动手实验室 — CCA-F Claude Certified Architect — Foundations

概述

先决条件

步骤

1. 1.安装 Claude Code 并确认它能连接到 API

   配置服务：

   • Claude Code
   • Anthropic API

   Claude Code 是 CCA-F 考试最直接测试的 CLI 工具——配置、钩子、斜杠命令和权限模式都集中于此。我们首先将其全局安装，以便后续所有步骤都可以调用 claude。

   安装后，快速运行 claude --version 可以证明二进制文件位于您的 PATH 中，而 claude --print-system-prompt 则确认 API 密钥有效（它在底层进行了一次轻量级调用）。

   复制

   # Install Claude Code globally
   npm install -g @anthropic-ai/claude-code
   
   # Verify installation
   claude --version
   
   # Confirm API connectivity (prints the built-in system prompt)
   claude --print-system-prompt | head -20

2. 2.设置 `CLAUDE.md` 项目级和用户级文件

   配置服务：

   • CLAUDE.md
   • Project Config

   考试会重点测试三层 CLAUDE.md 结构：用户级 (~/.claude/CLAUDE.md)、项目级（仓库根目录）和模块级（子目录）。每个级别都继承自其上级，并且可以覆盖上级的设置。

   我们将创建一个项目级文件来设置编码规范，并创建一个用户级文件来设置个人偏好。当 Claude Code 读取项目时，它会合并所有这三层——这正是考试中会问到的行为。

   复制

   # Create project root CLAUDE.md
   cat > CLAUDE.md << 'EOF'
   # Project: CCA-F Lab
   
   ## Coding conventions
   - Use TypeScript for all new files
   - Prefer const over let
   - Use single quotes for strings
   - No semicolons (Prettier default)
   
   ## Testing
   - Run `npm test` before committing
   - All new functions must have at least one test
   EOF
   
   # Create user-level CLAUDE.md (applies to ALL your projects)
   mkdir -p ~/.claude
   cat > ~/.claude/CLAUDE.md << 'EOF'
   # User preferences
   - Be concise in responses
   - Prefer functional programming patterns
   - Always explain the "why" before showing code
   EOF
   
   # Verify Claude Code sees both files
   claude "What instructions do you see in CLAUDE.md?" --print

3. 3.构建一个自定义斜杠命令来搭建新模块

   配置服务：

   • Slash Commands
   • Claude Code

   自定义斜杠命令以 Markdown 文件的形式存储在 .claude/commands/ 中。文件名成为命令名称（例如 scaffold.md → /scaffold）。文件内容是提示模板——它可以包含 $ARGUMENTS 用于用户输入。

   这是一个高频考点：命令的存放位置、参数的传递方式以及它们与钩子 (hooks) 的区别。

   复制

   # Create the commands directory
   mkdir -p .claude/commands
   
   # Create a scaffold command
   cat > .claude/commands/scaffold.md << 'EOF'
   Create a new TypeScript module named $ARGUMENTS with:
   1. A main source file at src/$ARGUMENTS/index.ts
   2. A test file at src/$ARGUMENTS/__tests__/index.test.ts
   3. Export the module from the project root index.ts
   
   Follow the coding conventions in CLAUDE.md.
   EOF
   
   # Test the command (Claude Code will execute it interactively)
   claude /scaffold calculator

4. 4.配置钩子以在每次文件写入前自动 Linter

   配置服务：

   • Hooks
   • Settings

   钩子会响应 Claude Code 事件运行 shell 命令。考试会测试三种钩子类型：PreToolUse（工具运行前）、PostToolUse（工具运行后）和 Notification（状态变更时）。钩子在 .claude/settings.json 中定义。

   我们将添加一个 PostToolUse 钩子，它在每次文件编辑后运行 Linter —— 这是一个常见的生产模式，也是考试经常会问到的。

   复制

   # Initialize a package.json so we have a lint script
   npm init -y
   npm install --save-dev typescript @typescript-eslint/parser
   
   # Create project-level settings with a post-tool hook
   mkdir -p .claude
   cat > .claude/settings.json << 'EOF'
   {
     "hooks": {
       "PostToolUse": [
         {
           "matcher": "Write|Edit",
           "command": "npx tsc --noEmit 2>&1 | head -5"
         }
       ]
     },
     "permissions": {
       "allow": [
         "Read",
         "Write",
         "Edit"
       ]
     }
   }
   EOF
   
   # Verify settings are recognized
   cat .claude/settings.json

5. 5.启动一个本地 MCP 服务器并连接 Claude Code 到它

   配置服务：

   • MCP
   • stdio Transport

   模型上下文协议 (MCP) 是考试的核心领域。MCP 服务器通过标准化的 JSON-RPC 协议向 LLM 客户端公开工具、资源和提示。stdio 传输是最简单的——服务器从 stdin 读取并写入 stdout。

