adversarial-code-review
/adversarial-code-review
Use when reviewing a PR, diff, or proposed change and the user wants a strict adversarial pass — not a rubber stamp. Default antagonistic stance: find what is wrong in the PR itself and what new problems merge could introduce. Soft Karpathy-inspired checks and Prompt 六大(辩论、预演失败). Review only unless the user asks to fix.
- Author
- rongxinzy ↗
- Repository
- rongxinai ↗
- Licence
- AGPL-3.0
- Last repository change
- 27.10.2026
Assessment and security
- Files in this package
- 1
- Scripts
- 0
- HTML comments
- 0
- URLs in SKILL.md
- 0
No signals detected. Repository changed on 27.10.2026, without references to retired models or APIs.
AGPL-3.0, copyleft: przeróbki udostępniane innym muszą mieć tę samą licencję.
What it does and when to use it
Use when reviewing a PR, diff, or proposed change and the user wants a strict adversarial pass — not a rubber stamp. Default antagonistic stance: find what is wrong in the PR itself and what new problems merge could introduce. Soft Karpathy-inspired checks and Prompt 六大(辩论、预演失败). Review only unless the user asks to fix.
Original author description
Use when reviewing a PR, diff, or proposed change and the user wants a strict adversarial pass — not a rubber stamp. Default antagonistic stance: find what is wrong in the PR itself and what new problems merge could introduce. Soft Karpathy-inspired checks and Prompt 六大(辩论、预演失败). Review only unless the user asks to fix.
How it works
- Overview
- Typical Requests
- Soft Stance
- Soft Karpathy Map
- Soft Prompt 心法
- Review Dimensions
- Output Format
- Workflow
- Anti-Patterns
- Composition
SKILL.md preview
--- name: adversarial-code-review description: Use when reviewing a PR, diff, or proposed change and the user wants a strict adversarial pass — not a rubber stamp. Default antagonistic stance: find what is wrong in the PR itself and what new problems merge could introduce. Soft Karpathy-inspired checks and Prompt 六大(辩论、预演失败). Review only unless the user asks to fix. --- # 对抗式 Code Review(Adversarial Code Review) ## Overview 以**反对者**立场审查变更:默认「这版改动可能有问题」,要用证据才能放过。目标是找出真实风险,不是堆严厉语气,也不是帮作者圆场。 本 skill 是**软约束**:优先按对抗清单走;若变更极小且风险面清晰,可缩短输出,但不要无证据直接说「没问题」。 ## Typical Requests - 「帮我 review 这个 PR / 这段 diff」 - 「站在反对者角度挑刺」 - 「合并前看看会不会埋雷」 - 父代理审查 subagent 提交的改动 ## Soft Stance 1. **默认对抗**:不假设作者正确;每个关键结论要有 diff / 行为 / 测试上的依据。 2. **两层提问**(必须都想一遍): - **本 PR 自身**:逻辑、做法、边界、测试、可读性有什么问题? - **合并之后**:会引入哪些新问题(回归、耦合、运维、安全、性能、后续维护成本)? 3. **只审查**:默认只给结论与修改建议;**除非用户明确说「按你的意见改」**,否则不擅自改代码。 4. **可降级**:纯文档错字、单行重命名等,可用短评;仍要扫一眼合并后影响。 ## Soft Karpathy Map 审查时软性对照四条(不必逐条刻板打分,但心里要过一遍): | 条 | 问什么 | | --- | --- | | 假设 | 作者默认了什么?这些假设在代码/环境里成立吗? | | 简单 | 有没有更简单、同效的做法?复杂度是否与问题匹配? | | 手术式改动 | 是否只改该改的?有无顺手重构、无关文件、scope 膨胀? | | 目标可验证 | 成功标准是否清楚?