How to Pass the CCAR-F Exam: 13 Wrong-Answer Patterns I Found Building 360 Practice Questions

A practical guide to the Claude Certified Architect – Foundations exam: the format, the five domains, the six scenarios and the traps that cost the most points.

Over the last few weeks I built 360 practice questions for the Claude Certified Architect – Foundations exam, which Anthropic codes CCAR-F (many people still write CCA-F). Every question had to map to a task statement in the official exam guide, and every wrong option had to be tempting enough to fool someone who almost knows the answer.

Designing wrong answers teaches you something that reading the guide doesn't. The exam keeps coming back to a small set of traps, and once you can name them, a lot of "two plausible options" questions get easier.

I build practice material for AI certifications, and this project started with a gap. The biggest CCAR-F practice set I found used only single-answer questions, while the real exam puts every question inside a production scenario and mixes in questions with more than one correct answer. So I built a set that follows the exam's actual shape, and I checked every answer against the official guide and the current Claude documentation.

Below you'll find what the exam looks like, what each domain really tests, the 13 wrong-answer patterns that kept coming back, one question solved step by step, and a four-week plan. Every exam fact here comes from the public exam guide (version 1.0, July 2026) and the official certification page.

What the CCAR-F exam looks like

• 60 questions in 120 minutes, so about two minutes each. Most questions open with a paragraph of production context, so reading speed matters.
• Two question types. Multiple choice with one correct answer, and multiple response. A multiple-response question always tells you how many answers to pick ("Select TWO").
• Every question belongs to a scenario. Each exam draws 4 scenarios at random from a bank of 6.
• Scaled scoring from 100 to 1,000, with 720 to pass. There is no penalty for guessing, so never leave a question blank.
• The score report shows pass or fail, your scaled score and your percentage in each domain. If you have to retake, that breakdown tells you where to work.
• Logistics. The fee is $125. The exam is delivered by Pearson VUE, online or in a test center, in English or Japanese. The credential is valid for 12 months, and renewal is a free, non-proctored assessment.
• Retakes open after 14 days, then 30, then 90, with at most four attempts in 12 months.

Registration goes through the official CCAR-F certification page, where you'll also find the exam guide and the free preparation courses. Check who can register there before you plan a date. Scheduling itself happens on Pearson VUE's Anthropic page.

The five CCAR-F domains, and what each one really tests

The weights tell you where the points are. The domain names, though, hide what the questions actually ask.

1. Agentic Architecture & Orchestration (27%, about 16 questions). How an agent loop runs and stops, how a coordinator hands work to subagents and what context they need, and when a rule has to be enforced in code rather than in the prompt. If you only have time to go deep on one domain, make it this one.
2. Tool Design & MCP Integration (18%, about 11 questions). Tool descriptions, structured errors, how many tools a single agent should get, and where MCP servers are configured. A surprising number of questions come down to "the tool description is the real problem."
3. Claude Code Configuration & Workflows (20%, about 12 questions). Where configuration lives (the CLAUDE.md hierarchy, .claude/rules/ files with path globs, commands and skills), plan mode versus direct execution, and running Claude Code in CI with -p.
4. Prompt Engineering & Structured Output (20%, about 12 questions). Explicit review criteria, few-shot examples, JSON schemas through tool use, validation and retry loops, and when the Message Batches API is the right call.
5. Context Management & Reliability (15%, about 9 questions). What gets lost in long sessions, when to escalate to a human, how errors should travel between agents, and how to keep sources attached to claims.

The domain I see people underrate is the last one. It has the fewest questions of its own, but its ideas (case facts, escalation triggers, provenance) also show up inside scenario questions from the other domains.

The guide also lists what is out of scope, which saves real study time: fine-tuning, computer use, vision, streaming, rate limits and pricing calculations, model benchmarking and comparisons, embeddings and vector databases, and prompt caching beyond knowing that it exists.

