Home / Claude Certified Architect Foundations

Anthropic · CCAR-F

Claude Certified Architect Foundations Practice Exam (CCAR-F)

Claude Certified Architect Foundations practice exam with answers from all 5 official domains, with every answer option explained.

Start practicing free Take a mock examSee plans

Claude Certified Architect – Foundations (CCAR-F)

Questions
60

Multiple-choice and multiple-response.

Time
120 min
To pass
720 / 1000

Scored on a 100–1000 scale.

Exam Price
$125

What is on the CCAR-F exam?

Open a domain for its objectives and their practice questions.

Try free CCAR-F practice questions

Pick an answer to see why each option is right or wrong. A preview of 4 of the 10 free questions, which cover every domain.

Domain 1: Agentic Architecture & Orchestration

1.2 Orchestrate multi-agent systems with coordinator-subagent patternseasy

An architect is deciding where a lead agent with parallel subagents is worth its extra token cost. Which workload fits that architecture best?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Multi-agent runs use far more tokens than chat, so they only make sense when the task's value covers that cost.

Multi-agent systems pay off on valuable, heavily parallel work whose information exceeds a single context window: subagents explore separate directions at once, each in its own context.

Parallel subagents are overkill for a simple lookup. One agent with a few tool calls covers it.

Code changes look like divisible work, which makes this tempting. But tightly coupled steps need shared context, and agents are not yet good at coordinating in real time, so this is named as a poor fit.

1.1 Design and implement agentic loops for autonomous task executioneasy

Claude asks for two currency conversions in one response. Your code runs both and returns two tool_result blocks, but Claude's final answer swaps the two figures. What should you check first?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

A good description helps Claude decide when and how to call a tool, so the advice is real. Here Claude already called the right tool twice; the failure is in pairing results with calls.

Order is not the matching mechanism. The IDs let Claude pair results with requests even when results arrive in a different order.

Every tool call gets a unique ID, and each tool_result must echo the ID of the call it answers. That pairing is how Claude knows which output belongs to which request, so a mismatch produces exactly this swap.

Claude can legitimately request several tools in one response, and all their results go back together. Splitting them is unnecessary; matching IDs is what keeps them apart.

Domain 2: Tool Design & MCP Integration

2.1 Design effective tool interfaces with clear descriptions and boundarieseasy

An agent loads tools from a ticketing service and a wiki service. Both services expose a tool called 'search', and Claude sometimes queries the wrong system. Which naming change does the guidance recommend?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Version labels describe history, not purpose. Claude still cannot tell which tool reaches tickets and which reaches the wiki.

Every tool definition needs a name, and list order carries no meaning about what a tool does. Selection would get worse, not better.

This removes the name clash but not the ambiguity: a random suffix tells Claude nothing about which system each tool searches.

When tools span several services, the guidance recommends namespacing: prefix each name with its service. Two tools with the same vague name are exactly the overlap that confuses agents, and the prefix makes the boundary visible at selection time.

2.2 Implement structured error responses for MCP toolseasy

Why does the course recommend that tool functions validate their inputs and raise errors with clear messages, such as 'start_date is required'?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

A tool error does not end the conversation. It is reported back so Claude can continue, often by retrying.

The point of the recommendation is the opposite: Claude sees the error and learns from it.

Error messages from a tool reach Claude. A clear one tells it what was wrong, so it can call the tool again with a fixed value instead of guessing.

A failed call still gets a result block; validation improves what that block says, it does not remove it.

The other 6 free questions, by domain:

Domain 3: Claude Code Configuration & Workflows · 2 questions
3.1 Configure CLAUDE.md files with appropriate hierarchy, scoping, and modular organizationeasy

A developer starts Claude Code at the root of a monorepo. packages/billing/CLAUDE.md holds billing conventions, but right after launch /context does not list it under Memory files. What explains this?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

That variable is real, but it governs directories added with --add-dir, which sit outside the working directory. A package inside the repository does not need it.

CLAUDE.md files do not replace one another; discovered files are concatenated. A root file could not suppress the billing file.

CLAUDE.local.md is the personal, gitignored variant. Its name has nothing to do with whether a nested file is discovered.

Only files at and above the launch directory load at startup. A CLAUDE.md below it joins the conversation once Claude reads, writes or edits a file in that subdirectory, so its absence from the launch list is expected.

