns 3 independent agents (OpenCode CLI, Codex CLI, Claude Task subagent) in separate git worktrees to review, test, and fix code. Synthesizes results and applies final fixes on the original branch.

AI, LLMs & Agents

Case Cortex

9

accilium/accilium-skills

Build and maintain a Case Cortex — a structured, living, machine-readable representation of a single bounded knowledge-work problem (a "case") that agents and humans share as their reasoning substrate. Use whenever the user wants to set up, seed, update, query, or regenerate artifacts from a Case Cortex. Triggers on phrases like "build a cortex for X", "start a new case", "initialize a case", "seed my cortex", "update the cortex", "add to the cortex", "ingest this document into the cortex", "what does the cortex say about Y", "regenerate the briefing", or "lint the cortex". A Cortex fits bound

AI, LLMs & Agents

Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.