One person. 22 AI specialists. Full product lifecycle on autopilot.
一个人。22 个 AI 专家。全流程自动驾驶。
CrewPilot turns GitHub Copilot CLI into a virtual engineering team. Describe an idea, and it autonomously drives through the entire product lifecycle — ideation → design → architecture → implementation → testing → code review → security audit → documentation → deployment — pausing only at critical decision gates for your approval.
Inspired by: garrytan/gstack · CrewAI · Aider · CodeRabbit
# 1. Clone CrewPilot
git clone https://github.com/kylehuang0323-ai/CrewPilot-dev.git
# 2. Copy skills into your project
cp -r CrewPilot-dev/.github/skills/ your-project/.github/skills/
# 3. Launch Copilot CLI and start autopilot
copilot
> /autopilot
> "Build a todo API with JWT auth and real-time sync"
# 4. Type "go" at decision gates, or "fast" to auto-approve all| Skill | Role | What it does |
|---|---|---|
/autopilot |
Team Lead | Full lifecycle autopilot: 7 stages, 7 gates. Type "fast" for full auto. |
| Skill | Role | What it does |
|---|---|---|
/office-hours |
YC Partner | Six forcing questions + 5-lens innovation analysis → design doc |
/plan-ceo-review |
CEO | Four scope modes (expand/selective/hold/reduce), 11-section review |
/plan-eng-review |
Eng Manager | Architecture diagrams + data flow + test strategy + edge cases |
/multi-debate |
Design Panel | CEO/Engineer/User Advocate debate simultaneously, surface disagreements |
| Skill | Role | What it does |
|---|---|---|
/design-system |
Senior Designer | Colors/typography/spacing/components → DESIGN.md + Tailwind config |
/design-review |
Designer Who Codes | 6-dimension audit (0-10): accessibility/responsive/consistency/CSS/perf/compliance |
/ui-gen |
Frontend Engineer | Detect framework → read DESIGN.md → generate full-state components + tests |
| Skill | Role | What it does |
|---|---|---|
/test-gen |
Test Engineer | Bootstrap test framework + analyze source + generate unit/integration/edge tests |
/doc-gen |
Tech Writer | Staleness detection + auto-update README/API docs/Changelog |
| Skill | Role | What it does |
|---|---|---|
/review |
Staff Engineer | 7 parallel scanners (21 grep patterns) + auto-fix |
/adversarial-review |
Dual Reviewer | Defender + Attacker dual-agent review, cross-compare blind spots |
/cso |
CISO | OWASP Top 10 + STRIDE + secrets archaeology + dependency audit |
/investigate |
Senior Debugger | Iron Law debugging: git bisect + 5 pattern scanners + 3-Strike escalation |
/api-compat |
Contract Guardian | Breaking change detection: removed endpoints, type changes, new required params |
/benchmark |
Perf Engineer | Bundle size/API latency/memory before-after comparison + regression thresholds |
/preferences |
