How to Pass the Claude Certified Architect (CCA-F) Exam in 2026

If you build with Claude — through the API, Claude Code, or agent pipelines — the Claude Certified Architect Foundations (CCA-F) exam is the clearest signal you can put on a resume or LinkedIn profile to prove it. Launched by Anthropic in March 2026, this certification has quickly become the go-to credential for AI engineers, solutions architects, and developers who work in the Claude ecosystem.

This guide covers everything you need to know: exam format, the five domains, a realistic prep timeline, and the study strategies that separate people who pass on the first attempt from those who retake it.

What Is the Claude Certified Architect (CCA-F) Exam?

The CCA-F (Claude Certified Architect Foundations) is Anthropic's first professional certification. It validates that you can design, configure, and deploy production-ready Claude-based systems — not just prompt the chat interface.

Exam format at a glance:

The exam is deliberately scenario-based. You won't see trivia questions like "what is the maximum context window of Claude Sonnet?" Instead, you'll get a description of a real production system — a customer support bot, a CI/CD pipeline, a multi-agent research system — and be asked which architectural choice is correct.

This means you cannot memorize your way through it. You need to understand why certain patterns work.

The 5 Domains You'll Be Tested On

Anthropic publishes the domain weights in the official Exam Guide. Here they are, ordered by weight:

Domain 1: Agentic Architecture & Orchestration (27%)

The heaviest domain. It covers how to design multi-agent systems, when to use orchestrator vs. subagent patterns, how to handle failures gracefully, and how to manage long-running tasks.

Key topics:

• Orchestrator-agent hierarchies and delegation patterns
• Handling tool call failures and retry logic
• When to use parallel agents vs. sequential pipelines
• Context passing between agents without blowing the token budget
• Human-in-the-loop checkpoints and interruption handling

Study tip: Build something. Seriously. If you haven't written a multi-agent pipeline with Claude, do it before the exam. The questions in this domain are about real tradeoffs — latency vs. cost, autonomy vs. safety — and you need practical intuition to pick the right answer quickly.

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

This domain catches many candidates off guard with how specific it gets. It's not enough to know that CLAUDE.md exists — you need to know how Claude Code reads it, how the hierarchy works across root/subdirectory/project, and how skills and hooks are configured.

Key topics:

• CLAUDE.md file structure, hierarchy, and inheritance rules
• Writing and installing custom skills with proper frontmatter
• Hook types: pre-tool, post-tool, on-error, and when to use each
• Permission modes and what they allow/restrict
• CI/CD integration with --headless flag and automated pipelines
• Worktrees for parallel agent tasks

Study tip: Read through your own CLAUDE.md (or write one from scratch) and understand every configuration choice. Then read the official Claude Code docs on settings and permissions. Domain 2 rewards people who use Claude Code daily.

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

A well-weighted domain that tests your ability to write prompts that produce reliable, parseable output at production scale — not just responses that look good in a demo.

Key topics:

• Role/task/context framing patterns
• Few-shot examples and when they help vs. hurt
• Structured output: JSON mode, XML tags, and validation strategies
• System prompt design for multi-turn conversations
• Reducing hallucination with grounding techniques
• Prompt chaining vs. long single prompts (when to split)

Study tip: Practice generating structured JSON from Claude with and without schema constraints. Understand the failure modes: when does Claude drift from the schema, and how do you catch it? This domain also overlaps with context engineering concepts — the practice of shaping all the context around a prompt, not just the prompt itself.

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

The MCP (Model Context Protocol) section of the exam has tripped up candidates who know about MCP but haven't actually wired up a server. You need to understand tool definition schemas, server types (local stdio vs. remote SSE), and the security boundary between tools and the host system.

Key topics:

• Anatomy of an MCP tool definition (name, description, input schema)
• Writing descriptions that Claude actually follows (quality of description matters)
• Local vs. remote MCP servers and when to choose each
• Tool authorization and principle of least privilege
• Composing multiple MCP servers in a single workflow
• Common anti-patterns: tools with side effects not declared, ambiguous parameter names

Study tip: Build one MCP server from scratch. Use the TypeScript or Python SDK. Run it against Claude Code and see how Claude interprets your tool descriptions. If you've read our guide on building MCP servers, you're already ahead.

Domain 5: Context Management & Reliability (15%)

The lightest domain by weight, but the questions are subtle. It's about operating at scale — keeping costs predictable, outputs consistent, and context meaningful across long conversations.

Key topics:

• Token budget management and context window strategies
• Prompt caching: what gets cached, cache TTL, and cost impact
• Conversation summarization patterns for long-running sessions
• System-level reliability: retries, exponential backoff, fallbacks
• Evaluating Claude outputs programmatically (evals)
• A/B testing prompts in production

Study tip: Read Anthropic's documentation on prompt caching carefully. The exam will test you on what types of content benefit from caching and where caching breaks (dynamic content, high-churn contexts). Also understand how to design lightweight eval frameworks.

