The Agentic Startup：给 Claude Code 装一套创业团队工作流

by Mycelium Protocol

---

Claude Code 很强，但没有结构的时候很容易跑偏：直接开始写代码，需求没讲清楚，多轮对话之后上下文丢失，实现和最初想法越来越远。

The Agentic Startup 的回答是：先写规格，再写代码（spec-driven development）。它把 Claude Code 改造成一个「创业团队」——需求分析师、架构师、工程师、QA、文档工程师各司其职，按阶段流转，有质量门控。

GitHub: https://github.com/rsmdt/the-startup | ⭐ 367 | MIT License | Shell

---

10 个命令，3 个阶段

SETUP（可选）
  /constitution ──► 建立项目治理规则（在 BUILD 全程自动执行）

BUILD（主流程）
  /specify     ──► 生成规格（需求文档 + 解决方案设计）
  /validate    ──► 校验质量（3C 框架）
  /implement   ──► 按复杂度分级执行
  /test        ──► 运行测试，强制所有权
  /review      ──► 多智能体代码评审
  /document    ──► 生成/同步文档

MAINTAIN（按需）
  /analyze     ──► 发现代码模式和规律
  /refactor    ──► 安全重构（保留行为）
  /debug       ──► 根因分析式调试

---

三级复杂度自动分发

/specify 执行完会自动分类复杂度，/implement 按分类选择执行策略：

级别	场景	执行方式
Direct	修复/重构/单一验收标准功能	直接读 requirements + solution，无分解产物
Incremental	单一功能，1-2 个组件	生成 plan/ 目录，按 phase-N.md 分阶段循环
Factory	多功能/并行工作	生成 manifest.md + units/*.md，并行执行原子单元

规格存储结构：

.start/specs/001-feature-name/
├── requirements.md     # 构建什么，为什么
├── solution.md         # 技术上怎么构建
├── plan/               # (Incremental) README.md + phase-N.md
└── manifest.md + units/ + scenarios/  # (Factory)

---

关键机制

跨会话恢复：规格写到磁盘。上下文用完了重开对话，/specify 001 或 /implement 001 接着上次的继续——Claude 读规格文件恢复状态。

漂移检测：/implement 执行过程中自动对比实现和规格，发现不一致时提示：更新规格还是改代码。

代码所有权强制：/test 发现失败测试，不接受「这是预存的」借口——碰了代码库就要修。

自适应代码评审：/review 根据改动内容自动追加专项评审视角——有 async 代码→并发评审，有依赖变更→供应链检查，有 UI 变更→可访问性审计。5 个基础视角 + 条件专项。

---

安装

# 一行安装（推荐）
curl -fsSL https://raw.githubusercontent.com/rsmdt/the-startup/main/install.sh | sh

# 或通过 Marketplace 手动安装
/plugin marketplace add rsmdt/the-startup
/plugin install start@the-startup      # 核心工作流（必装）
/plugin install team@the-startup       # 专项智能体（可选）

---

两个插件

start@the-startup（核心）：10 个用户命令 + 5 个自主技能 + 2 种输出风格

team@the-startup（可选）：8 个角色，20 个活动专项智能体

角色	负责领域
Chief	复杂度评估、活动路由、并行执行
Analyst	需求分析、优先级、项目协调
Architect	系统设计、技术研究、质量评审、文档
Software Engineer	API、组件、领域建模、性能
QA Engineer	测试策略、探索性测试、负载测试
Designer	用研、交互设计、设计系统、无障碍
Platform Engineer	IaC、容器、CI/CD、监控、数据管道
Meta Agent	智能体设计与生成

---

v3 新增：Agent Teams（实验性）

专项智能体之间可以自主协作，共同处理复杂任务：

// ~/.claude/settings.json
{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

安装脚本会自动询问是否配置。

---

两种输出风格

The Startup：高能量执行风格，YC 路演氛围，「现在就交付」 The ScaleUp：冷静专业风格，教育深度，「可持续的速度」

任何时候切换：/output-style start:The Startup 或 /output-style start:The ScaleUp

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

The Agentic Startup: A Full Startup-Team Workflow for Claude Code

by Mycelium Protocol

---

Claude Code is powerful, but without structure it drifts easily: you jump straight to code before requirements are clear, context gets lost across turns, and what you build keeps diverging from what you intended.

The Agentic Startup’s answer: spec before code. It turns Claude Code into a startup team — analyst, architect, engineer, QA, and doc writer each playing their role, flowing through stages with quality gates at every step.

GitHub: https://github.com/rsmdt/the-startup | ⭐ 367 | MIT License | Shell

---

10 Commands, 3 Phases

SETUP (optional)
  /constitution ──► Project governance rules (auto-enforced throughout BUILD)

BUILD (primary flow)
  /specify     ──► Generate spec (requirements + solution design)
  /validate    ──► Quality check (3C framework)
  /implement   ──► Auto-dispatched by complexity tier
  /test        ──► Run tests, enforce code ownership
  /review      ──► Multi-agent code review
  /document    ──► Generate / sync documentation

MAINTAIN (as needed)
  /analyze     ──► Discover patterns and rules
  /refactor    ──► Safe refactor (preserve behavior)
  /debug       ──► Root-cause analysis debugging

---

Three-Tier Complexity Dispatch

/specify classifies complexity automatically; /implement routes to the matching strategy:

Tier	When	Execution
Direct	Fixes, refactors, single-AC features	Reads requirements + solution directly, no decomposition artifacts
Incremental	Single feature, 1–2 components	Generates plan/ with phase-N.md files, executed loop per phase
Factory	Multi-feature, parallel work	Generates manifest.md + units/*.md, parallel atomic unit execution

Specs persist on disk in .start/specs/001-feature-name/ — this is what makes resumption work.

---

Key Mechanisms

Resume across sessions: Specs live on disk. When context runs out, start a new session and /implement 001 picks up exactly where you left off.

Drift detection: During /implement, the framework auto-compares implementation against spec. When they diverge, you choose: update the spec or update the code.

Code ownership mandate: /test finds a failing test — “pre-existing failure” is not an acceptable response. You touched the codebase, you own it.

Adaptive code review: /review reads what changed and adds specialist perspectives automatically. Async code triggers concurrency review. Dependency changes trigger supply-chain checks. UI changes trigger accessibility audits.

---

Install

# One-line install (recommended)
curl -fsSL https://raw.githubusercontent.com/rsmdt/the-startup/main/install.sh | sh

# Or via Marketplace
/plugin marketplace add rsmdt/the-startup
/plugin install start@the-startup      # core workflow (required)
/plugin install team@the-startup       # specialist agents (optional)

---

Two Plugins

start@the-startup (core): 10 user skills + 5 autonomous skills + 2 output styles

team@the-startup (optional): 8 roles, 20 activity-based agents covering Chief, Analyst, Architect, Software Engineer, QA Engineer, Designer, Platform Engineer, and Meta Agent.

---

v3: Agent Teams (Experimental)

Specialist agents can now coordinate autonomously on complex tasks. Enable via:

{ "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" } }

The installer configures this automatically if you opt in.

---

Two Output Styles

The Startup: High-energy, Y Combinator intensity, “let’s ship this NOW” The ScaleUp: Calm confidence, engineering excellence, “sustainable speed at scale”

Switch anytime: /output-style start:The Startup or /output-style start:The ScaleUp

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.