   我们将构建一个最简 MCP 服务器，它公开一个 get_weather 工具，然后配置 Claude Code 连接到它。这将实践考试测试的完整 MCP 生命周期：工具定义、传输连接、工具调用和结果处理。

   复制

   # Install the MCP SDK
   npm install @modelcontextprotocol/sdk
   
   # Create the MCP server
   cat > mcp-weather.mjs << 'EOF'
   import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
   import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
   
   const server = new McpServer({ name: "weather", version: "1.0.0" });
   
   server.tool("get_weather", { city: { type: "string" } }, async ({ city }) => ({
     content: [{ type: "text", text: `Weather in ${city}: 72°F, sunny` }],
   }));
   
   const transport = new StdioServerTransport();
   await server.connect(transport);
   EOF
   
   # Register the MCP server in Claude Code settings
   cat > .claude/settings.json << 'EOF'
   {
     "mcpServers": {
       "weather": {
         "command": "node",
         "args": ["mcp-weather.mjs"]
       }
     }
   }
   EOF
   
   # Ask Claude Code to use the MCP tool
   claude "What is the weather in Tokyo? Use the get_weather tool."

6. 6.使用工具定义和结构化输出调用 Anthropic API

   配置服务：

   • Tool Use API
   • Structured Output

   考试会详细测试 Messages API 的 tool_use 流程：您在请求中发送 tools，模型以 tool_use 内容块响应，您执行它并返回 tool_result，然后模型生成最终答案。

   此步骤使用 curl 直接调用 API —— 这是最符合考试要求的做法，因为考试测试的是原始请求/响应格式，而非 SDK 封装。

   复制

   # Call the Messages API with a tool definition
   curl -s https://api.anthropic.com/v1/messages \
     -H "x-api-key: $ANTHROPIC_API_KEY" \
     -H "anthropic-version: 2023-06-01" \
     -H "content-type: application/json" \
     -d '{
       "model": "claude-sonnet-4-6",
       "max_tokens": 1024,
       "tools": [{
         "name": "calculate",
         "description": "Evaluate a math expression",
         "input_schema": {
           "type": "object",
           "properties": {
             "expression": { "type": "string", "description": "Math expression to evaluate" }
           },
           "required": ["expression"]
         }
       }],
       "messages": [{
         "role": "user",
         "content": "What is 15% tip on a $85.50 dinner bill?"
       }]
     }' | python3 -m json.tool

7. 7.启用扩展思考以进行复杂的多步骤推理

   配置服务：

   • Extended Thinking
   • Anthropic API

   扩展思考允许模型在响应前使用 thinking 块进行逐步推理。考试会测试何时使用它（复杂推理、数学、代码生成）以及如何配置 budget_tokens。

   我们将发送一个刻意设计的复杂逻辑谜题，并比较启用和不启用扩展思考时的响应，以便亲身体验质量差异。

   复制

   # Request WITH extended thinking
   curl -s https://api.anthropic.com/v1/messages \
     -H "x-api-key: $ANTHROPIC_API_KEY" \
     -H "anthropic-version: 2023-06-01" \
     -H "content-type: application/json" \
     -d '{
       "model": "claude-sonnet-4-6",
       "max_tokens": 16000,
       "thinking": {
         "type": "enabled",
         "budget_tokens": 10000
       },
       "messages": [{
         "role": "user",
         "content": "A farmer has 17 sheep. All but 9 run away. How many does he have left? Explain your reasoning step by step."
       }]
     }' | python3 -m json.tool
   
   # The response includes a "thinking" block showing the model's
   # internal reasoning before the final answer.

8. 8.使用提示缓存来降低重复系统提示的成本

   配置服务：

   • Prompt Caching
   • Anthropic API

   提示缓存允许您将请求的一部分标记为可缓存。缓存的前缀在后续调用中成本降低 90%，并且没有额外的延迟。考试会测试 cache_control 块的放置位置以及缓存命中时的 HTTP 头。

   我们将两次发送相同的长系统提示，并检查响应头以确认第二次调用命中了缓存。

   复制

   # First call — creates the cache entry
   curl -s -D /dev/stderr https://api.anthropic.com/v1/messages \
     -H "x-api-key: $ANTHROPIC_API_KEY" \
     -H "anthropic-version: 2023-06-01" \
     -H "content-type: application/json" \
     -d '{
       "model": "claude-sonnet-4-6",
       "max_tokens": 256,
       "system": [{
         "type": "text",
         "text": "You are a helpful assistant specialized in cloud architecture. You follow AWS Well-Architected Framework principles. You always recommend infrastructure as code. You prefer Terraform over CloudFormation. You prioritize security and cost optimization.",
         "cache_control": { "type": "ephemeral" }
       }],
       "messages": [{ "role": "user", "content": "How should I set up VPC peering?" }]
     }' > /dev/null 2>&1
   