Preference Learner | Remember accept/reject decisions, auto-filter known false positives |
| Skill | Role | What it does |
|---|---|---|
/deploy |
Release Engineer | Semantic versioning + changelog + deploy + health check + rollback |
/dep-upgrade |
Supply Chain | Per-package vulnerable dependency upgrade + test + atomic commit + rollback |
| Skill | Role | What it does |
|---|---|---|
/dashboard |
Project Board | Zero-dependency HTML dashboard: pipeline status/timeline/findings/file tree |
/knowledge-base |
Knowledge Mgr | chronicle.jsonl + decision records + session summaries + KNOWLEDGE.md |
You: "Build a calendar AI assistant"
↓
Stage 1: Ideation ──→ 🚦 Gate 1: Approve design
Stage 2: Strategy ──→ 🚦 Gate 2: Scope decision (may auto-skip)
Stage 3: Architecture──→ 🚦 Gate 3: Approve architecture
Stage 4: Build ──→ 🚦 Gate 4: Review implementation
Stage 5: Review ──→ 🚦 Gate 5: Only [ASK] items need approval
Stage 6: Security ──→ 🚦 Gate 6: Only CRITICAL findings stop
Stage 7: Ship ──→ 🚦 Gate 7: Final confirmation
↓
✅ PR opened + docs updated + dashboard generated
Commands: go · skip · back to {stage} · pause · status · detail · history · fast · abort
_shared (1010 lines — shared infrastructure)
├── §0 Voice & writing standards
├── §0.5 Anti-hallucination guardrails (7 rules)
├── §0.9 Chronicle auto-logging protocol
├── §0.95 Atomic commit protocol (every change is revertable)
├── §0.96 Context memory (cross-session knowledge)
├── §0.97 GitHub Actions integration templates
├── §1 Project detection (4 project modes)
├── §2 Finding schema (universal JSON format)
├── §3 Confidence scale (anchored 1-10)
├── §4 Stage artifact contract (autopilot file paths)
├── §5 Greenfield mode (new project handling)
├── §6 Error handling standards
└── §7 Grep pattern library (21 named patterns)
Per-project (recommended):
cp -r CrewPilot-dev/.github/skills/ your-project/.github/skills/Global:
# Windows
xcopy /E /I CrewPilot-dev\.github\skills %USERPROFILE%\.copilot\skills
# macOS/Linux
cp -r CrewPilot-dev/.github/skills/ ~/.copilot/skills/CrewPilot 将 GitHub Copilot CLI 变成一支虚拟工程团队。描述一个想法，它会自动驾驶走完完整的产品生命周期 — 构思 → 设计 → 架构 → 实现 → 测试 → 代码审查 → 安全审计 → 文档 → 部署 — 仅在关键决策点暂停等你拍板。
灵感来源： garrytan/gstack · CrewAI · Aider · CodeRabbit
# 1. 克隆 CrewPilot
git clone https://github.com/kylehuang0323-ai/CrewPilot-dev.git
# 2. 复制技能到你的项目
cp -r CrewPilot-dev/.github/skills/ your-project/.github/skills/
# 3. 启动 Copilot CLI，运行自动驾驶
copilot
> /autopilot
> "做一个带用户认证的 Todo API，支持实时同步"
# 4. 在决策节点输入 "go" 继续，或 "fast" 全自动到底| 技能 | 角色 | 做什么 |
|---|---|---|
/autopilot |
团队编排器 | 全流程自动驾驶（7 阶段 7 Gate），输入 "fast" 全自动 |
| 技能 | 角色 | 做什么 |
|---|---|---|
/office-hours |
YC 合伙人 | 六个强制性问题 + 五透镜创新分析 → 设计文档 |
/plan-ceo-review |
CEO | 四种范围模式（扩展/选择/保持/缩减），11 节评审 |
/plan-eng-review |
