Claude Certified Architect — Foundations (CCA-F) Study Guide

Domain 1: Agentic Architecture & Orchestration (27%)

This is the heaviest domain on the exam. Anthropic wants to ensure architects understand how to build systems where Claude interacts with environment hooks, loops, and other models.

Key Architecture Patterns

• Single Agentic Loop: An agent that receives a prompt, uses a tool, observes the outcome, and repeats until the task is complete. Be prepared to choose the right stop_reason (e.g., tool_use, stop_sequence, or end_turn) to control the flow.
• Hub-and-Spoke Pattern: A supervisor model orchestrating multiple specialized subagents. You must know when to delegate a task to a subagent and how to aggregate their results to avoid bloated context windows.
• Sequential Processing vs. Parallel Execution: Choosing whether tasks must run step-by-step or can run concurrently. Candidates must resolve latency vs. cost tradeoffs here.

Domain 2: Claude Code Configuration & Workflows (20%)

Claude Code is Anthropic's agentic CLI tool. This domain assesses your ability to configure developer projects so that agents operate efficiently and safely.

Understanding CLAUDE.md

• Hierarchy: How user-level config (e.g., ~/.claude.json) relates to workspace-specific CLAUDE.md instructions.
• Sections: Standard parts of a CLAUDE.md file, including **Build Commands**, **Formatting Guidelines**, and **Project Rules**.

# CLAUDE.md Example Structure
## Build & Test Commands
- Run tests: `npm test`
- Build app: `npm run build`

## Code Style Guidelines
- Use TypeScript for all modern utilities
- Always handle API exceptions with try/catch blocks
- Prefix helper functions with a descriptive comment

You will also face scenarios about **Plan Mode vs. Direct Execution**. You must determine when it is appropriate for Claude Code to write an implementation plan to get user approval before writing code, and how to set up CI/CD workflows so that agent scripts do not hang waiting for keyboard inputs.

Domain 3: Prompt Engineering & Structured Output (20%)

While this covers prompting, it is tailored for system designers. You must know how to guarantee that Claude returns structured, deterministic responses suitable for programmatic parsing.

XML Tags & Context Placement

Anthropic models are pre-trained to respond well to XML tags (e.g., <rules>, <context>, <examples>). You will be quizzed on where to place context. For instance, the **system prompt** should contain instructions and schemas, whereas the **user message** should contain the immediate task data.

JSON Schema Enforcement

1. Describing the JSON schema in the tool definitions or system prompt.
2. Instructing Claude to skip conversational preambles and output only raw JSON.
3. Implementing parsing validation on the client side with a retry loop that feeds the error trace back to Claude.

Domain 4: Tool Design & MCP Integration (18%)

The Model Context Protocol (MCP) is an open-source standard for connecting LLMs to data and tools. The exam tests your practical understanding of MCP architecture.

Here is a basic JSON representation of an MCP tool configuration schema that you might see on the exam:

{
  "name": "fetch_incident_log",
  "description": "Retrieves incident details from the database. Use this tool only when you have a specific incident ID.",
  "input_schema": {
    "type": "object",
    "properties": {
      "incident_id": {
        "type": "string",
        "description": "The unique ID of the incident to fetch (e.g., INC-404)"
      }
    },
    "required": ["incident_id"]
  }
}

Domain 5: Context Management & Reliability (15%)

Designing production-grade AI systems requires managing model errors, latency, and tokens.

Optimizing for Long Context & Caching

• Prompt Caching: How to mark static blocks of prompt data (like system guidelines or large documents) with cache_control tags so that subsequent calls are faster and cost up to 90% less.
• Error Propagation & Escalation: Setting up deterministic thresholds. If Claude fails a tool validation 3 times, the system must trigger an escalation hook for human-in-the-loop (HITL) intervention.