Claude Certified Architect · Foundations (CCAF)

Master the CCA-F Certification Exam

The most focused free practice platform for the Claude Certified Architect — Foundations (CCAF) exam. Study with real-style questions, detailed explanations, and domain-specific quizzes — all at no cost.

Start Practice Exam → Learn About the Exam

Exam Overview

What Is the Claude Certified Architect — Foundations (CCAF)?

The CCAF is Anthropic's entry-level certification for professionals who want to demonstrate a solid understanding of Claude AI systems, responsible deployment, and foundational architectural principles.

About the Certification

The Claude Certified Architect — Foundations (CCAF) certification is designed for developers, solution architects, product managers, and AI practitioners who work with or intend to build applications powered by Claude. The credential validates that a candidate understands how Claude models function, how to integrate them responsibly into real-world systems, and how to evaluate their outputs effectively.

This is a vendor-specific certification aligned with Anthropic's Claude API ecosystem, covering topics from model architecture and capabilities to safety considerations and prompt engineering best practices. Earning the CCAF demonstrates professional competency in one of the most widely adopted large language model platforms available today.

Who Should Take This Exam?

The CCAF exam is ideal for candidates in one or more of the following roles:

Software engineers and developers building applications on the Claude API
Solutions architects designing AI-powered workflows and enterprise integrations
Product managers overseeing AI feature development and model capabilities
AI/ML practitioners transitioning to large language model application development
Technical consultants advising organizations on Claude adoption strategies
Students and researchers seeking formal recognition of their Claude knowledge

Recommended Prerequisites

While there are no strict prerequisites, candidates are strongly encouraged to have working familiarity with basic programming concepts, REST APIs, and general machine learning principles. Prior hands-on experience using the Claude API or Claude.ai will significantly benefit your preparation. Reading Anthropic's official documentation before exam day is considered essential.

How to Register

Candidates register for the CCAF exam through Anthropic's official certification portal. After registration, you will receive access to an exam guide and a list of official study resources. We recommend using this site's practice questions alongside official materials to maximize your readiness.

💡 Pro Tip

Focus heavily on understanding the reasoning behind correct answers, not just memorizing facts. The CCAF exam tests applied knowledge — including how you would approach real deployment scenarios and responsible AI considerations.

Quick Exam Facts

FormatMultiple Choice

Questions~60 Questions

Duration90 Minutes

Passing Score~70%

DeliveryOnline Proctored

Validity2 Years

LanguageEnglish

Exam Domains at a Glance

Claude Fundamentals25%

Prompt Engineering20%

Safety & Responsible AI22%

API & Deployment18%

Ethics & Governance15%

Knowledge Domains

CCAF Exam Domains Explained

The CCAF exam is organized into five core domains. Understanding each domain's scope and weight is the first step in building a targeted study plan.

25%

Domain 1: Claude AI Fundamentals

This domain covers how Claude models are built, how they process language, how they differ from other LLMs, and what their core capabilities and limitations are. Candidates must understand context windows, model versioning (e.g., Claude Haiku vs. Opus), tokenization, and sampling parameters like temperature.

Model Architecture Context Windows Tokenization Model Versions Capabilities & Limits

20%

Domain 2: Prompt Engineering

This domain tests your ability to craft effective prompts that produce accurate, safe, and useful outputs. Topics include zero-shot vs. few-shot prompting, chain-of-thought reasoning, system prompts, role prompting, and common failure modes like hallucination triggers and ambiguous instructions.

System Prompts Few-Shot Examples Chain-of-Thought Output Formatting Failure Modes

22%

Domain 3: Safety & Responsible AI

One of the highest-weighted domains, this section covers Anthropic's Constitutional AI (CAI) framework, content moderation, harm avoidance strategies, and how Claude is trained to be helpful, harmless, and honest. Candidates must understand how to evaluate model safety in different deployment contexts.

Constitutional AI Harm Avoidance Content Policy Bias Mitigation Safety Evaluation

18%

Domain 4: API Integration & Deployment

This domain covers practical integration patterns for the Claude API: authentication, rate limits, streaming responses, multi-turn conversation management, tool use (function calling), and best practices for deploying Claude in production environments. Basic REST and JSON handling is assumed.

API Authentication Rate Limits Streaming Tool Use Production Patterns

15%

Domain 5: Ethics & Governance

Candidates must demonstrate knowledge of AI governance frameworks, data privacy considerations, regulatory compliance (such as the EU AI Act), responsible disclosure, and organizational AI policies. This domain also covers how to document AI system decisions and the ethical responsibilities of AI architects.

