Adversarial Review

开发完成后的三方对抗式代码评审闭环:蓝军(敌意审查,须给证据链)→ 第三方(独立审计,不信双方文档、只读代码,专找"修改引入的新缺陷"并稽核蓝军漏审的维度)→ 中立裁定(逐条采纳/驳回、审查双方共同前提)。当用户说"蓝军评审""第三方复核""对抗性审查""红蓝互搏""三方评审""交叉验证""发布前审查""帮我挑刺""adversarial review""red team review""pre-release review""critique this""find bugs in my code"或要求提升代码质量、准备发布前把关时使用。也适用于用户想把当前会话变成"多个 Agent 互搏"来提升代码水平的场景。

5,442 stars 816 forksSecurity

Install

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

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

From the SKILL.md

# 三方对抗式代码评审(Blue Team / Third Party / Adjudication)

## 这个 skill 要解决什么

单人或单 AI 写代码,最大的风险不是"不会写",而是:

1. **用未验证的乐观假设说服自己推迟修复**("这个兜底应该够了吧");
2. **测试提供假安全感**(测试全绿,但功能其实早坏了);
3. **为了修 A 而引入 B**(新写的修复代码本身没被独立审视)。

这三个问题**靠"再仔细看一遍"解决不了**——写代码的人和审代码的人是同一个,就一定会为自己辩护。只能靠**结构化的对抗关系**:

| 角色 | 立场 | 硬性约束 |
|---|---|---|
| **蓝军** | 敌意审查——假设代码会出事,去**证明** | 每条结论必须附证据链(`文件:行号` / 实测命令 / 上游源码);🔴/🟠 必须是**实测**结论,仅凭推理最高 🟡;禁止空话 |
| **第三方** | 独立审计——**不信任任何一方文档**,直接读代码 | 逐条复核"声称修了的是否真修了"、**专找修改引入的新缺陷**、并稽核蓝军**漏掉了哪些维度** |
| **中立裁定** | 对双方结论**逐条采纳 / 驳回 / 改级** | 必须由**与蓝军不同的 agent** 担任;当事人不能当法官 |

**关键不是"多找两个 AI 看看",而是顺序与互不信任**:第三方不信开发团队,裁定方不信蓝军和第三方。任一环节缺失,闭环就破。

### 配套文件(按需读取)

- **`references/review-dimensions.md`** —— **完整质量维度清单(19 项)**。第 0 步按改动类型选维度时必须读,否则审查会漏掉大片区域。
- **`references/prompt-templates.md`** —— **三个角色的完整提示词模板 + 占位符约定 + 「设计要点」表(改模板前必读)**。第 1/2/3 步派发子代理时直接取用,不要临时自编。
- **`references/quickstart.md`** —— 面向用户的上手说明(档位选择、什么时候不该用它)。
- **`examples/sample-review.md`** —— 一次真实评审的产出样例。

> ⚠️ **能力边界(必须向用户如实说明)**
> 子代理与主代理通常是**同一模型**,这不是真正独立的第三方。其价值来自**角色约束 + 强制证据**,而非"另一个 AI 的观点"。
> - ✅ 能有效抓出:代码级错误、逻辑漏洞、测试造假、自相矛盾、遗漏分支、声明与实现不符;
> - ❌ 抓不出:**双方共有的知识盲区**(例如对某个上游行为的一致误解)。若某个结论依赖外部系统行为,必须要求**实测验证**,而不是两个 agent 互相点头。
>
> 向用户汇报时,不要说"经独立第三方验证无误",要说"经同模型不同角色的对抗审查"。

> 🔁 **降级模式(宿主无子代理能力时)**
> 若宿主不提供 `subagent`,只能由**同一个 agent 串行扮演三角色**。此时:
> - 独立性**已被显著削弱**,第 2、3 步的价值大幅下降;
> - 必须在报告开头显式声明"本次为降级模式,角色由同一 agent 串行扮演";
> - 不得使用"独立复核""第三方验证"等措辞;
> - 有条件时优先换一个有子代理能力的宿主,而不是将就。

---

## 执行流程

### 第 0 步:确定档位、范围与维度
Read the full skill on GitHub

Details

Skill name
adversarial-review
Path
builtin-skills/skills/adversarial-review/SKILL.md
License
Apache-2.0
Repo updated
18 November 2026

FAQ

What does the Adversarial Review skill do?

开发完成后的三方对抗式代码评审闭环:蓝军(敌意审查,须给证据链)→ 第三方(独立审计,不信双方文档、只读代码,专找"修改引入的新缺陷"并稽核蓝军漏审的维度)→ 中立裁定(逐条采纳/驳回、审查双方共同前提)。当用户说"蓝军评审""第三方复核""对抗性审查""红蓝互搏""三方评审""交叉验证""发布前审查""帮我挑刺""adversarial review""red team review""pre-release review""critique this""find bugs in my code"或要求提升代码质量、准备发布前把关时使用。也适用于用户想把当前会话变成"多个 Agent 互搏"来提升代码水平的场景。

How do I install the Adversarial Review skill?

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

Is Adversarial Review popular?

It lives in iflytek/skillhub, which has 5,442 GitHub stars and 816 forks.

Related skills

Frontend Design

5.4k

iflytek/skillhub

Guidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.

Frontend & Design816

AI Claim Checker

5.4k

iflytek/skillhub

Evaluate factual claims in AI-generated text and teach a lightweight verification habit. Use when a learner wants to fact-check an AI answer, identify uncertainty, choose appropriate independent sources, or practise critical AI literacy.

AI, LLMs & Agents816

Cue Omni Reader

5.4k

iflytek/skillhub

Parse and understand an HTTP(S) URL or an authorized local document, audio, or video source through Cue Omni Reader when the Agent has the official Omni MCP tools.

Images, Video & Audio816

Best Practices

7.2k

tech-leads-club/agent-skills

Apply modern web development best practices for security, compatibility, and code quality. Use when asked to "apply best practices", "security audit", "modernize code", "code quality review", or "check for vulnerabilities". Do NOT use for accessibility (use web-accessibility), SEO (use seo), performance (use core-web-vitals), or comprehensive multi-area audits (use web-quality-audit).

Security568

Security Threat Model

7.2k

tech-leads-club/agent-skills

Repository-grounded threat modeling that enumerates trust boundaries, assets, attacker capabilities, abuse paths, and mitigations, and writes a concise Markdown threat model. Use when the user asks to threat model a codebase or path, enumerate threats or abuse paths, or perform AppSec threat modeling. Do NOT use for general architecture summaries, code review, security best practices (use security-best-practices), or non-security design work.

Security568

Security Best Practices

7.2k

tech-leads-club/agent-skills

Perform language and framework specific security best-practice reviews and suggest improvements. Use when the user explicitly requests security best practices guidance, a security review or report, or secure-by-default coding help. Supports Python, JavaScript/TypeScript, and Go. Do NOT use for general code review, debugging, threat modeling (use security-threat-model), or non-security tasks.

Security568

Browse the most popular Claude skills, all security skills, or write your own with the skills generator.