The six scenarios

Every question sits inside one of these settings, and each one has favorite topics:

• Customer Support Resolution Agent: loop control, refund limits enforced in code, structured tool errors, escalation rules and handoff summaries.
• Code Generation with Claude Code: the CLAUDE.md hierarchy, path-scoped rules, team commands and skills, plan mode.
• Multi-Agent Research System: how the coordinator splits work, spawning subagents in parallel, passing context explicitly, keeping sources through synthesis.
• Developer Productivity with Claude: built-in tools (Grep, Glob, Read, Edit), MCP servers and resources, exploring a large codebase without flooding the context.
• Claude Code for Continuous Integration: non-interactive runs, JSON output with a schema, review precision and independent reviewers.
• Structured Data Extraction: tool-use schemas, nullable fields, validation and retry, batch processing and human review.

The 13 wrong-answer patterns

After a few hundred questions, the same kinds of wrong answers kept coming back. Each one is wrong for a reason the exam cares about, and most of them would be right in some other situation, which is exactly why they are tempting. Learn to spot them and you can often cross out two options before you finish reading.

Image generated with Gemini

1. A bigger model or a bigger context window. It rarely fixes attention, tool selection or missing information. It's only the answer when the question is about capacity itself.
2. Saying it louder. Capital letters, repeating the rule, "be very careful." A prompt is still probabilistic. When a rule must always hold, prompt-only options are out.
3. Confidence or sentiment as a trigger. Self-rated confidence isn't calibrated, and an angry customer is not, on its own, a reason to escalate. Calibrated per-field confidence used to route documents to human review is a different idea, and a valid one.
4. A keyword router or a classifier. Usually far too much for a question that asks for the "first step," and it bypasses what the model already understands.
5. Merging everything. Combining tools or agents tends to hide the real problem, which is often a vague tool description or a missing scope.
6. Fixing it after the fact. Audits, alerts, weekly reviews, reversals. If the scenario needs prevention, detection is the wrong layer.
7. Blind retries. The same prompt again, a higher temperature, retrying a permanent error, or retrying when the information isn't in the source at all.
8. Hiding the failure. Returning an empty result as a success, filling in default values, quietly dropping the items that failed.
9. Shutting everything down. Aborting a whole workflow because one step failed, or removing a tool that is useful most of the time.
10. Pushing the work onto people. Asking customers or developers to do what the system should do for them.
11. Parsing the model's text. Ending a loop or routing a request based on words in the reply instead of stop_reason or structured output.
12. Voting. Running the same review three times and keeping only what the runs agree on throws away real issues that appear intermittently.
13. Features that don't exist. Made-up flags, environment variables or config keys. If an option names something you've never seen in the docs, be suspicious.

One question, solved step by step

Here is a question from the free readiness check I created for this exam. It is my own question, not one from the real exam.

A platform team uses the Claude Agent SDK to build a deployment agent that can run shell commands. Company policy says the agent must never run "terraform apply" against the production workspace unless the request carries an approved change ticket. In 2% of test runs the agent ignored this rule, even though the system prompt states it. Which TWO changes enforce the policy? (Select TWO.)

A. Repeat the rule at the end of the system prompt in capital letters, so it is the last thing Claude reads

B. Add a PreToolUse hook that denies the shell call when it targets production without an approved ticket

C. Add few-shot examples in which the agent asks for an approved ticket before it applies any change

D. Post an alert to the on-call channel whenever a production apply runs without a ticket ID

E. Replace the generic shell tool with a deploy tool whose code checks the ticket status before applying

The word "never" does most of the work. A rule that must always hold rules out anything probabilistic, so A goes first (pattern 2). C goes too. Few-shot examples make the behavior more consistent, and the 2% failure rate shows that consistent isn't the same as guaranteed. D is pattern 6, since the alert fires after production has already changed.

That leaves B and E, and they act at two different layers. The hook blocks the call before it runs, and the narrower tool removes the unsafe path entirely. On multiple-response questions, two correct answers that cover two complementary layers are a common shape.