AI Governance Data Privacy EU AI Act Responsible Disclosure Documentation

Study Guide

How to Prepare for the CCAF Exam

A structured, domain-aware study plan will dramatically improve your chances of passing. Here is exactly how to approach your preparation from day one.

Step 1 — Read the Official Exam Blueprint First

Before opening any study material, download and read the official CCAF Exam Guide from Anthropic's certification portal. This document describes what each domain covers, what level of understanding is expected, and how questions are weighted. Many candidates skip this step and study the wrong things in the wrong proportions. The blueprint is your map — treat it as the most important document in your preparation.

Pay particular attention to the action verbs used in each domain objective. Words like "identify," "compare," "evaluate," and "design" indicate very different cognitive levels. A domain objective that says "evaluate the trade-offs between model versions" requires deeper thinking than one that says "identify the primary context window sizes for each Claude model."

Step 2 — Build Your Claude Fundamentals First

Domain 1 (Claude AI Fundamentals) is worth 25% of the exam and underpins your understanding of every other domain. Start here. Read Anthropic's model cards, the documentation on Claude's model families, and the research papers on Constitutional AI. Make sure you thoroughly understand the following:

How transformer-based language models generate output one token at a time, and why this matters for response consistency
What a context window is, why it matters, and how different Claude versions differ in context window size
The difference between Claude Haiku, Sonnet, and Opus model tiers and when each is the right choice
How sampling parameters like temperature and top-p affect output diversity and creative variation
The concept of "grounding" and why factual accuracy is probabilistic rather than guaranteed
How Claude's outputs are deterministic at temperature=0 and increasingly random at higher values

📌 Key Insight

Claude models are trained using a technique called Constitutional AI (CAI). Understanding CAI — what it is, why Anthropic developed it, and how it shapes Claude's behavior — is critical for both Domain 1 and Domain 3. This is one of the most frequently tested topic areas across the entire exam.

Step 3 — Master Prompt Engineering Through Hands-On Practice

Reading about prompt engineering is not enough — you must practice writing prompts and analyzing outputs. Open the Claude API or Claude.ai and experiment deliberately with each of these techniques:

Zero-shot prompts: Ask Claude a question with no examples. Note the output quality, format, and consistency across repeated runs.
Few-shot prompts: Provide 2–3 examples of the desired input-output pattern before your actual query. Compare how the presence of examples changes output quality.
System prompts: Write a system prompt that defines a persona, restricts scope, or requires a specific output format. Test how it constrains behavior across varied user inputs.
Chain-of-thought prompting: Add "Think step by step before answering" to a complex reasoning question. Observe how accuracy and explanation quality change.
Role prompting: Assign Claude a professional role (e.g., "You are a senior software engineer reviewing code for security vulnerabilities") and observe changes in tone, depth, and specificity.

Document what works and what fails in a notebook. The exam will present scenarios where you must choose the best prompting approach — and that judgment comes from hands-on experimentation, not theory alone.

Step 4 — Study Safety and Responsible AI Deeply

Domain 3 is the second-highest weighted domain at 22%. More importantly, safety topics appear embedded within other domains as well — making your understanding of responsible AI genuinely cross-cutting throughout the exam. Study Anthropic's published guidelines carefully, focusing on:

The three pillars of Claude's training objective: Helpful, Harmless, and Honest (HHH)
How Constitutional AI uses a set of written principles to guide the model in self-critiquing and revising responses
Common categories of harmful content and how Claude is trained to recognize and appropriately refuse or redirect them
The distinction between hard safety limits (absolute refusals that cannot be overridden) and soft defaults (behaviors that can be adjusted through system prompts in legitimate contexts)
How to evaluate AI outputs for factual accuracy, demographic bias, and potential real-world harms before deploying in production

Step 5 — Get Hands-On with the Claude API

Domain 4 (API Integration) is the most practical domain and the one where preparation through actual building pays off most clearly. Even a simple Python script that calls the API, manages a multi-turn conversation, and handles errors will teach you far more than an hour of passive reading. Focus your API study on:

Authenticating API calls with API keys and understanding security best practices such as never hardcoding keys in source code
Understanding the structure of both the request body and response payload in the Claude Messages API
Implementing conversation history management by correctly appending and passing prior messages in each successive API call
Using the streaming API to handle long responses incrementally and improve perceived response time
Implementing tool use (function calling) to allow Claude to invoke external functions and return structured results
Understanding rate limits, error codes (such as 429 Too Many Requests and 529 Overloaded), and appropriate retry and backoff strategies