   # Second call — should hit the cache (check usage.cache_read_input_tokens)
   curl -s https://api.anthropic.com/v1/messages \
     -H "x-api-key: $ANTHROPIC_API_KEY" \
     -H "anthropic-version: 2023-06-01" \
     -H "content-type: application/json" \
     -d '{
       "model": "claude-sonnet-4-6",
       "max_tokens": 256,
       "system": [{
         "type": "text",
         "text": "You are a helpful assistant specialized in cloud architecture. You follow AWS Well-Architected Framework principles. You always recommend infrastructure as code. You prefer Terraform over CloudFormation. You prioritize security and cost optimization.",
         "cache_control": { "type": "ephemeral" }
       }],
       "messages": [{ "role": "user", "content": "What about NAT gateway sizing?" }]
     }' | python3 -c "import sys,json; d=json.load(sys.stdin); print(f\"Cache read tokens: {d['usage'].get('cache_read_input_tokens', 0)}\")"

9. 9.在无头模式下运行 Claude Code 以用于 CI/CD 流水线

   配置服务：

   • Headless Mode
   • CI/CD

   考试会测试如何非交互式地运行 Claude Code —— --print 标志会将结果输出到 stdout 而不启动交互式会话，而 --output-format json 则提供结构化输出，便于脚本处理。

   这是 CI 流水线中使用的模式：输入提示，获取结果，然后以编程方式解析它。

   复制

   # Headless mode: prompt in, text out
   claude --print "List 3 TypeScript best practices, one per line"
   
   # JSON output for programmatic use
   claude --print --output-format json "What is 2+2?"
   
   # Pipe a file for code review (CI pattern)
   echo 'const x: any = "hello";' > example.ts
   claude --print "Review this TypeScript file for type safety issues: $(cat example.ts)"

10. 10.构建一个多智能体主管，将任务委托给专业子智能体

    配置服务：

    • Multi-Agent
    • Supervisor Pattern
    • Anthropic SDK

    领域 1（智能体架构，占 27%）是考试中权重最高的。主管模式——一个编排器调度到专业子智能体，每个子智能体都有自己的系统提示和工具集——是考试中最常测试的架构。

    我们将构建一个独立的 Node.js 脚本，实现一个主管来协调两个专家：一个研究员负责读取文件并总结，一个作者负责生成输出。主管会分解任务，通过单独的 API 调用委托给每个专家，并汇总结果。从终端运行它，并实时观察任务的委托过程。

    复制

    # Install the Anthropic SDK
    npm install @anthropic-ai/sdk
    
    # Create sample data for the agents to work with
    cat > sales-data.csv << 'EOF'
    month,revenue,units
    Jan,12500,150
    Feb,15200,180
    Mar,11800,140
    Apr,18900,220
    May,22100,260
    Jun,19500,230
    EOF
    
    # Create the multi-agent supervisor script
    cat > supervisor.mjs << 'SCRIPT'
    import Anthropic from "@anthropic-ai/sdk";
    import { readFileSync } from "fs";
    
    const client = new Anthropic();
    
    async function callAgent(role, systemPrompt, userMessage) {
      console.log(`\n--- ${role} agent ---`);
      const response = await client.messages.create({
        model: "claude-sonnet-4-6",
        max_tokens: 1024,
        system: systemPrompt,
        messages: [{ role: "user", content: userMessage }],
      });
      const text = response.content
        .filter((b) => b.type === "text")
        .map((b) => b.text)
        .join("\n");
      console.log(text);
      return text;
    }
    
    // --- Supervisor: decompose, delegate, aggregate ---
    const data = readFileSync("sales-data.csv", "utf-8");
    
    console.log("=== SUPERVISOR: Decomposing task ===");
    console.log("Task: Analyze sales data and write a brief executive summary.\n");
    console.log("Step 1: Delegate analysis to Researcher agent");
    console.log("Step 2: Delegate writing to Writer agent");
    console.log("Step 3: Combine results\n");
    
    // Sub-agent 1: Researcher — extracts insights
    const analysis = await callAgent(
      "Researcher",
      "You are a data analyst. Extract key trends, highs, lows, and growth rates. Be precise with numbers. Output bullet points only.",
      `Analyze this CSV data:\n\n${data}`
    );
    
    // Sub-agent 2: Writer — produces the summary from the analysis
    const summary = await callAgent(
      "Writer",
      "You are a business writer. Write a 3-sentence executive summary from the analysis provided. Use a professional tone. Include specific numbers.",
      `Write an executive summary based on these findings:\n\n${analysis}`
    );
    
    console.log("\n=== SUPERVISOR: Final aggregated output ===");
    console.log(summary);
    SCRIPT
    
    # Run the supervisor
    node supervisor.mjs

清理

# Remove the lab directory
cd .. && rm -rf cca-f-lab

# Optionally remove the user-level CLAUDE.md we created
rm -f ~/.claude/CLAUDE.md

# Optionally uninstall Claude Code
npm uninstall -g @anthropic-ai/claude-code