A five-step method for every question

1. Name the symptom in a few words: wrong tool selected, invented data, lost facts, duplicated work, a loop that doesn't stop.
2. Find the layer that is actually broken: the prompt, a tool description, the tool interface, the schema, the loop code, a hook, the orchestration, where a configuration lives, or how context is managed.
3. Look for words that demand a guarantee, such as "must," "never" or "always," or stakes like money, identity and compliance. If you find one, keep only the deterministic options.
4. Cross out the 13 patterns above.
5. Pick the smallest change that fixes the cause at its own layer. On multiple-response questions, judge each option on its own.

One more habit: read the question sentence before the scenario. "Most effective first step," "most likely cause" and "which TWO" each rule out options that would be correct for a different question.

How I would prepare in four weeks

This plan assumes a few hours a week and some hands-on time with Claude.

• Week 1: the source. Read the official exam guide twice, including its sample questions. Then take Anthropic's free Academy courses that match the domains: Building with the Claude API, Introduction to Model Context Protocol and Claude Code in Action.
• Week 2: build small things. An agent loop that stops on stop_reason, an MCP tool that returns structured errors, a Claude Code project with a path-scoped rule and a PreToolUse hook. Ten minutes of building makes a configuration question obvious.
• Week 3: timed practice. Do a full set of 60 questions in 120 minutes, then read every explanation, including the ones you got right. Keep a list of the patterns that caught you.
• Week 4: close the gaps. Rework your two weakest domains, do one more timed set, then rest. On exam day, aim for about 90 seconds per question on the first pass, flag anything you're unsure about, and keep 15 to 20 minutes to come back to it.

Two details caught me out while I checked answers against the current Claude documentation. The exam guide calls the subagent tool "Task," while the current Claude Code and Agent SDK docs call it "Agent." And some models reject the forced tool_choice setting that the guide presents as standard API behavior. The exam follows the guide, so answer in its terms, and don't let the newer docs make you doubt a correct answer.

A timing note: Anthropic has announced updated content for the Associate, Developer and Architect – Professional exams from 7 December 2026. Foundations isn't in that list on the official page, but check it before you book.

CCAR-F exam FAQ

How many questions are on the CCAR-F exam?

60 questions in 120 minutes. Each one is either multiple choice with one answer or multiple response, and multiple-response questions tell you how many answers to select.

What is the passing score?

720 on a scaled score from 100 to 1,000. The scoring is criterion-referenced, so your result doesn't depend on how other candidates did.

Is CCA-F the same exam as CCAR-F?

Yes. CCAR-F is the official exam code, and many people still write CCA-F.

Are computer use or prompt caching on the exam?

No. The exam guide lists computer use as out of scope, and prompt caching beyond knowing that it exists.

How much does it cost, and how long is it valid?

The fee is $125 and the credential is valid for 12 months. Renewal is a free, non-proctored assessment.

Is the CCAR-F exam hard?

Most questions ask you to choose the right fix for a production problem, so hands-on experience helps more than memorized API details. The difficulty comes from the time limit and from options that are almost right, which is what the 13 patterns are for.

Free study guide and practice exams

I turned these notes into a free 17-page pocket guide: the 10-question readiness check (including the question above) with answers, one line for each of the 30 task statements, decision tables, drills and an exam-day plan. Download the CCAR-F Pocket Guide for free.

For full timed practice, I also published six 60-question practice exams on Udemy. They follow the official blueprint, include multiple-response questions, and get harder from the first exam to the sixth: CCAR-F practice exams on Udemy. Disclosure: both are mine.

Good luck with your preparation. If something in your own exam surprised you, tell me in the comments, and I'll add it to the next version of this guide.

This article is independent and not affiliated with or endorsed by Anthropic. Exam details come from the public exam guide (version 1.0, July 2026) and the official certification page, checked in October 2026. Confirm them on the official page before you book.

I drafted this article with the help of an AI assistant, then checked every exam fact against the official guide and edited it myself.

Originally published on Medium.