Step 6 — Review Ethics, Governance, and Compliance

Domain 5 (Ethics and Governance) is the smallest domain by weight at 15%, but it is the domain where technically strong candidates frequently lose points due to underpreparation. Study current AI governance frameworks with enough depth to apply them to realistic scenarios, including:

The EU AI Act — how it classifies AI systems by risk tier and what specific obligations apply at each level (minimal, limited, high, and unacceptable risk)
The NIST AI Risk Management Framework (AI RMF) — its four core functions: Govern, Map, Measure, and Manage
Data minimization and privacy-by-design principles and how they apply to LLM applications that process personal information
The concept of model cards and system cards as transparency tools and how to evaluate whether an AI system's documentation is adequate
Responsible disclosure practices when discovering safety-relevant issues or unexpected behaviors in AI systems

Step 7 — Take Full Practice Exams Under Timed Conditions

In the final two weeks before your exam, switch to full-length timed practice tests using our QuizMaster app. Time pressure is a significant variable — the real exam gives you 90 minutes for approximately 60 questions, which works out to about 90 seconds per question. Practice flagging uncertain questions and returning to them efficiently rather than getting stuck. After each practice test, spend as much time analyzing wrong answers as you did taking the test itself.

⏱ Time Management Strategy

On the real exam, answer every question you're confident about first, then flag the uncertain ones. Never leave a blank answer — most certification exams have no penalty for guessing, so a considered guess is always better than no answer. On review, trust your first instinct unless you can identify a specific logical reason to change your answer.

Sample Questions

Practice Questions with Full Explanations

Try these sample questions spanning all five CCAF domains. Each includes the correct answer and a detailed explanation — just like every question in the full app.

Domain 1 · Claude Fundamentals

A developer is building a document summarization application that must handle legal contracts up to 150,000 tokens in length in a single call. Which model selection criterion is MOST important for this use case?

Selecting the model with the highest reasoning capability regardless of context window size
Selecting a model with a context window large enough to accommodate 150,000 tokens in a single API call
Selecting the fastest model available to minimize end-user wait time
Selecting the least expensive model to optimize cost efficiency before other factors

Explanation

When input size is a hard technical constraint, context window capacity must be the primary deciding factor. A model that cannot fit the entire document in context would require chunking strategies that add development complexity, risk losing cross-document context, and may produce less coherent summaries. For long legal contracts, you must first confirm the model can handle the full input — only then should you optimize for reasoning quality, latency, or cost. Choosing a faster or cheaper model that cannot hold the full document would cause functional failures, not just trade-offs.

Domain 2 · Prompt Engineering

A team is experiencing inconsistent output formatting when using Claude to generate structured JSON responses for a downstream data pipeline. Which prompting technique is MOST effective for ensuring consistent, parseable output?

Increasing the temperature parameter to encourage more creative and flexible output structure
Adding polite and deferential language to the request to improve model compliance behavior
Providing a concrete JSON schema example directly in the prompt and explicitly instructing Claude to follow it exactly
Using a longer prompt with more context about the application's general purpose and business goals

Explanation

Few-shot examples that demonstrate the exact desired output format are the most reliable technique for achieving consistent structured outputs. By embedding a concrete JSON schema example and requiring it be followed precisely, you eliminate ambiguity about structure, field names, and data types. Increasing temperature would do the opposite — it increases output variation, which is exactly the problem being solved. Polite language has no measurable effect on structural consistency. Adding general context about the application does not provide the formatting constraints the model needs to produce reliably parseable JSON.

Domain 3 · Safety & Responsible AI

Which of the following BEST describes Anthropic's Constitutional AI (CAI) training methodology?

A framework where human moderators manually review and approve every model response before it is delivered to the end user
A technique that trains Claude to refuse all requests that contain sensitive keywords by default
A training method where a set of human-written guiding principles is used to have the model critique and revise its own responses, reducing harmful outputs without requiring human review of every example
A post-processing system that applies external content filters to model outputs after generation is complete

Explanation

Constitutional AI is a training methodology developed by Anthropic in which a model is given a written set of principles — a "constitution" — and is trained to evaluate its own outputs against those principles, identify problematic aspects, and revise them. This self-critique process occurs during training, not at inference time, making it fundamentally different from post-hoc filtering or human moderation of individual responses. One of CAI's key advantages is that it allows safety alignment to scale without requiring a human reviewer for every training example. On the CCAF exam, candidates must distinguish CAI from real-time content filtering, keyword blocking, and RLHF-based approaches.

Domain 4 · API Integration & Deployment

When implementing a multi-turn conversational application using the Claude Messages API, how must conversation history be managed across successive API calls?

