1. EXAM OVERVIEW
Released: March 12, 2026
Format: 60 scenario-based multiple-choice questions
Duration: 120 minutes (2 hours)
Pass Score: 720 out of 1000
Cost: $99 USD (free for first 5,000 CPN employees)
Register: https://anthropic.skilljar.com
What This Exam Tests
The CCA-F is an architecture-level certification that validates your ability to:
- Design production-grade Claude systems
- Implement agentic loops correctly
- Manage tools and MCP integration
- Engineer prompts for specific outcomes
- Handle context and reliability at scale
Who Should Take This
- Backend engineers building Claude systems
- Solution architects designing agent workflows
- DevOps engineers deploying Claude at scale
- Technical leads evaluating Claude for production
- Anyone with 6+ months Claude API experience
2. OFFICIAL EXAM BLUEPRINT
| Domain | Weight | Key Focus Areas |
|---|---|---|
| D1: Agentic Architecture | 27% | Agentic loops, stop_reason, multi-agent patterns, orchestration |
| D2: Tool Design & MCP | 18% | Tool definitions, MCP primitives, error handling |
| D3: Claude Code | 20% | CLAUDE.md, skills, commands, hooks, subagents |
| D4: Prompt Engineering | 20% | System prompts, JSON output, few-shot examples |
| D5: Context Management | 15% | Token budgeting, prompt caching, reliability |
3. 4-WEEK STUDY ROADMAP
Weekly Commitment: 25–30 hours
Best Timeline: 4 consecutive weeks
Week 1: Foundation (20–25 hours)
Focus: D1 basics, D4 fundamentals
Week 2: Core Domains Deep Dive (30–35 hours)
Focus: D1 (heavy), D4, D2
Week 3: Advanced Domains (30–35 hours)
Focus: D3 (Claude Code), D5 (Context)
Week 4: Practice & Exam (20–25 hours)
Focus: All domains, intensive practice
4. OFFICIAL RESOURCES
Free Resources (Mandatory)
| Resource | Time | Why It Matters |
|---|---|---|
| Anthropic Academy | ~20 hours | Official courses, covers all domains |
| Official Exam Guide PDF | 4–5 hours | Sample scenarios, scoring rubric |
| Claude API Docs | As needed | Reference material |
| MCP Specification | 2–3 hours | Required for D2 |
| Official Practice Exam | 2 hours | Mirrors real exam format |
Paid Resources (Recommended)
| Resource | Cost | Questions | Value |
|---|---|---|---|
| TutorialsDojo | $49 | 500+ | Best for exam prep |
| CertSafari | $49 | 600+ | Good for domain drills |
| Whizlabs | $39 | 400+ | Scenario-based |
To Ace It: Add paid question banks = 95%+ pass rate
5. DOMAIN BREAKDOWN & KEY CONCEPTS
Domain 1: Agentic Architecture & Orchestration (27%)
The Agentic Loop (Most Important Concept)
User Request
|
Claude Agent (with system prompt + tools)
|
Check stop_reason:
- "end_turn" -> respond to user, done
- "tool_use" -> extract tool calls, execute
- "max_tokens" -> context overflow, truncate/summarize
|
If tool_use: Execute tool, get result
|
Pass result back to Claude as new message
|
Repeat until stop_reason = "end_turn"
Stop_reason Values (Memorize These)
- end_turn: Claude finished responding, ready to return to user
- tool_use: Claude called a tool, waiting for result
- max_tokens: Claude hit token limit, response incomplete
- stop_sequence: Claude hit a stop sequence (rare)
Multi-Agent Patterns
- Coordinator-Subagent: One coordinator decides, multiple specialists execute
- Hub-and-Spoke: Central agent with many tools (simpler than multi-agent)
- Sequential Pipeline: Agents run one after another (research -> outline -> draft -> edit)
Domain 2: Tool Design & MCP Integration (18%)
Tool Definition Quality
The description is what Claude reads. A clear description > a clever function name.
GOOD: "Execute read-only SELECT queries. Use for: finding orders, checking account status. DO NOT use for: INSERT, UPDATE, DELETE. Always include WHERE clause." BAD: "Query database"
MCP Primitives
- Tools: Callable functions
- Resources: Data Claude can read (docs, config)
- Prompts: Reusable instructions
- Sampling: LLM-powered logic in MCP server
Domain 3: Claude Code Configuration & Workflows (20%)
CLAUDE.md Hierarchy
Project Root/
|-- CLAUDE.md (applies everywhere)
|-- subdir1/
| |-- CLAUDE.md (applies only to subdir1)
|-- subdir2/
|-- CLAUDE.md (applies only to subdir2)
Key Components
- Skills: Reusable markdown instructions (stored in .claude/skills/)
- Commands: Custom slash commands (/review, /test, etc.)
- Hooks: Lifecycle events (before_command, after_command, on_error)
- Subagents: Isolated agents with restricted tools
Exam Tip
Questions often ask: "How to configure CLAUDE.md for a team?" or "What's the scope of this rule?" Study path-scoping and frontmatter syntax.
