Claude Certified Architect practice exam (CCAR-F Foundations)

Question 1

Agentic Architecture & Orchestration

Northline Ferries has a customer assistant that calls the Claude Messages API with a tool that looks up ferry schedules; the application runs the tool whenever Claude asks for it. In some conversations the application stops right after the first tool call and shows the customer nothing. In others, Claude asks for a schedule lookup that has already run successfully. The tool never fails, and every response finishes normally. Which two changes let the application keep going until Claude gives its final answer, using the lookups already done?

Choose two.

• A. End the loop when the schedule tool returns without an error, and show its result to the customer.
• B. Check stop_reason on each response: run the requested tools on tool_use, and finish on end_turn.
• C. Keep Claude's tool_use message in the history and add a tool_result for each call, matched by tool_use_id.
• D. Replace each earlier tool call and its result with a short prose summary before sending the next request.

Question 2

Claude Code Configuration & Workflows

Clearwater Utilities keeps its code-review conventions for Claude Code in one senior engineer's personal instruction file. A new engineer has joined, and Claude Code does not apply the conventions in her sessions. The team wants every contributor to get the same conventions automatically, with changes tracked in the repository. Where should the conventions go?

• A. In each engineer's user-level ~/.claude/CLAUDE.md file, copied from the original.
• B. In a team wiki page that each engineer pastes into Claude Code at the start of a session.
• C. In the project's .claude/CLAUDE.md file, committed to the repository.
• D. In an --append-system-prompt flag that each engineer adds to a personal shell alias.

Question 3

Prompt Engineering & Structured Output

Rivermark Council uses Claude to review building-permit applications and flag clauses that create legal or safety risk. The prompt says 'flag anything risky', and staff now spend most of their time dismissing flags on standard boilerplate, such as the council's own indemnity wording. The council has a list of risk categories it cares about and a list of routine clauses that never need review. How should the team change the prompt?

• A. List the risk categories to report and the routine clauses to skip, with an example of each.
• B. Tell Claude to flag a clause only when it is highly confident that the clause is risky.
• C. Explain how the council's permit process works and who reads the flags, and keep the instruction.
• D. Ask Claude to score each clause's risk from 1 to 10 and report only scores of 7 or higher.

Question 4

Tool Design & MCP Integration

Heritage Seed Bank gives its research agent two tools, search_catalogue and lookup_catalogue, and both have the one-line description 'Searches the seed catalogue'. In fact, one takes a species name and returns a list of matches, and the other takes an accession number and returns one full record. Analysts see the agent pick either tool at random for the same kind of question. What should the tool owners change first?

• A. Add a system-prompt line telling the agent to prefer search_catalogue unless told otherwise.
• B. Rewrite each description to say what input it takes, what it returns and when to use it.
• C. List search_catalogue first in the tools array, since it answers most analyst questions.
• D. Add optional filter fields to both input schemas so that either tool can answer any query.

Question 5

Context Management & Reliability

Harbour Mutual's claims assistant takes insurance claims in long chat sessions. Customers often correct the claimed amount partway through, and to save space the application summarises older messages. Near the end of long chats, the assistant sometimes quotes an old amount or none at all. Which design keeps the latest confirmed amount available for every later decision?

• A. Store the confirmed amount in application state, update it on each correction, and send it with every request.
• B. Pin the customer's first message, which states the claimed amount, and summarise all messages after it.
• C. Ask Claude to carry the amount forward in each summary, and keep only the summaries as the record.
• D. Keep more recent messages in full before summarising, so that corrections stay in the window longer.

Question 6

Agentic Architecture & Orchestration

Prairie Grain's inventory assistant uses the Claude Messages API. When Claude asks for the warehouse stock tool, the application runs it and then sends the next request with the conversation so far. Claude's reply repeats its earlier recommendation as if the stock figures never arrived, even though the tool returned correct data. What must the application do before sending the next request?

• A. Resend the user's original question on its own, so that Claude starts again with current stock.
• B. Ask the tool to return its figures as a plain-language sentence instead of as JSON.
• C. Add the stock figures to the history as a tool_result matched to the call's tool_use_id.
• D. Start a new conversation for each stock lookup and pass Claude only the lookup result.