工程经理 | 架构图 + 数据流 + 测试策略 + 边界情况 |
/multi-debate |
设计评审团 | CEO/工程师/用户代言人三角色同时辩论，暴露分歧 |
| 技能 | 角色 | 做什么 |
|---|---|---|
/design-system |
高级设计师 | 颜色/字体/间距/组件规范 → DESIGN.md + Tailwind 配置 |
/design-review |
会写代码的设计师 | 6 维度 0-10 评分：可访问性/响应式/一致性/CSS/性能/合规 |
/ui-gen |
前端工程师 | 检测框架 → 读 DESIGN.md → 全状态组件（loading/error/empty/success） |
| 技能 | 角色 | 做什么 |
|---|---|---|
/test-gen |
测试工程师 | 引导测试框架 + 分析源码 + 生成单元/集成/边界测试 |
/doc-gen |
技术文档 | 过期检测 + 自动更新 README/API 文档/Changelog |
| 技能 | 角色 | 做什么 |
|---|---|---|
/review |
Staff 工程师 | 7 并行扫描器（21 个 grep 模式）+ 自动修复 |
/adversarial-review |
双审查员 | Defender + Attacker 双 Agent 对抗审查，发现盲点 |
/cso |
首席安全官 | OWASP Top 10 + STRIDE + 密钥考古 + 依赖审计 |
/investigate |
高级调试员 | Iron Law 根因调试：git bisect + 5 模式扫描器 + 3-Strike 升级 |
/api-compat |
契约守护者 | API 破坏性变更检测（删端点/改类型/加必填参数） |
/benchmark |
性能工程师 | Bundle 大小/API 延迟/内存/前后对比 + 回归阈值 |
/preferences |
偏好学习器 | 记住 accept/reject 偏好，自动过滤已知误报 |
| 技能 | 角色 | 做什么 |
|---|---|---|
/deploy |
发布工程师 | 语义版本 + Changelog + 部署 + 健康检查 + 回滚 |
/dep-upgrade |
供应链安全 | 逐包升级有漏洞依赖 + 测试 + 原子 commit + 回滚 |
| 技能 | 角色 | 做什么 |
|---|---|---|
/dashboard |
项目看板 | 零依赖 HTML 看板：pipeline 状态/时间线/发现/文件树/决策 |
/knowledge-base |
知识库 | chronicle.jsonl + 决策记录 + 会话摘要 + KNOWLEDGE.md |
你: "做一个日历 AI 助手"
↓
阶段 1: 构思 ──→ 🚦 Gate 1: 审批设计方案
阶段 2: 战略 ──→ 🚦 Gate 2: 范围决策（可自动跳过）
阶段 3: 架构 ──→ 🚦 Gate 3: 确认架构
阶段 4: 实现 ──→ 🚦 Gate 4: 确认实现
阶段 5: 审查 ──→ 🚦 Gate 5: 仅 ASK 项需确认
阶段 6: 安全 ──→ 🚦 Gate 6: 仅 CRITICAL 需确认
阶段 7: 发布 ──→ 🚦 Gate 7: 最终确认
↓
✅ PR 已开 + 文档已更新 + 看板已生成
控制命令： go · skip · back to {stage} · pause · status · detail · history · fast · abort
_shared (1010 行 — 共享基础设施)
├── §0 写作规范 + 禁用词
├── §0.5 防幻觉护栏（7 条规则）
├── §0.9 Chronicle 自动日志协议
├── §0.95 Atomic Commit 全程原子提交
├── §0.96 Context Memory 跨会话记忆
├── §0.97 GitHub Actions 集成模板
├── §1 项目检测（4 种项目模式）
├── §2 Finding Schema（统一 JSON 格式）
├── §3 置信度标尺（锚定 1-10）
├── §4 阶段产物约定（autopilot 文件路径）
├── §5 Greenfield 模式（新项目处理）
├── §6 错误处理标准
└── §7 Grep 模式库（21 个命名模式）
项目级安装（推荐）：
cp -r CrewPilot-dev/.github/skills/ your-project/.github/skills/全局安装：
# Windows
xcopy /E /I CrewPilot-dev\.github\skills %USERPROFILE%\.copilot\skills
# macOS/Linux
cp -r CrewPilot-dev/.github/skills/ ~/.copilot/skills/Contributions welcome! Please read the skill files in .github/skills/ to understand the structure. Each skill is a self-contained SKILL.md with YAML frontmatter.
欢迎贡献！请阅读 .github/skills/ 中的技能文件了解结构。每个技能都是一个独立的 SKILL.md，带有 YAML 前置元数据。
MIT — see LICENSE.
| Project | Inspiration |
|---|---|
| garrytan/gstack | Core methodology: office-hours, review workflow, security audit |
| CrewAI | Multi-agent adversarial review pattern |
| Aider | Atomic commits, full git integration |
| CodeRabbit | Preference learning, adaptive review filtering |
| MetaGPT | Multi-role design debate |