Domain 4: Prompt Engineering & Structured Outputs (20%)
System Prompt Design
Be specific. Bad example: "You're a helpful assistant." Good example: "You are a customer support specialist with 10 years experience. Keep responses under 150 words. If billing is involved, escalate to senior agent."
Structured Output (JSON)
Constrain Claude to return valid JSON. Example schema:
{
"resolution": "string",
"confidence": "number (0-1)",
"escalate": "boolean",
"reason": "string"
}
Few-Shot Examples
Show Claude 2–5 examples of input -> desired output. More effective than lengthy instructions for consistent output.
Domain 5: Context Management & Reliability (15%)
Token Budgeting
Each message token costs money. Conversation history grows. At some point, you hit the token limit. Solution: prune old messages, summarize.
Prompt Caching
Cache system prompt + reference docs. Reuse across requests (~90% cheaper for cached tokens).
"Lost in the Middle" Problem
Claude pays less attention to middle of long prompts. Solution: put important info at start/end, verbose content in middle.
Error Recovery
- Retry: If transient failure (timeout), retry with backoff
- Fallback: If tool fails, use alternative or skip
- Escalation: If repeated failure, escalate to human
6. CRITICAL TERMINOLOGY (Flashcards)
7. HANDS-ON PROJECTS
Project 1: Simple Chatbot (Week 1, 2 weeks)
Domains: D1, D4
Tech: Flask/FastAPI + Claude API
Deliverable: Working chatbot with 3 conversations
Success Metrics: <2s latency, <$0.01/request, context doesn't grow unbounded
Project 2: Customer Support Agent (Weeks 3–5, 3 weeks)
Domains: D1, D2, D5
Tools: check_ticket_history, get_customer, search_knowledge_base, escalate
Deliverable: Production agent with monitoring
Success Metrics: 70%+ resolve without escalation, <5s latency, <$0.05/ticket
Project 3: RAG System (Weeks 6–8, 3 weeks)
Domains: D2, D5
Tech: Chroma/Pinecone + embeddings + Claude
Deliverable: Working RAG with evaluation
Success Metrics: Precision >80%, latency <3s, cost <$0.10/query
8. PRACTICE STRATEGY
Practice Exam Schedule
The authoritative blueprint: domains & weights, all task statements, the 6 scenarios, the 12 official sample questions with answers, and the in/out-of-scope topic lists. Read this first.
77 scenario questions across all five domains at the official weightings (D1 27% · D2 18% · D3 20% · D4 20% · D5 15%), single-best-answer, pass ≥ 720/1000. Instant scoring, full explanations, and options reshuffle each attempt. Use it as your Week-4 dress rehearsal.
60 questions · 120-minute countdown · no feedback until submit — mirrors the real exam. All-new questions (not Mock #1's 77), scaled-score estimate, and a per-domain breakdown so you can spot a weak domain. Your final dress rehearsal under real conditions.
77 scenario-focused questions · instant feedback. All-new items covering the complete 6-scenario bank: Customer Support Resolution Agent (Q1–12: D1/D2/D5), Code Generation with Claude Code (Q13–25: D3/D5), Multi-Agent Research System (Q26–38: D1/D2/D5), Developer Productivity with Claude (Q39–51: D2/D3/D1), Claude Code for Continuous Integration (Q52–64: D3/D4), and Structured Data Extraction (Q65–77: D4/D5). All five domains represented.
60 questions · 120 minutes · answer shown on click (optional exam mode) · topic tags hidden. Built because the real exam felt harder than Mocks #4 and #5. Every question has a runner-up: a good option that would be right if one fact in the question were different. Options are length-matched, so the longest one is rarely correct. The report shows how many of your misses were the runner-up.
60 questions · short, clear questions · close answers. Pitched at the real exam as candidates describe it: easy to read, hard to choose. Every question has a runner-up and full plain-language and think-it-through explanations. Start here, then use Mock #6 as a stretch.
Question Types You'll See
- Design Questions: "Design an agent for X task" (tools, escalation, success metrics)
- Constraint-Based: "Given budget/latency/compliance, choose architecture"
- Fix-The-Problem: "What's wrong with this design? How to fix?"
- Explain-Code: "What does this code do? What's missing?"
- Ranking: "Rank these options by trade-offs (Opus vs. Sonnet)"
9. EXAM DAY CHECKLIST
48 Hours Before
1 Hour Before Exam
During Exam
Strategy: Don't overthink. Most correct answer is obvious.
Flag: If stuck, flag for review, continue
Keywords to watch: "best practice" (robust), "cost-effective" (Sonnet), "least latency" (don't add features)
After Exam
Summary: Your Path to CCA-F
| Week | Hours | Focus | Deliverable |
|---|---|---|---|
| 1 | 20–25 | D1, D4 basics | Chatbot project |
| 2 | 30–35 | D1 (heavy), D4, D2 | Support agent project |
| 3 | 30–35 | D3, D5 | RAG system project |
| 4 | 20–25 | All domains review | Practice exams 850+ score |
| Total | 100–120 | – | CCA-F Pass |