How to Pass the Claude Architect Exam (CCAR-F): A 4-Week Study Plan

About 40 hours over four weeks, weighted to the official blueprint, with the 27% gap that Anthropic’s own prep list leaves open. One verified passer did it in 60 hours over two weeks and scored 843.

The Claude Certified Architect: Foundations exam is 60 questions in 120 minutes, and you need 720 out of 1,000 to pass. It is scenario-based, which means you get a broken agentic system and four plausible fixes, and you pick the one a good architect would actually choose.

That framing matters more than any fact you memorize. Plan accordingly.

Before week one: two things

Download the official exam guide PDF. It is free, linked from Anthropic’s Partner Academy certification page, and it contains the full blueprint plus 12 sample questions with answer rationales. That is four times what the other exams publish. It is better than anything you can buy.

Know the weights, because this plan follows them.

Domain	Weight
Agentic Architecture & Orchestration	27%
Claude Code Configuration & Workflows	20%
Prompt Engineering & Structured Output	20%
Tool Design & MCP Integration	18%
Context Management & Reliability	15%

Week 1: close the biggest gap first (about 12 hours)

Agentic Architecture is 27% of the exam. It is also the domain that Anthropic’s own curated prep list does not really cover, so start here rather than saving it.

Take the free Introduction to subagents course. About 45 minutes.
Read the Agent SDK agent loop documentation. It is the densest single page for this domain.
Build an agent. An agentic loop, a coordinator with two subagents, and one hook that intercepts a tool call. This is the week’s real work.

By Friday you should be able to explain, without looking it up, what a subagent inherits and what it does not. It gets its own system prompt, CLAUDE.md and git status. It does not get the parent’s message history, and only its final response comes back.

Week 2: Claude Code, for real (about 10 hours)

Twenty percent of the exam, and it is the domain where reading gets you the least.

Work through Claude Code in Action.
Configure a real repository. A CLAUDE.md at both user and project level, one custom skill, a .claude/rules/*.md file with paths: frontmatter scoping it to a glob, and a claude -p step in a CI pipeline.

Do all four. Each maps to a numbered task statement in the blueprint, and the exam asks about them at a level of detail that assumes you have done it.

Week 3: tools, MCP, and structured output (about 10 hours)

Two domains, 38% combined.

Introduction to MCP, then MCP Advanced Topics.
Build a structured data extraction pipeline using tool-use JSON schemas, with validation and retry.
Implement structured error responses with the isError flag so a failing tool tells the model enough to correct itself.

Do not shortcut prompt engineering this week. A candidate who scored 886 overall got 50% on this domain, and it is 20% of the exam. The trap is assuming you have it because you prompt daily. What gets tested is enforcement: tool use with JSON schemas, validation loops, few-shot for consistency.

Week 4: the six scenarios and the sample questions (about 8 hours)

Four of these six appear on your exam form. You will not know which four.

Customer Support Resolution Agent
Code Generation with Claude Code
Multi-Agent Research System
Developer Productivity with Claude
Claude Code for Continuous Integration
Structured Data Extraction

Sketch an architecture for each. Then work all 12 sample questions from the guide and write out your reasoning, not just your answer. If your reasoning matches the rationale, you are ready. If you got the right answer for the wrong reason, that is the gap to close.

Finish the week on Context Management and Reliability, which is 15% and mostly conceptual.

The five principles to carry in

Published free by a candidate who passed with 843. They generalize across every scenario, and they are worth more than any fact list.

Constrain, do not add. More tools, more agents, bigger context window is almost always wrong.
Prompts suggest. Systems enforce. If a rule must always hold, put it in code or a hook.
Fix the earliest layer. Most wrong answers are sound fixes applied too late in the pipeline.
Match the fix to the failure. Vague rule needs explicit criteria. Inconsistent output needs few-shot. Invalid JSON needs a schema. Valid but wrong needs semantic validation.
Do not collapse states. Escalate on policy, not vibes. A tool failing is not the same as finding nothing.

Underneath all five: you are building a reliable system around a fallible model, so fine-tune it and tell it to be more careful are almost never the answer.

And one calibration. The same passer noted that correct answers usually felt reasonable, efficient and balanced, not too aggressive and not too conservative. So do not reflexively pick the most locked-down option. If an answer bolts a heavyweight control onto a low-stakes failure, it is probably the distractor.

Exam day

Check in 30 minutes early. That window covers a tech check, photo ID, and a 360-degree room scan.
Unplug the second monitor. Single display only, and no headphones or headsets.
No VPN, no corporate network. Run the OnVUE system test on the exact machine and connection you will use.
Clear the desk. No paper, no pens. A digital whiteboard is provided in the exam interface, so practice thinking on screen beforehand if you normally sketch on paper.
Bring unexpired government photo ID with a name matching your registration. Digital IDs are refused.

You get your score on screen at the end, and the report breaks down percentage correct per domain. If you fail, the wait is 14 days, then 30, then 90, with four attempts allowed per rolling twelve months and no retake discount.

Drill it week by week

Free CCAR-F practice questions, weighted to the official blueprint, with a plain-English explanation on every one.

Free practice questions →

Keep reading: The CCAR-F study guide · CCAR-F vs CCAR-P · All four Claude exams explained

HOW TO // AI is not affiliated with or endorsed by Anthropic. CCAR-F, Claude Certified Architect, Claude and Claude Code are trademarks of Anthropic PBC; we reference them descriptively. All content is original.