What the Exam Scenarios Look Like

The CCA-F uses six recurring production scenarios as the backdrop for questions. On any given exam, four of the six appear. Study all six so you're not caught off guard:

Customer Support Resolution Agent — a ticket triage + resolution pipeline

Code Generation with Claude Code — an autonomous PR review and generation workflow

Multi-Agent Research System — orchestrator + specialized subagents searching and synthesizing

Developer Productivity with Claude — IDE integration, CLAUDE.md setup, workflow automation

Claude Code for CI/CD — headless Claude in a build pipeline

Structured Data Extraction — parsing documents into validated JSON schemas

Each scenario has ~10 questions layered through the exam. Recognizing the scenario type quickly helps you apply the right mental model — instead of treating each question in isolation.

Realistic Study Timeline

Here's a timeline that works for most candidates who already build software but are new to Claude's production patterns:

Week 1 — Foundation

• Read the official Anthropic Academy courses (free, on Skilljar) — all four modules
• Download and read the official Exam Guide PDF
• Build something: a simple Claude Code workflow or one MCP server

Week 2 — Domain depth

• Focus on Domains 1 and 3 (the two highest-weight areas)
• Read the Anthropic docs on agents, tool use, and structured outputs
• Take notes on tradeoffs: when to use X vs. Y

Week 3 — Weak spots + practice

• Take the official practice exam (60 questions, same format)
• Identify your lowest-scoring domains
• Re-study those specific topics with the domain breakdown above

Week 4 — Consolidation

• Take a full practice test under real conditions (120 min, no interruptions)
• Aim for 850+ before scheduling the real exam
• Review every wrong answer until you understand why it's wrong

Most people who pass on the first attempt report 20–30 hours of prep over 2–4 weeks if they already code with Claude daily.

The 5 Most Common Mistakes (and How to Avoid Them)

1. Treating it like a knowledge quiz. This is a scenario exam. The right answer is always about sound architectural judgment — read the scenario carefully before looking at the options. 2. Underestimating Domain 2. Claude Code configuration feels like a small topic until you realize 20% of your score depends on knowing exactly how CLAUDE.md hierarchies, hooks, and permissions work. 3. Guessing on MCP tool boundaries. The exam has questions where two answers seem reasonable but one violates the principle of least privilege. When in doubt, the more restrictive tool design is usually correct. 4. Ignoring the official practice exam. Anthropic's official practice exam is 60 questions in the exact same format as the real exam. There is no better signal for your readiness. Take it early enough that you have time to fix weak spots. 5. Over-optimizing for cost over reliability. Several exam questions present a cost-saving architecture that introduces a failure mode. The exam rewards reliability-first thinking — budget optimization is a secondary concern.

Resources That Actually Help

For additional practice questions beyond the official 60, the AI for Anything CCA Practice Test Bank has 200+ scenario-based questions across all five domains with detailed explanations for every answer — including why the wrong answers are wrong.

What Passing Actually Gets You

The CCA-F certification is still relatively new (launched March 2026), but the signal value is strong for a few reasons:

• It's hard to fake — scenario-based exams filter out people who've only read docs
• Anthropic controls the brand — it won't get diluted by thousands of competing prep courses the way AWS certs did
• Demand is growing — companies deploying Claude in production want engineers who understand the architecture, not just the chat interface
• It builds on itself — the CCA-F is explicitly designed as the foundation for more advanced certifications Anthropic has signaled are coming

If you're an engineer who already works with Claude, this certification formalizes the knowledge you've built. If you're newer to the ecosystem, it's a forcing function that gets you to a production-level understanding faster than you would organically.

Key Takeaways

• The CCA-F is 60 scenario-based questions, 120 minutes, passing score 720/1000
• Five domains: Agentic Architecture (27%), Claude Code Config (20%), Prompt Engineering (20%), MCP & Tools (18%), Context Management (15%)
• You cannot memorize your way through — you need production intuition
• Build something real before the exam (an agent, an MCP server, a CLAUDE.md workflow)
• Aim for 850+ on the official practice exam before scheduling the real thing
• Most candidates pass in 20–30 hours of prep over 2–4 weeks

Start Preparing Today

The best way to prepare for scenario-based questions is to practice with scenario-based questions.

Explore the AI for Anything CCA-F Practice Test Bank →

Our test bank includes 200+ questions across all five domains, modeled after the real exam format, with detailed explanations for every answer. You'll know exactly which domains to strengthen before you walk into the real exam.

If you're still exploring whether the CCA-F is worth pursuing, read our breakdown of the best AI certifications for salary impact in 2026 to see where the CCA-F ranks against AWS, Google, and OpenAI credentials.