现有测试/验证步骤能否证明? | ## Soft Prompt 心法 - **辩论**:主动站在反方,攻击方案弱点;再看反方论点是否被证据化解。 - **预演失败**:假设已合并上线,设想 2~3 个最可能的失败路径(错误输入、并发、空数据、权限、回滚困难等),看本 PR 是否挡住。 ## Review Dimensions 按需覆盖(与改动相关的优先): - 正确性与根因是否对症(是否只治症状) - 边界条件、错误处理、空值/并发/重试 - 安全(注入、权限、密钥、不可信输入) - 性能与资源(热路径、N+1、无界增长) - 可读性与命名、模块边界 - 测试缺口与验证是否可复现 - 兼容性 / 迁移 / 特性开关 - 可观测性(日志、指标、失败是否可诊断) ## Output Format 用中文,结构固定,便于扫读: ```markdown ## 结论 通过 | 要改 | 再讨论 (一句话说明为什么) ## 致命 - …(必须修才能合;写清位置/证据/风险) ## 重要 - …(强烈建议修;写清原因) ## 建议 - …(可选改进,不阻塞) ## 合并后风险(预演失败) - …(合并后可能出现的新问题) ## 测试缺口 - …(缺什么用例 / 怎么补验证) ``` 规则: - **没有某类问题时**:写「无」或省略该节,不要凑数。 - **每条意见尽量带锚点**:文件、符号、行为路径,或引用 diff 片段。 - **区分事实与推测**:不确定标 `[推测]`,并写还需要什么证据。 - 结论为「通过」时,仍可列少量建议;不要为了显得严格而捏造致命项。 ## Workflow 1. 读清范围:目标分支、PR 描述、commit 列表、实际 diff(不要只看标题)。 2. 先复述:作者声称解决了什么、根因/解法是否自洽。 3. 第一层:本 PR 内的正确性与做法问题。 4. 第二层:合并后新问题 + 预演失败。 5. 软对照 Karpathy 四条,收束到 致命 / 重要 / 建议。 6. 给出结论;若用户要你修,再另开实现流程(可配合 `clarify-before-coding` 与 `git-pr-workflow`)。 ## Anti-Patterns - 客气到只说「整体不错」却无任何具体检查痕迹 - 风格偏好冒充致命缺陷 - 只挑命名/格式,不碰逻辑与合并后风险 - 未受请求就直接改作者分支 - 把软清单当成必须填满的表格(无问题硬写三条) ## Composition - 开发澄清与小 PR:`clarify-before-coding` - 提交/PR 规范:`git-pr-workflow` - 需要作者补计划时:要求 PR 描述含 问题 / 根因 / 解法 / 验证
Versions
| Repository | Score | Licence |
|---|---|---|
| This page · rongxinzy/rongxinai | 100 | AGPL-3.0 |
| aws-solutions/konductor | 100 | Apache-2.0 |
Database snapshot: 13.11.2026 · package version 96ce9dd55b7e (SHA-256 digest of the entire package)
Downloads in the last 30 days: 0 · ranking
Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.
Similar skills
- 100adversarial-review
Adversarially review finished work that another model family built - hunt for bugs, risky assumptions, missed requirements and security problems, rank findings by severity with path:line evidence, and never rubber-stamp. Never edits, commits or pushes. Use when asked for an… - 100adversarial-review
Adversarially hunt for correctness bugs and regressions in a change set. Use when reviewing a diff before finalizing, during a cleanup pass, or when asked to find bugs, pressure-test, or stress-test recently modified code. Triggers on "find bugs", "what could break", "review for… - 90adversarial-review
Runs a risk-first review after design and implementation. Use when a change is medium or high risk, customer-facing, security-sensitive, or hard to roll back safely. - 100codex-review
Run Codex code review with Claude's independent double-check. Use when asked \"codex review\", \"review my code with codex\", or wants Codex to review code changes. For adversarial review use /codex-adversarial. - 96code-review
Use when the user asks for a code review of a change, branch or PR. - 90adversarial-qa
Use only when the user explicitly requests adversarial, hostile, end-to-end, or release-gate QA beyond the project's normal verification suite.