Question 7

Claude Code Configuration & Workflows

Crestline Hotels' project CLAUDE.md has grown past 900 lines, covering testing, security, API design and deployment. Each topic has a different owner, and their edits keep colliding. The team wants each owner to maintain their standards in a separate version-controlled file, and Claude Code must still load those standards during project work. How should the team organise the instructions?

• A. Give each service directory its own copy of the full standards file, and review the copies every quarter.
• B. Move each topic to its own file, imported with a line such as @docs/testing.md or kept in .claude/rules/.
• C. Move each topic to its own page in the project wiki, and link each page from the project CLAUDE.md.
• D. Keep one CLAUDE.md split into owned sections, and have each owner approve changes to their section.

Question 8

Prompt Engineering & Structured Output

Orchid Dental Labs uses Claude to check incoming lab orders and label each problem as low, medium or high severity. The same order can get 'medium' on one run and 'high' on the next, and the technicians who act on the labels disagree about what each level means. The lab wants each level to mean the same thing on every run and to every technician. What change meets this requirement?

• A. Define what each severity level means for an order, with one real order problem as an example of each.
• B. Tell Claude to report a problem only when it is highly confident of the severity it has assigned.
• C. Send more orders in each request so that Claude can grade problems relative to each other.
• D. Ask Claude to explain its reasoning before each label so that technicians can see why it chose it.

Question 9

Tool Design & MCP Integration

Seaway Freight has one document tool that extracts fields, classifies documents and compares two documents. Its input schema is a single flat list of many optional fields, and Claude often sends requests that mix fields from different operations, such as a comparison target on an extraction request. Each operation needs different inputs and returns a different result. The team wants the tool definitions themselves to show those differences. Which redesign best meets that requirement?

• A. Keep the flat schema and add a worked example of each operation to the tool description.
• B. Keep the flat schema and add a required operation field that names the chosen operation.
• C. Keep the optional fields and have the tool reject invalid combinations with an error message.
• D. Replace it with three tools, each with its own required inputs, result format and description.

Question 10

Context Management & Reliability

Alpine Resorts' maintenance assistant calls an asset lookup tool that returns more than 40 fields for each ski-lift component. For the repair decisions the assistant makes, only the fault code and the last inspection date matter. After a few lookups the context is full of unused fields, and answers become slower and less focused. What should the team do?

• A. Switch to a model with a larger context window so that all 40 fields fit for longer sessions.
• B. Have the application keep only the fault code and inspection date before adding each result.
• C. Tell Claude in the system prompt to disregard every field except the fault code and the date.
• D. Summarise the conversation after every five lookups to remove the fields that were not used.

Question 11

Agentic Architecture & Orchestration

The Lantern Museum's archive research system has a coordinator agent and specialist subagents for letters, photographs and exhibition records. Specialists have started passing results straight to each other, and the final report sometimes leaves out findings the coordinator never saw. The team wants a single component to hand out every assignment and receive every finding before the report is written. Which communication pattern fits?

• A. Each specialist passes its findings to the next one in a fixed order, and the last one writes the report.
• B. Specialists read and write a shared memory store so that each can pick up the others' work.
• C. After each result, the specialist that produced it chooses which specialist should work next.
• D. Specialists return findings only to the coordinator, which assigns the next task and passes on context.

Question 12

Claude Code Configuration & Workflows

Morrow Bakeries wants a /release-notes command in Claude Code that drafts release notes from recent commits. Every contributor should get the command when they clone the repository, and changes to it should go through code review. Where should the command definition live?

• A. As a Markdown file in the project's .claude/commands/ directory.
• B. As a Markdown file in each contributor's ~/.claude/commands/ directory.
• C. As a paragraph in the project CLAUDE.md that explains how to write release notes.
• D. As a saved prompt in the team's snippet tool, pasted in before each release.

Question 13

Prompt Engineering & Structured Output

LoomWorks Textile Mill uses Claude to classify fabric defect reports and return a JSON object with a defect type and severity. Despite two pages of detailed prose instructions, the output changes shape between reports: fields are renamed, nested differently or added. The team has several past reports with correct, reviewed classifications. What should it add to the prompt?

• A. A more detailed explanation of each output field, with its name, type and allowed values.
• B. Two to four of the reviewed reports, each shown with its correct output in the required shape.
• C. A closing instruction to follow the output format exactly and to add no extra fields.
• D. Twenty recent defect reports without classifications, so that Claude sees the range of inputs.

