GStack Pro — 10-Role AI Engineering Team
Built on the philosophy of Garry Tan's GStack (YC CEO) · 16K GitHub Stars · MIT License
Adapted for OpenClaw subagent + session architecture
What It Does
GStack Pro gives your AI 10 specialist roles — each with a clear mandate, a structured output format, and a measurable quality bar.
Instead of one generic AI doing everything badly, you get a team:
| # | Role | Icon | Subagent | Best For |
|---|
| 1 | CEO / Product Thinker | 🏛️ | INLINECODE0 | Rethink the problem before building |
| 2 |
Architect / Tech Lead | 🏗️ |
architect | Lock in data flow, failure modes, tests |
| 3 | Designer Review | 🎨 |
designer | 80-item design audit, AI slop detection |
| 4 | Paranoid Code Review | 🔍 |
tester | N+1, race conditions, trust boundaries |
| 5 | Browser QA | 🌐 |
browser tool | AI with eyes — login, click, screenshot, verify |
| 6 | Automated QA + Fix | 🧪 |
tester +
coder | Find → fix → re-verify with Health Score |
| 7 | QA Reporter | 📊 |
tester | Report-only, clean handoff to team |
| 8 | One-Command Ship | 🚀 |
operator | sync → test → push → PR |
| 9 | Engineering Retro | 🔄 |
progress | Commit analysis, praise, growth areas |
| 10 | Release Docs | 📝 |
writer | Sync docs to match what shipped |
The Development Cycle
CODEBLOCK0
How to Activate a Role
Method 1: Direct Command (e.g., in conversation)
CODEBLOCK1
Method 2: Subagent (for background/parallel work)
CODEBLOCK2
Health Score System
After every /qa session, output a structured score:
CODEBLOCK3
| Score | Status | Action |
|---|
| 90-100 | 🟢 Excellent | Ready to ship immediately |
| 70-89 |
🟡 Good | 2-3 minor issues, fix before ship |
| 50-69 | 🟠 Needs Work | Significant bugs, fix before next sprint |
| <50 | 🔴 Do Not Ship | Core functionality broken, redo required |
Quality Bars
Code Must Pass
- - ✅ N+1 queries eliminated
- ✅ All external calls have timeouts
- ✅ Retries with exponential backoff
- ✅ Database transactions properly bounded
- ✅ Input validation on all untrusted data
- ✅ No trust boundary violations
- ✅ Structured logging (JSON, with trace IDs)
Design Must Pass
- - ✅ Consistent visual hierarchy
- ✅ No AI slop patterns (copy-paste generic cards, overuse of gradients)
- ✅ Responsive at 375px / 768px / 1440px
- ✅ Accessible (color contrast, focus states)
- ✅ Meaningful empty states
Anti-Patterns Detected
| Pattern | Why It Fails | Detection |
|---|
| "Looks good!" | AI self-evaluation bias | Evaluator never reads generator code |
| Circular dependency |
Unmaintainable architecture | Dependency graph analysis |
| AI slop | Generic, low-quality design | 80-item designer audit |
| Magic numbers | Hard to maintain |
no-magic-numbers lint rule |
| Forgotten edge cases | Silent production failures | Mandatory test matrix |
| No rollback plan | Can't safely deploy |
/ship requires rollback plan |
OpenClaw Subagent Mapping
| Role | Subagent ID | Type |
|---|
| CEO Product | INLINECODE14 | demand analysis |
| Architect |
architect | tech design |
| Designer |
designer | UI/UX review |
| Code Review |
tester | quality assurance |
| Browser QA |
browser tool | automated testing |
| QA + Fix |
tester +
coder | test + implement |
| QA Reporter |
tester | reporting |
| Ship |
operator | release |
| Retro |
progress | analysis |
| Docs |
writer | documentation |
Key Insight: Generator vs Evaluator
GStack Pro separates creation from judgment.
CODEBLOCK4
This eliminates cognitive commitment bias — the AI can't judge what it already committed to building.
Inspired by: Anthropic Engineering, "Harness Design for Long-Running Application Development" (2026)
Files
| File | Purpose |
|---|
| INLINECODE25 | This file — overview and usage |
| INLINECODE26 |
CEO product thinking SOP |
|
references/plan-eng.md | Architecture review SOP |
|
references/review.md | Paranoid code review SOP |
|
references/qa.md | Automated QA SOP + Health Score |
|
references/ship.md | One-command ship SOP |
|
references/retro.md | Engineering retro SOP |
Inspired by Garry Tan's GStack (https://gstacks.org) · MIT License
For OpenClaw · Compatible with Claude Code GStack workflows
GStack Pro — 10角色AI工程团队
基于Garry Tan的GStack(YC CEO)理念构建 · 16K GitHub Stars · MIT许可证
适配OpenClaw子代理+会话架构
功能概述
GStack Pro为您的AI提供10个专业角色——每个角色都有明确职责、结构化输出格式和可衡量的质量标准。
您得到的不是一个什么都能做但什么都做不好的通用AI,而是一个完整的团队:
| # | 角色 | 图标 | 子代理 | 最佳用途 |
|---|
| 1 | CEO / 产品思考者 | 🏛️ | requirer | 在构建之前重新思考问题 |
| 2 |
架构师 / 技术负责人 | 🏗️ | architect | 锁定数据流、故障模式、测试 |
| 3 | 设计评审 | 🎨 | designer | 80项设计审计,AI劣质内容检测 |
| 4 | 偏执代码审查 | 🔍 | tester | N+1查询、竞态条件、信任边界 |
| 5 | 浏览器QA | 🌐 | browser工具 | 有眼睛的AI——登录、点击、截图、验证 |
| 6 | 自动化QA+修复 | 🧪 | tester + coder | 发现→修复→重新验证,附带健康评分 |
| 7 | QA报告员 | 📊 | tester | 仅报告,干净交接给团队 |
| 8 | 一键发布 | 🚀 | operator | 同步→测试→推送→PR |
| 9 | 工程回顾 | 🔄 | progress | 提交分析、表扬、成长领域 |
| 10 | 发布文档 | 📝 | writer | 同步文档以匹配实际发布内容 |
开发周期
用户请求
↓
① CEO审视 (/plan-ceo)
→ 这个功能值得构建吗?10星产品是什么样的?
