agent-skill-tdd MCP — 使用场景, 安装 & 实时演示

为什么要用

核心特性

6-step workflow: Serena -> Investigate -> Clarify -> Document -> Design Review -> TDD
Timestamped .requirements/ docs per task
Explicit requirements confirmation before code
Architecture review via Lad MCP before implementation
Per-change peer review on every edit
Agent-agnostic — Claude Code, Codex, Cursor, Windsurf, Antigravity

实时演示

实际使用效果

agent-skill-tdd-skill.replay ▶ 就绪

0/0

安装

选择你的客户端

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "agent-skill-tdd-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Shelpuk-AI-Technology-Consulting/agent-skill-tdd",
        "~/.claude/skills/agent-skill-tdd"
      ],
      "_inferred": true
    }
  }
}

打开 Claude Desktop → Settings → Developer → Edit Config。保存后重启应用。

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "agent-skill-tdd-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Shelpuk-AI-Technology-Consulting/agent-skill-tdd",
        "~/.claude/skills/agent-skill-tdd"
      ],
      "_inferred": true
    }
  }
}

Cursor 使用与 Claude Desktop 相同的 mcpServers 格式。项目级配置优先于全局。

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "agent-skill-tdd-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Shelpuk-AI-Technology-Consulting/agent-skill-tdd",
        "~/.claude/skills/agent-skill-tdd"
      ],
      "_inferred": true
    }
  }
}

点击 Cline 侧栏中的 MCP Servers 图标，然后选 "Edit Configuration"。

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "agent-skill-tdd-skill": {
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Shelpuk-AI-Technology-Consulting/agent-skill-tdd",
        "~/.claude/skills/agent-skill-tdd"
      ],
      "_inferred": true
    }
  }
}

格式与 Claude Desktop 相同。重启 Windsurf 生效。

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "agent-skill-tdd-skill",
      "command": "git",
      "args": [
        "clone",
        "https://github.com/Shelpuk-AI-Technology-Consulting/agent-skill-tdd",
        "~/.claude/skills/agent-skill-tdd"
      ]
    }
  ]
}

Continue 使用服务器对象数组，而非映射。

~/.config/zed/settings.json

{
  "context_servers": {
    "agent-skill-tdd-skill": {
      "command": {
        "path": "git",
        "args": [
          "clone",
          "https://github.com/Shelpuk-AI-Technology-Consulting/agent-skill-tdd",
          "~/.claude/skills/agent-skill-tdd"
        ]
      }
    }
  }
}

加入 context_servers。Zed 保存后热重载。

claude mcp add agent-skill-tdd-skill -- git clone https://github.com/Shelpuk-AI-Technology-Consulting/agent-skill-tdd ~/.claude/skills/agent-skill-tdd

一行命令搞定。用 claude mcp list 验证，claude mcp remove 卸载。

使用场景

实战用法： agent-skill-tdd

Implement a new feature with full TDD discipline

👤 Teams where AI coders jump to implementation and miss requirements ⏱ ~120 min intermediate

何时使用： Non-trivial feature where 'just write it' has burned you before.

前置条件

Skill cloned — git clone https://github.com/Shelpuk-AI-Technology-Consulting/agent-skill-tdd ~/.claude/skills/agent-skill-tdd
Serena available — Install Serena MCP server
Lad MCP for design review — Install Lad MCP server

步骤

Kick off with the skill

Use agent-skill-tdd. I want to add a 'saved searches' feature. Start from step 1 — activate Serena and investigate.✓ 已复制

→ Claude reads code, summarizes current state
Clarify requirements

Confirm what I just said maps to these acceptance criteria — don't code yet.✓ 已复制

→ Back-and-forth until confirmed; .requirements/ file saved
Design review via Lad

Run design review — architecture fit, hotspots.✓ 已复制

→ Design critique, not just approval
TDD loop

Now begin TDD — write a failing test, minimum code to pass, refactor. Peer review each step.✓ 已复制

→ Red-green-refactor cycles, each reviewed

结果： Feature shipped with tests, requirements history, and design rationale preserved.

注意事项

Skipping clarify to 'save time' — That's where most AI coding value is lost — hold the line

搭配使用： filesystem · github

Fix a bug after writing down what correct behavior should be

👤 Devs tired of AI 'fixes' that break other paths ⏱ ~60 min intermediate

何时使用： A subtle bug where the reproduction is half the battle.

步骤

Investigate + document

Use agent-skill-tdd bug mode. Symptom: <description>. Write the reproduction + expected behavior to .requirements/ first.✓ 已复制

→ Timestamped doc with As-Is and To-Be
Write failing test

Test that expresses the expected behavior — currently failing.✓ 已复制

→ Red test
Fix + peer review

Minimum fix to go green. Peer review.✓ 已复制

→ Green + review notes

结果： A bug fix with a test preventing regression.

Stop before implementing — validate you understand the spec

👤 Teams where spec-misreading causes rework ⏱ ~30 min beginner

何时使用： The ticket is ambiguous and you suspect you're about to build the wrong thing.

步骤

Only run steps 1–4

Use agent-skill-tdd — investigate + clarify + document + design review. Stop before coding.✓ 已复制

→ Requirements doc + design review without touching code

结果： You either discover you misread the ticket or greenlight with confidence.

组合

与其他 MCP 搭配，撬动十倍杠杆

agent-skill-tdd-skill + filesystem

Keep .requirements/ in the repo for future agents to read

Commit .requirements/ to git so the next session has context.✓ 已复制

agent-skill-tdd-skill + github

Link requirements doc in the PR description

When opening the PR, paste the .requirements/ file path in the description.✓ 已复制

工具

此 MCP 暴露的能力

工具	输入参数	何时调用	成本
activate_serena	repo path	Step 1	0 — local
investigate	task description	Step 2	0
clarify	task, user answers	Step 3	0
document_requirements	confirmed requirements	Step 4	0
design_review	requirements + repo	Step 5	0
tdd_cycle	requirements	Step 6	0

成本与限制

运行它的成本

API 配额: None from the skill; Serena and Lad have their own footprints
每次调用 Token 数: Moderate — each step is a round trip
费用: Free
提示: Full 6-step is overkill for trivial tasks — run steps 1–4 only on small work

安全

权限、密钥、影响范围

凭据存储： None

数据出站： Only via Serena/Lad if configured

故障排查

常见错误与修复

Serena not available

Install the Serena MCP server and register it with your agent

Requirements doc not created

Ensure filesystem write access to .requirements/

验证： ls .requirements/

Design review step feels rubber-stamp

Force an adversarial framing: 'what's the strongest objection to this design?'

替代方案

agent-skill-tdd 对比其他方案

替代方案	何时用它替代	权衡
claude-gpt-workflow	You want cross-model review instead of self-TDD discipline	Higher API cost
planning-with-files-skill	You want persistence across sessions but less TDD rigor	No test-first enforcement

agent-skill-tdd