What the CCAR-F Exam Actually Tests
The Claude Certified Architect: Foundations (CCAR-F) certification is Anthropic's entry-level certification for Claude Code practitioners. It tests whether you understand how to use Claude Code effectively, safely, and in alignment with Anthropic's documented best practices.
The exam is not a coding test. You will not be asked to write code. You will be asked to demonstrate that you understand the concepts, tools, workflows, and safety principles that govern how Claude Code operates.
This matters for healthcare professionals specifically because the agentic architecture domain, the heaviest on the exam, covers human-in-the-loop gating and programmatic enforcement patterns that map directly onto the compliance requirements you already work within.
The Five Domains
Domain 1: Agentic Architecture & Orchestration (27% of the exam, the heaviest domain)
An agent is Claude running in a loop: call tools, get results, reason, call more tools, stop when done. The critical signal is stop_reason, not the content type of the response. This domain also covers multi-agent orchestration, subagent context isolation (subagents get zero automatic memory of the parent session), human-in-the-loop gating, fallback design, and RAG pipelines.
The most tested concept: for compliance-critical rules with financial or safety consequences, use programmatic hooks, not system prompt instructions. Prompts are probabilistic. Hooks are deterministic.
Domain 2: Tool Design & MCP Integration (18%)
MCP (the Model Context Protocol) is the standard for giving Claude access to external tools, and tool descriptions are the primary routing mechanism, not just documentation. This domain also tests stdio vs Streamable HTTP transport selection (SSE is deprecated), authentication patterns, and structured error handling in tool responses.
The most common miss: candidates still learn SSE as a current transport option. It was deprecated in the MCP spec, replaced by Streamable HTTP for remote or multi-client use. stdio stays correct for local, single-session development.
Domain 3: Claude Code Configuration & Workflows (20%, more specific than most candidates expect)
Claude Code is controlled through a hierarchy of config files, and this domain tests exact file paths, not general familiarity. Project memory lives at ./CLAUDE.md in the project root, committed to git, not inside the .claude/ folder. That folder holds settings, rules, and commands. This domain also covers the hooks event model, slash command frontmatter, and monorepo patterns.
Domain 4: Prompt Engineering & Structured Output (20%)
This domain tests when each prompting technique applies and its limits, It also covers model selection tradeoffs (Haiku for volume, Sonnet for most production work, Opus for the highest-consequence tasks) and validation retry loop design.
Domain 5: Context Management & Reliability (15%)
This is often the easiest set of points on the exam once you know the material. It tests context engineering: compaction, structured note taking, context rot, prompt caching with cache_control markers, token budget management, and provenance tracking across multiple sources.
Where Most Candidates Lose Points
Based on the actual structure and weighting of the exam, the most common failure points are:
Choosing prompts over programmatic enforcement. This is the single most tested concept. When a business rule has financial or safety consequences, prompt instructions have a non-zero failure rate. The exam always wants a hook or a programmatic gate instead.
Checking response content type instead of stop_reason. A common code pattern checks response.content[0].type === 'text' to decide whether the agentic loop is done. This breaks in production because Claude can return text and a tool_use block in the same response. Always check stop_reason.
Confusing SSE and stdio transports. A confirmed Domain 2 topic. stdio is for local, single-client development. Streamable HTTP is for remote or multi-client use. Many candidates don't know SSE was deprecated.
Assuming subagents inherit context. Subagents start with zero automatic memory of the parent session or sibling agents. Every piece of information they need has to be explicitly passed in their prompt.
Skipping context engineering and prompt caching. Domain 5 is the smallest slice of the exam at 15 percent, which makes it the easiest to under study. The caching minimum is model dependent rather than a flat number, and cached prefixes carry a 5 minute TTL by default with a 1 hour option.
How to Prepare Efficiently
The most efficient preparation path is:
1. Read the official Claude Code documentation completely. Every answer on the exam is sourced from this documentation. There is no trick content.
2. Work through practice questions by domain, not by difficulty. Identify which domains you are weakest in and focus there.
3. For each question you get wrong, find the specific documentation page it came from. Do not just memorize the correct answer. Understand why it is correct.
4. Pay special attention to Domain 1 (Agentic Architecture & Orchestration). At 27% of the exam, it carries more weight than any other domain, and its programmatic-enforcement concepts show up as distractors throughout the other domains too.
5. Do not rely on general AI knowledge. Claude Code has specific behaviors that differ from other AI tools. What is true for GPT-4 or Gemini may not be true for Claude Code.
The 226 practice questions in the CCAR-F Exam Prep product cover all five domains with full explanations and documentation citations. Every question was written against the official Anthropic documentation, not from memory or general AI knowledge.
The Practical Value of the Certification
The CCAR-F certification signals to employers, clients, and collaborators that you understand Claude Code at a documented, tested level. For healthcare professionals building AI-powered tools, it is also a credibility signal that you have engaged seriously with the safety and compliance dimensions of AI development.
The exam is not difficult if you have read the documentation. The challenge is knowing which parts of the documentation are most heavily tested and how to apply the concepts to scenario-based questions.
That is exactly what the practice questions are designed to address.