Claude automatically stores conversation history server-side and retrieves it using a session ID passed in each new request
The application must pass the complete conversation history as an ordered array of prior messages in every new API request, because the API is completely stateless
Only the most recent user message needs to be sent — Claude uses a compressed internal memory mechanism to recall earlier turns
Conversation state is managed automatically through HTTP cookies that Claude's API servers set and maintain

Explanation

The Claude Messages API is completely stateless — it has no memory of previous API calls whatsoever. To maintain a coherent multi-turn conversation, your application must append each new user message and the corresponding Claude response to a messages array, and then send the full updated array with every subsequent API request. This is a fundamental and frequently tested architectural pattern for all LLM-powered conversational applications. Developers must also account for the fact that this approach consumes tokens from the context window for every prior message included, which has both cost and context-limit implications that must be managed in long conversations.

Domain 5 · Ethics & Governance

An organization is deploying a Claude-powered hiring assistant to automatically screen and rank résumés before human review. Under the EU AI Act risk classification framework, how would this system most likely be classified?

Minimal risk — the system is simply processing text and does not make final employment decisions autonomously
Limited risk — requiring only transparency obligations, such as disclosing to applicants that AI was involved
High risk — employment screening systems that influence individuals' access to work are explicitly listed as high-risk AI use cases in Annex III of the EU AI Act
Unacceptable risk — making the system prohibited from deployment across the European Union entirely

Explanation

The EU AI Act explicitly designates AI systems used in employment contexts — including CV screening, candidate ranking, interview scheduling, and performance evaluation — as high-risk AI systems under Annex III. High-risk systems face substantial compliance obligations: mandatory conformity assessments, comprehensive risk management documentation, human oversight requirements, detailed logging of system decisions, transparency obligations to affected individuals, and registration in an EU database. This classification applies regardless of whether a human makes the final hiring decision — the fact that the AI materially influences which candidates are seen is sufficient to trigger high-risk status. This is one of the most widely tested governance facts on the CCAF exam.

Access 200+ Full Practice Questions →

Exam Day Tips

8 Tips to Pass the CCAF on Your First Attempt

Hard-won advice for navigating AI certification exams with confidence.

Read Every Option Before Answering

Many CCAF questions include a "most correct" option alongside plausible-sounding distractors. Committing to the first answer that sounds right is one of the most common exam mistakes. Read all four options, eliminate the obvious wrong answers, then choose carefully between the remaining ones.

Understand "Best" vs. "Correct"

Certification exams often ask what the "BEST" or "MOST appropriate" action is — not what is technically possible. Multiple answers may be valid, but only one reflects recommended professional practice. Always think from the perspective of an architect following established best practices.

Anchor Your Answer to the Scenario

The CCAF frequently presents scenario-based questions where the context changes the right answer. A prompting strategy ideal for creative writing may be wrong for structured data extraction. Always ground your answer in the specific use case described, not general knowledge.

Trust HHH for Safety Questions

For Domain 3 questions, Claude's training principles — Helpful, Harmless, Honest — and Anthropic's published safety guidelines are your guide. If an answer option appears to undermine human oversight or bypass safety controls, it is almost certainly wrong, even if it sounds pragmatically efficient.

Memorize API Concepts Precisely

Domain 4 questions often hinge on exact technical knowledge — knowing that the Messages API is stateless, that streaming uses server-sent events, or that tool use requires a specific request format. Make flashcards for key API behaviors and test yourself until they are automatic.

Flag and Return Strategically

Use the exam's flagging feature to mark uncertain questions and continue forward. When reviewing flagged items, trust your first instinct unless you can identify a specific and logical reason to switch. Research consistently shows that changing answers on a vague feeling tends to reduce scores.

Rest the Night Before

Cognitive performance is measurably impaired by sleep deprivation. A well-rested brain outperforms a cramming-exhausted one every time. Stop studying at least 24 hours before your exam. Use that time to rest, do a light notes review, and build confidence rather than attempting to learn new material.

Review Wrong Answers Actively

Do not just take practice exams to see your score — analyze every wrong answer carefully. For each mistake, write down precisely why the correct answer is right and why your chosen answer is wrong. This active review process is the single most effective preparation technique for improving your score.

Ready to Test Your Knowledge?

Launch the full QuizMaster app for 200+ practice questions, timed exams, and detailed scoring across all five CCAF domains.

Open the Practice App →

About This Site

An Independent CCAF Study Resource

This site is an independent educational resource built by AI practitioners dedicated to helping professionals prepare for the Claude Certified Architect — Foundations (CCAF) certification. We are not affiliated with, sponsored by, or endorsed by Anthropic, PBC.

