Claude Certified Architect practice exam: 60 questions, free
For the certification I wrote myself 60 exam questions. I passed: 833 out of 1000. Now I'm giving them away for free. They run in the browser and you don't need an account.
On July 15 I passed Claude Certified Architect Foundations with a score of 833 out of 1000, against a pass line set at 720. I described the exam itself in a separate post, and put together the full Claude Certified Architect certification guide. Today I'm handing you the tool I prepared with. A full practice exam, 60 questions. I wrote them myself, from scratch, based on the official exam guide. You can take it for free on the practice exam page or grab the whole thing from GitHub.
What exactly you get
These are not questions leaked from the exam. They're not reworked material from anyone else either. I wrote all 60 questions myself while studying, modeled on Anthropic's public exam guide: the same domains, the same weights and the same format of four options with one correct. Then I tried to break every question. I checked two things. Does the correct answer actually hold up. And is none of the distractors accidentally correct too.
The entire exam is literally one HTML file. It runs in the browser. You install nothing and send data to no one. Progress is saved locally on your side, so you can close the tab halfway through and come back tomorrow exactly where you left off.
What the real exam covers
But first, the facts. Anthropic launched the certification on March 12, 2026. The exam carries the code CCAR-F and consists of 60 multiple-choice questions, for which you get 120 minutes under full proctoring. You pass from 720 points on a scale from 100 to 1000. The material splits into five domains with official weights:
| Domain | Weight | Questions in my mock |
|---|---|---|
| D1 · Agentic Architecture & Orchestration | 27% | 16 |
| D2 · Tool Design & MCP Integration | 18% | 11 |
| D3 · Claude Code Configuration & Workflows | 20% | 12 |
| D4 · Prompt Engineering & Structured Output | 20% | 12 |
| D5 · Context Management & Reliability | 15% | 9 |
I took one thing away from the real exam. It doesn't ask about definitions. It asks about decisions. Almost every question gives you several technically correct options and makes you pick the one that fits the situation, the budget and the requirements. I built my questions the same way. Each has a production scenario and four answers. The differences between them are subtle.
This is what I do hands-on — advising on AI strategy and building agents that survive the demo.
Three questions to try
Instead of talking, I'll show you. Here are three questions from the mock, from three different domains. Can you get them without hints?
QUESTION · D1 · AGENTIC ARCHITECTURE
Company policy states that any refund above $500 must be approved by a manager. An audit shows the support agent has autonomously called process_refund for higher amounts several times. What is the best way to guarantee the policy holds?
- A. Add a PreToolUse hook on
process_refundthat blocks any call above $500 and redirects the flow to human escalation. - B. Add a system-prompt line instructing the agent to always escalate refunds above $500.
- C. Remove
process_refundfrom the agent's toolset so every refund is handled by a human. - D. Fine-tune the model on historical tickets so it learns the $500 threshold.
Correct: A. A rule that must never be broken calls for a deterministic mechanism, not a request. The hook blocks the call regardless of what the model decides. A prompt instruction is exactly what the audit already exposed, and removing the tool throws away all the legitimate automation below the threshold.
QUESTION · D4 · STRUCTURED OUTPUT
For documents that lack a purchase-order number, Claude invents a plausible-looking number about 11% of the time instead of signaling absence. Downstream systems treat the fabricated values as real. What is the best schema-level fix?
- A. Add a required confidence float to every field and discard low values downstream.
- B. Switch
tool_choiceto "auto" so the model can skip extraction when the number is missing. - C. Add a few-shot example with a correctly formatted purchase-order number.
- D. Make the
po_numberfield nullable/optional and instruct the model to return null when the value is not present in the source.
Correct: D. The model fabricates because a required field gives it no sanctioned way to say "absent". A nullable field plus an instruction to return null removes the cause. A confidence score only filters the symptoms, and a few-shot example reinforces the format without solving the missing-data case.
QUESTION · D5 · CONTEXT MANAGEMENT
Midway through a coherent multi-hour refactor, the session approaches its context limit. The history is full of superseded intermediate diffs, but the task and the goal are unchanged and the latest file states are still valid. What should you do to keep going without losing the thread?
- A. Start a brand-new session with no prior context and re-describe the task.
- B. Fork the session to branch off a divergent line of work.
- C. Request a larger context window and continue as-is.
- D. Run
/compactto condense the accumulated history into a summary and continue the same task.
Correct: D. There is one thread, and the only dead weight is the history of intermediate changes. /compact removes it and preserves the goal. A fresh session would discard still-valid state, forking is for divergent branches, and a larger window only postpones the problem.
Hesitated somewhere between two answers? Exactly. That's the point. That kind of hesitation wore me down for the full two hours of the real exam.
Study mode and exam mode
The app has two modes, because you study differently than you test yourself. In study mode you pick an answer, can still change your mind, and once you hit “Check answer” you get the full picture: why that answer wins and what makes the other three tempting. The clock stops while you read the feedback. Nothing rushes you. Exam mode recreates the real CCAR-F: 120 minutes on the clock, feedback only after you hand in the sheet, and until the very end you can flag questions and change answers.
When you finish, you get a score on Anthropic's scale: 100 to 1000, with a 720 pass line. You also see how you did in each of the five domains. You know right away whether MCP or context is your weak spot. At the end you review every question with the answers and the reasoning. You can filter down to the incorrect and the flagged ones.
How I would use it
If I were starting from zero, here is what I would do:
PREP PLAN
- 1. First pass in study mode. No time pressure, question by question. Read every explanation, even when you got it right. The goal is understanding why the distractors are wrong.
- 2. A full sitting in exam mode. One session, 120 minutes, no breaks. You're simulating the conditions, not just checking knowledge.
- 3. Study the domain breakdown. The per-domain score tells you where to read up. Go back to the docs for your two weakest areas.
- 4. Treat 720 on the mock as the floor, not the target. The real exam plays a class above sample material. Aim for a score with a solid margin.
One more piece of advice beyond the app. Read my 12 rules for building agents. I wrote them down while studying for this same exam. They cover a good chunk of D1 and D2.
Important disclaimer. This is unofficial study material. It is not affiliated with, authorized or endorsed by Anthropic. I wrote all 60 questions myself, modeled solely on the publicly available exam guide. Claude and Claude Certified Architect are names owned by Anthropic.
These questions carried me to 833 out of 1000. You can beat that score. When you pass, write to me and show off how you did in each of the five domains.
Ready for 60 questions?
Free, in the browser, no signup. Progress is saved locally.
