Anti-Vibecoding — 一个开发前门禁型 Agent Skill。在任何人动手写代码之前，强制先通过一套审查逻辑，杜绝“随时随地、随心所欲的冲动式编程”。 Anti-Vibecoding — a pre-development gatekeeper Agent Skill that forces a review before any code is written, to stop impulsive “vibe coding”.

中文 · English

---

中文

这是什么

绝大多数“我想做个应用”的冲动，要么已有成熟方案可复用，要么本质只是一段提示词 / 一个 Skill / 一个专家角色就能解决——并不需要一整个应用。Anti-Vibecoding 把这个判断变成一套可被强制执行的审查流程：通关前，绝不写任何代码。

两种模式

模式	适用场景	产出
Mode 1 · 快速自检	想法较轻，只想“停一停”	15 道多选题（5 维度 × 3 题）→ 冲动指数 0–100 + 开发者人格画像 + 红/黄/绿裁决；高分自动升级到 Mode 2
Mode 2 · 深度梳理	想法较重，需要认真论证	adversarial 五关 “grill me” 访谈 → 开源工具清单 + 初步 PRD，供二次开发

五道关卡（Mode 2）

1. G1 动机溯源 — 用一句话描述真正痛点（不含方案字眼）。
2. G2 市场调研 — 现成商用 / 开源为何不用（必须实际检索，不接受“应该没有”）。
3. G3 能力评估 — 技术栈是否真正掌握，维护成本能否承受。
4. G4 使用场景 — 开发完你实际会用吗？频率？ROI 是否值得？
5. G5 方案对比 — 为何必须是“应用”而非提示词 / Skill / 专家角色？更轻量方案能否达到同样目的？

裁决逻辑：五关全过 → PROCEED；有任意 FAIL → ABORT / 转轻量方案；混合 → PAUSE。通关前绝不写任何代码。

安装

# WorkBuddy 用户级 skills 目录（macOS 示例）
git clone https://github.com/anyeduke11/anti-vibecoding.git ~/.workbuddy/skills/anti-vibecoding

或直接下载 Releases 中的 anti-vibecoding-v1.2.0.zip 解压到 skills 目录。

使用

• 当你表达构建意图时会被自动触发（如“帮我做个应用”“我想开发…”“写个脚本”，或尚未说清问题就跳到框架选型）；
• 也可显式调用：用 anti-vibecoding 审一下 / grill me / /anti-vibecoding。

HTML 报告

每次审查结束会自动生成一份科技风 light 的自包含 HTML 报告（anti-vibecoding-report-YYYYMMDD.html），由 scripts/build_report.py（纯标准库，零依赖）渲染：

python3 scripts/build_report.py --data review.json
# 或： cat review.json | python3 scripts/build_report.py
# Windows 用 python 或 py

配色由 assets/design-tokens.css 统一驱动（tech-light：浅蓝 + 淡青），与遵循同一 Token 约定的其它 skill 对齐。

文件结构

anti-vibecoding/
├── SKILL.md                  # 门禁主逻辑 + 双模式 + 五关 + 裁决
├── assets/
│   ├── design-tokens.css     # 统一 design token（tech-light 配色）
│   └── readme/               # README 视觉资产（hero / workflow SVG）
├── references/
│   ├── grill-bank.md         # Mode 2 每关深挖提问 / 反驳话术 / 红旗启发
│   ├── quick-check-bank.md   # Mode 1 题库（15 题）+ 6 对话风格 + 计分/人格
│   ├── decision-record.md    # 裁决记录（markdown）模板
│   └── report-schema.md      # HTML 报告 JSON 数据契约
└── scripts/
    └── build_report.py       # 渲染 tech-light HTML 报告

版本

当前 v1.2.0。变更详见仓库 Releases。

License

MIT

---

English

What is this

Most “I want to build an app” urges either already have a mature solution, or are really just a prompt / a skill / an expert role in disguise — not a full application. Anti-Vibecoding turns that judgment into an enforceable review flow: no code until you pass.

Two modes

Mode	When	Output
Mode 1 · Quick Check	a light idea, just “pause and think”	15 multiple-choice questions (5 dimensions × 3) → Impulse Index 0–100 + developer persona + red / yellow / green verdict; high scores auto-escalate to Mode 2
Mode 2 · Deep Dialogue	a heavy idea that needs real justification	adversarial five-gate “grill me” interview → existing OSS tools + a preliminary PRD for secondary development

The five gates (Mode 2)

1. G1 Motivation — one sentence, real pain point (no solution words).
2. G2 Market — why not use an existing product / OSS (actually search; “probably none” is rejected).
3. G3 Capability — do you truly master the stack; can you afford maintenance.
4. G4 Use-case — will you actually use it; frequency; ROI.
5. G5 Solution — why must it be an app vs. a prompt / skill / expert role?

Verdict: all pass → PROCEED; any FAIL → ABORT / lighter alternative; mixed → PAUSE. No code until you pass.

Install

git clone https://github.com/anyeduke11/anti-vibecoding.git ~/.workbuddy/skills/anti-vibecoding

Or download anti-vibecoding-v1.2.0.zip from Releases and extract into your skills directory.

Usage

• Auto-triggers when you signal intent to build something new (“帮我做个应用”, “我想开发…”, “写个脚本”), especially when under-specified or solution-first.
• Or invoke explicitly: 用 anti-vibecoding 审一下 / grill me / /anti-vibecoding.

HTML report

A tech-light, self-contained HTML report is generated after each review (anti-vibecoding-report-YYYYMMDD.html), rendered by scripts/build_report.py (stdlib-only, zero deps):

python3 scripts/build_report.py --data review.json

Styling is driven by assets/design-tokens.css (tech-light palette).

File structure

anti-vibecoding/
├── SKILL.md                  # gatekeeper logic + two modes + five gates + verdict
├── assets/
│   ├── design-tokens.css     # unified design tokens (tech-light palette)
│   └── readme/               # README visual assets (hero / workflow SVG)
├── references/
│   ├── grill-bank.md         # Mode 2 deep probes / pushback / red flags
│   ├── quick-check-bank.md   # Mode 1 question bank (15) + 6 styles + scoring
│   ├── decision-record.md    # decision-record (markdown) template
│   └── report-schema.md      # HTML report JSON schema
└── scripts/
    └── build_report.py       # renders the tech-light HTML report

Version

Current v1.2.0. See Releases for the changelog.

License

MIT