All practice questions are original, written and reviewed by individuals with hands-on experience using the Claude API and deploying LLM-powered applications in production. Our question bank is updated regularly to reflect changes in the exam blueprint, Anthropic's published documentation, and evolving best practices in responsible AI deployment.

Our mission is simple: make high-quality CCAF exam preparation freely accessible to everyone, regardless of budget or background. The core practice experience — including all sample questions and study guides — is and will remain free.

Original, handcrafted practice questions — not AI-generated filler
Aligned with the official CCAF exam blueprint and domain weightings
Detailed explanations for every question, correct and incorrect
No login required to start practicing immediately
Regularly reviewed and updated as the exam and documentation evolve

Topics Covered in the Practice App

Claude model families, versions, and selection criteria
Context windows, tokenization, and sampling parameters
Constitutional AI and the HHH training framework
Prompt engineering patterns and failure modes
System prompts, tool use, and multi-turn conversations
Claude Messages API structure, auth, and error handling
Streaming responses and rate limit management
EU AI Act risk classification and compliance obligations
NIST AI Risk Management Framework (AI RMF)
Bias evaluation and harm mitigation strategies
AI governance documentation and model cards
Responsible disclosure and incident response protocols

Get In Touch

Contact Us

Found an error in a question? Have a suggestion for new content or a topic that should be covered? We want to hear from you — every piece of feedback helps improve the resource for everyone.

ccaf@mtourk.com

Privacy Policy

Last updated: April 2026

1. Overview

mtourk.com ("we," "our," or "us") operates this website as a free educational resource for candidates preparing for the Claude Certified Architect — Foundations (CCAF) certification. We are committed to protecting your privacy and being transparent about how we handle information. This Privacy Policy explains what data we collect, why we collect it, and how it is used.

2. Information We Collect

This site does not require user registration and does not collect personally identifiable information (PII) by default. When you use our contact email address to send us a message, we receive only the information you voluntarily include. We do not store contact messages in any persistent database.

3. Automatically Collected Technical Data

Like most websites, our servers may log standard technical information including your IP address, browser type, operating system, referring URL, and the pages you visit. This data is used solely for diagnosing technical issues and understanding aggregate traffic patterns. It is not used to identify individual users and is not shared with third parties for marketing purposes.

4. Google AdSense and Advertising Cookies

We use Google AdSense to display third-party advertisements on this site. Google AdSense uses cookies and similar web technologies to serve ads based on your interests, including your prior visits to this website and other websites across the internet. The collection and use of information by Google is governed by Google's Privacy Policy.

You have the following options to control personalized advertising:

Visit Google Ads Settings to opt out of personalized advertising
Visit www.aboutads.info to opt out via the Digital Advertising Alliance
Use your browser's built-in cookie management settings to block or delete cookies

Opting out of personalized advertising does not remove all ads from the site — it means ads displayed will no longer be based on your browsing history and interests.

5. Google Analytics

We may use Google Analytics to collect anonymized, aggregated information about how visitors use this site, such as page view counts, session duration, and geographic region. This data does not identify individual users. You may opt out of Google Analytics tracking by installing the Google Analytics Opt-out Browser Add-on.

6. Third-Party Links and Companion Application

This website contains links to our companion application hosted at claude.mtourk.com and to external third-party resources including official Anthropic documentation and governance frameworks. We are not responsible for the privacy practices or content of those external sites. We encourage you to review the privacy policy of any third-party site you visit. Our companion application at claude.mtourk.com may have its own data handling practices.

7. Children's Privacy

This website is not directed at children under the age of 13. We do not knowingly collect or solicit personal information from children. If you believe a child has provided us with personal information, please contact us immediately and we will delete it promptly.

8. Data Retention

We do not maintain persistent user databases. Server log data is typically retained for no more than 30 days for technical diagnostic purposes and is then deleted or fully anonymized.

9. Changes to This Policy

We may update this Privacy Policy periodically to reflect changes in technology, applicable law, or our data handling practices. When we make material changes, we will update the "Last updated" date at the top of this page. Your continued use of this site after any changes are posted constitutes your acceptance of the updated policy.

10. Contact

For any questions or concerns regarding this Privacy Policy or our data practices, please contact us at ccaf@mtourk.com.

11. Disclaimer

This site is an independent educational resource. It is not affiliated with, sponsored by, or endorsed by Anthropic, PBC. "Claude" is a trademark of Anthropic, PBC. All other trademarks and registered trademarks referenced on this site are the property of their respective owners. Questions and content on this site represent our independent interpretation of publicly available information and should not be considered official exam preparation material from Anthropic.