3.2 Create and configure custom slash commands and skillseasy

A codebase-analysis skill prints hundreds of lines of intermediate findings into the main conversation every time it runs. Which frontmatter setting keeps that work out of the main session and returns only the result?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

With context: fork the skill runs in its own subagent, which does the noisy work in a separate context and hands back its result, so the main conversation stays clean.

This field is useful when you want to control when a skill runs, because it stops Claude from invoking it on its own. It does not change where the skill's output goes.

An argument hint only shows expected arguments during autocomplete. The output still lands in the main conversation.

This picks the shell for injected commands. It has no effect on how much output reaches the conversation.

Domain 4: Prompt Engineering & Structured Output · 2 questions
4.2 Apply few-shot prompting to improve output consistency and qualityeasy

A prompt pastes four sample ticket-and-answer pairs directly between the task instructions and the live ticket. Claude sometimes treats a sample ticket as part of the request it must answer. What fix does the guidance recommend?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Relocating unlabeled samples does not mark where each one starts and ends, so Claude still has nothing that separates them from the real request.

Making samples look like instructions blurs the boundary even more; the goal is to make them clearly distinct.

Dropping the examples throws away a reliable way to steer format. The problem is the missing boundaries, not the examples themselves.

Structured examples, each in its own example tag and grouped in an examples tag, let Claude tell samples apart from instructions and from the live input, which is exactly the confusion described.

4.1 Design prompts with explicit criteria to improve precision and reduce false positiveseasy

After enabling Code Review with default settings on a service repository, a lead notices that no findings mention missing unit tests or inconsistent formatting, even on PRs that clearly have both. What explains this?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

The trigger setting decides when a review runs and how much it costs, not what kinds of issues the review looks for.

The check run repeats the same findings as the inline comments in a severity table; it is not a separate home for style or test findings. If the review never produced them, they appear nowhere.

Out of the box, Code Review is calibrated for correctness: bugs that would break production. Formatting preferences and test coverage are outside that default, and a team that wants them flagged has to add guidance files.

Replying to an inline comment does not prompt Claude to respond or re-review, so replies cannot unlock new categories of findings.

Domain 5: Context Management & Reliability · 2 questions
5.1 Manage conversation context to preserve critical information across long interactionseasy

Hours into a session, an agent's history holds dozens of raw tool results from calls made early on that it no longer needs to reread. The team wants the least disruptive, low-risk way to reclaim context. What fits best?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Moving the same tokens elsewhere in the request does not reduce context. The results would still be read on every turn.

Once a tool result is far back in the history the agent rarely needs the raw output again, which makes clearing it one of the safest, lightest forms of compaction.

Expanding stale results adds tokens rather than freeing them, the opposite of what the team needs.

A blank restart frees space but throws away the decisions and progress in the history too. That is far more disruptive than removing only the stale tool output.

5.2 Design effective escalation and ambiguity resolution patternseasy

A team is building an agent that refactors a service across many files over several hours. The product owner wants to review direction at a few planned points without approving every individual action. Which design fits the agent pattern?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Checkpoints give the owner a say at the points they care about while the agent keeps operating independently between them, which is exactly what the pattern allows.

This keeps autonomy but drops the planned review points the owner explicitly wants during the run.

Per-call approval is the opposite of what the owner asked for and removes the agent's independent operation.

Iteration caps are a real control, but they end the run at an arbitrary count rather than at the review points the owner chose.

Practice all 10 free questionsGet 150+ questions

CCAR-F question bank coverage

DomainExam weightFree sampleMaster (full bank)
1. Agentic Architecture & Orchestration27%249
2. Tool Design & MCP Integration18%232
3. Claude Code Configuration & Workflows20%236
4. Prompt Engineering & Structured Output20%236
5. Context Management & Reliability15%227
Total100%10180

Looking for a Claude Certified Architect – Foundations practice exam? Start with the free questions above, then take the timed mock exam to rehearse the CCAR-F format under exam conditions.

How these CCAR-F questions are made

The domains, weights and objectives come from Anthropic's official CCAR-F exam guide (v1.0). Questions are original and written only from the vendor's official prep material, and each one cites the passage it is based on. Every question is then answered independently, without seeing its key, and every answer option's explanation is checked for accuracy before it is published.