Question 14

Tool Design & MCP Integration

Valley Voice Radio is adding an archive_search tool to its producer assistant, alongside tools for the music library and the advertising schedule. A developer suggests keeping the tool description to a few words, because the input schema already defines every parameter. Why is the description worth writing carefully?

• A. It sets the order in which Claude tries the tools when more than one of them could apply.
• B. It tells the API how to validate the arguments that Claude sends before the tool runs.
• C. It controls how the tool's result is formatted before the result is returned to Claude.
• D. Claude reads it to decide whether this tool fits the current request and when to use it.

Question 15

Context Management & Reliability

Pawswell Veterinary's support assistant answers pet owners' questions about appointments and bills. A customer writes: 'I don't want to do this with a bot. Please put me through to a person.' The request is clear, and nothing about the customer's issue has to be checked first. What should the assistant do?

• A. Look up the relevant policy first, so that it can offer an answer before deciding on a handoff.
• B. Ask the customer to describe the issue in more detail, so that the handoff note is complete.
• C. Hand the conversation to a staff member straight away, as the customer has asked.
• D. Try once to resolve the issue, and hand off if the customer is still not satisfied.

Question 16

Agentic Architecture & Orchestration

Harbourgate Port Authority uses a coordinator agent to produce a report on a planned container terminal. Specialist subagents returned well-sourced findings, but the draft covers only four of the seven topics in the commissioning brief. The budget allows two more subagent assignments. Which two coordinator actions make the best use of that budget?

Choose two.

• A. Compare the brief's topics with the draft and its findings to rank the topics that are still missing.
• B. Ask two specialists to rerun their original searches with a higher limit on the number of results.
• C. Have the synthesis agent expand each thin section of the draft using the evidence already collected.
• D. Send focused follow-up research to the top missing topics, then revise the report with the findings.

Question 17

Claude Code Configuration & Workflows

At Stonebridge Brewery, developers use a Claude Code skill for occasional deep investigations of the codebase. The skill's many search results fill the main conversation, although the developer needs only a short report of the findings. The same step-by-step investigation procedure is also copied into CLAUDE.md, so it loads during unrelated work. Which two changes reduce both sources of context use?

Choose two.

• A. Set context: fork on the skill, and have it finish with a short findings report that cites files.
• B. Keep the investigation procedure only in the skill body, and delete its copy from CLAUDE.md.
• C. Pre-approve only Grep and Read in the skill's allowed-tools, and keep it in the main conversation.
• D. Keep the skill in the main conversation, and tell it to leave search results out of its answer.

Question 18

Prompt Engineering & Structured Output

Northbridge Admissions uses Claude to extract applicant details from school transcripts. The prompt has two examples, both from the two most common transcript layouts, and extraction on those layouts is accurate. On a third layout, which puts grades in a side table, Claude misses fields. The team wants Claude to handle layouts it has not seen yet, not only this third one. Which change to the examples is best?

• A. Swap in examples that cover clearly different layouts, including the one with a side table.
• B. Add a prompt rule for the side-table layout, and add another rule for each new layout found.
• C. Add more examples of the most common layout so that Claude learns each field more thoroughly.
• D. Remove the examples and describe each field's meaning in more detail in the output schema.

Question 19

Tool Design & MCP Integration

Pinewatch Forestry's MCP server has a permit lookup tool. It returns the same text, 'Request failed', when the arguments are invalid, when the permit database is briefly down, and when the lookup is refused for lack of access. The client agent needs to decide whether to retry, fix its request or stop. What should the server return for these failures?

• A. An isError result whose content gives a failure category and says whether a retry can help.
• B. An error message with the full stack trace and the database's own error text for each failure.
• C. An empty result for any lookup that cannot be completed, whatever the reason for it.
• D. An error that tells the client to retry up to three times before it reports the failure.

Question 20

Context Management & Reliability

A Pawswell Veterinary customer asks for a refund after cancelling a pet's boarding stay two days before it began. The assistant's policy documents cover refunds for grooming and vet visits but say nothing about boarding. The assistant must not make up policy. What should it do?

• A. Apply the grooming cancellation rule, since it is the closest written policy to boarding.
• B. Pass the request to a staff member and explain that no written policy covers boarding refunds.
• C. Offer a partial refund as a goodwill gesture, since the customer cancelled with some notice.
• D. Tell the customer that no refund is available, because no written policy allows one for boarding.