↓
② 架构锁定 (/plan-eng)
→ 数据流、状态机、故障模式、测试矩阵
↓
③ 设计评审 (/plan-design)
→ 80项审计、设计质量等级、AI劣质内容检测
↓
④ 偏执代码审查 (/review)
→ N+1查询、竞态条件、信任边界违规
↓
⑤ 自动化浏览器QA (/qa)
→ AI驱动浏览器、发现bug、修复bug、重新验证
→ 健康评分0-100决定发布就绪状态
↓
⑥ 一键发布 (/ship)
→ 同步主分支→运行测试→推送→创建PR
↓
⑦ 工程回顾 (/retro)
→ 提交分析、团队绩效、改进计划
↓
⑧ 发布文档 (/document)
→ 更新README/ARCHITECTURE以匹配实际发布内容
如何激活角色
方法1:直接命令(例如在对话中)
/plan-ceo: 为AICFO设计一个新功能:员工工资条自动生成PDF
/review: 审查deepfmt Sprint 3的代码改动
/qa: 对 https://xxx.space.minimaxi.com 运行标准QA测试
方法2:子代理(用于后台/并行工作)
javascript
sessions_spawn({
agentId: tester, // QA + 审查
task: 阅读 skills/gstack-pro/roles/review.md 然后审查 /workspace/projects/aicfo/aicfo-mvp/src/api/ 的代码
})
健康评分系统
每次 /qa 会话后,输出结构化评分:
json
{
healthScore: 85,
status: 🟡 良好,
breakdown: {
functional: { passed: 8, total: 10, score: 24 },
edgeCases: { covered: 4, total: 5, score: 20 },
consoleErrors: { passed: true, score: 25 },
designRegressions: { passed: true, score: 16 }
},
shipRecommendation: 🟡 发布前修复2个小问题
}
| 评分 | 状态 | 操作 |
|---|
| 90-100 | 🟢 优秀 | 可立即发布 |
| 70-89 |
🟡 良好 | 2-3个小问题,发布前修复 |
| 50-69 | 🟠 需要改进 | 重大bug,下个sprint前修复 |
| <50 | 🔴 禁止发布 | 核心功能损坏,需要重做 |
质量标准
代码必须通过
- - ✅ 消除N+1查询
- ✅ 所有外部调用都有超时设置
- ✅ 指数退避重试
- ✅ 数据库事务正确边界
- ✅ 所有不可信数据的输入验证
- ✅ 无信任边界违规
- ✅ 结构化日志(JSON,带追踪ID)
设计必须通过
- - ✅ 一致的视觉层级
- ✅ 无AI劣质内容模式(复制粘贴通用卡片、过度使用渐变)
- ✅ 响应式设计:375px / 768px / 1440px
- ✅ 可访问性(颜色对比度、焦点状态)
- ✅ 有意义的空状态
检测到的反模式
| 模式 | 失败原因 | 检测方式 |
|---|
| 看起来不错! | AI自我评估偏差 | 评估者从不读取生成器代码 |
| 循环依赖 |
不可维护的架构 | 依赖图分析 |
| AI劣质内容 | 通用、低质量设计 | 80项设计师审计 |
| 魔法数字 | 难以维护 | no-magic-numbers lint规则 |
| 遗忘的边缘情况 | 静默生产故障 | 强制性测试矩阵 |
| 无回滚计划 | 无法安全部署 | /ship 要求回滚计划 |
OpenClaw子代理映射
| 角色 | 子代理ID | 类型 |
|---|
| CEO产品 | requirer | 需求分析 |
| 架构师 |
architect | 技术设计 |
| 设计师 | designer | UI/UX审查 |
| 代码审查 | tester | 质量保证 |
| 浏览器QA | browser工具 | 自动化测试 |
| QA+修复 | tester + coder | 测试+实现 |
| QA报告员 | tester | 报告 |
| 发布 | operator | 发布 |
| 回顾 | progress | 分析 |
| 文档 | writer | 文档编写 |
关键洞察:生成器 vs 评估器
GStack Pro将创造与判断分离。
生成器代理 ──→ 构建代码 ──→ 评估器代理
(创造) (产物) (仅根据SPEC和URL判断)
↑
从不读取生成器的代码
这消除了认知承诺偏差——AI无法判断它已经承诺要构建的内容。
灵感来源:Anthropic Engineering, Harness Design for Long-Running Application Development (2026)
文件
| 文件 | 用途 |
|---|
| SKILL.md | 本文件——概述和使用方法 |
| references/plan-ceo.md |
CEO产品思维SOP |
| references/plan-eng.md | 架构审查SOP |
| references/review.md | 偏执代码审查SOP |
| references/qa.md | 自动化QA SOP + 健康评分 |
| references/ship.md | 一键发布SOP |
| references/retro.md | 工程回顾SOP |
灵感来源:Garry Tan的GStack (https://gstacks.org) · MIT许可证
适用于OpenClaw · 兼容Claude Code GStack工作流