Skip to content

CCAR-F / Free practice questions

20 free CCAR-F practice questions, explained

A fixed sample from the Claude Certified Architect — Foundations (CCAR-F) course, spread across 5 domains. Choose an answer, check it, and read the full explanation. Your score stays in this page and nothing is saved or submitted.

  • AI & Anthropic Fundamentals
  • Claude Capabilities
  • Context Management
  • Prompt Engineering
  • Tools & Development
CCAR-F free practiceScore 0 / 0 answered · 20 questions

Question 1

AI & Anthropic Fundamentals
You are building a customer-support agent using the Claude Messages API with client-executed tools. After Claude requests a diagnose_issue tool call, the application executes it. How should the application continue so Claude can use the result when deciding whether to request apply_fix or escalate?

Show the answer and explanation

Correct answer: B. Preserve the assistant's tool_use message, return the matching tool_result in a user message, and send the next request; continue client-tool handling when requested and handle the eventual stop reason explicitly.

For client-executed tools in the Messages API, the application runs each requested operation and returns its output in a tool_result block associated with that tool_use ID. The next request preserves the assistant's tool-use message and follows it with the user message containing the results. Claude can then reason about those results. Continue when the response requests another client tool; an end_turn response indicates the model has finished its turn, while other stop reasons need their own handling. Keyword matching and a fixed action sequence do not implement this protocol, and an iteration cap alone does not establish successful completion. This manual client-tool loop is distinct from the higher-level Agent SDK, which supplies an agent loop and tool execution.

Open this question on its own page

Question 2

Claude Capabilities
A generated schema compiles, but expected foreign-key relationships are missing. Those relationships are explicitly listed in the source specification. What should the acceptance workflow compare?

Show the answer and explanation

Correct answer: B. The required relationships from the specification against the actual generated schema, rejecting unexplained omissions.

Check the required relationships in the artifact itself. Examples may help generation, but an output’s fluent explanation cannot replace that comparison.

Question 3

Context Management
Every new run needs a specified version of an API contract. The launcher currently asks the model to remember to load it, and some runs omit it. Which change makes startup availability verifiable?

Show the answer and explanation

Correct answer: B. Have the launcher load and verify the required contract version before starting the task.

The launcher can verify that the intended input was supplied. That proves availability, not that every later model judgment applies the contract correctly.

Question 4

Prompt Engineering
A completed extraction batch has 9,950 accepted results and 50 requests rejected for a repairable input-format error. The source versions are unchanged, and there is no urgent deadline. After fixing the input format, what avoids repeating the accepted work?

Show the answer and explanation

Correct answer: B. Submit the 50 failed requests and retain their document/version mapping across attempts.

Targeted retries retain the accepted work. Each new attempt still needs its own outcome handling and correlation with the original source.

Question 5

Tools & Development
A documentation search backend accepts phrases, code snippets and error codes, but its query parameter says only keyword. The model often truncates useful queries. Which parameter-level change should the team evaluate?

Show the answer and explanation

Correct answer: B. Describe the accepted query forms and add representative inputs to the tool definition.

Put the supported forms where the caller composes the query: the parameter description and valid input examples. Include a phrase, a snippet and an error code. Evaluate query retention and retrieval quality; accurately advertising accepted inputs does not establish that every query retrieves the right evidence.

Question 6

AI & Anthropic Fundamentals
A research application uses the Claude Messages API with a client-executed search tool. Claude has returned a tool_use block. The search produces a relevant JSON result that fits the current context budget. How should the application provide that result for Claude's next turn?

Show the answer and explanation

Correct answer: B. Preserve the assistant's tool_use message and follow it with a user message containing a tool_result block whose tool_use_id matches the request and whose content includes the relevant search result.

In a Messages API client-tool exchange, the application supplies the operation's result through a tool_result block linked to the preceding tool_use ID. Preserving that exchange lets Claude use the returned evidence; an inaccessible record ID or a generic success statement does not provide the result. This is the Claude API message format, not a universal MCP transport requirement. MCP tool results can contain content, structuredContent, or resource links that the host integrates as appropriate. Neither protocol requires every large raw payload to be inserted unaltered into every model request. Applications should supply the relevant evidence within their context budget and preserve a way to retrieve additional detail when needed.

Open this question on its own page

Question 7

Claude Capabilities
An organization requires human approval before a migration reaches production. Approval must cover the actual SQL and target database, and later edits must invalidate it. Which design enforces that requirement at execution?

Show the answer and explanation

Correct answer: C. Require the execution service to check approval bound to this SQL digest and target.

Authorization must match the operation being executed. Binding it to the artifact and target prevents a changed request from reusing approval for a different action.

Question 8

Context Management
A customer requests human support and automated resolution stops. The handoff service returns a confirmed failure before creating a ticket. What should the assistant communicate and preserve?

Show the answer and explanation

Correct answer: B. Report the failed handoff, preserve the human-support request, and check documented next steps.

Attempting a handoff is not a completed transfer. Report the observed failure without inventing a ticket or an unavailable fallback.

Question 9

Prompt Engineering
A review team wants to group adjudicated findings by the code construct involved, such as interpolated queries or unbounded regular expressions. Which additional field directly supports that grouping?

Show the answer and explanation

Correct answer: C. A construct category linked to the finding’s supporting source location.

A construct category supplies the grouping dimension, while the source link permits inspection. The label remains a claim to verify; it does not justify automatically allowlisting every matching construct.

Question 10

Tools & Development
A Claude Code project has approved, authenticated database and issue-tracker MCP connections. Both are configured for deferred tool loading on a supported host. What does tool search allow?

Show the answer and explanation

Correct answer: B. Definitions for needed tools from either server can be loaded on demand without switching connections.

Deferred loading manages when tool definitions enter model context. It does not require dropping the other server’s connection and does not grant additional permissions. Connection health, discovery and execution authorization remain separate concerns.

Question 11

AI & Anthropic Fundamentals
A migration tool may run only after validation of that exact migration against its intended target. Hiding the tool until validation succeeds is used as a UI aid. What additional enforcement is needed?

Show the answer and explanation

Correct answer: C. Enforce trusted validation evidence for this migration and target at execution time.

Tool discovery is not a substitute for enforcing the prerequisite when executing the operation. The check must bind validation to the migration and target, and account for intervening changes. An available tool name or model assertion is not trusted validation evidence.

Question 12

Claude Capabilities
A migration will change authentication flows, service boundaries and database compatibility. The team explicitly requires a design proposal before implementation. What is the appropriate initial workflow?

Show the answer and explanation

Correct answer: B. Investigate the current contracts, develop a reviewable migration plan, and address its unresolved decisions before implementing.

A shared plan can expose cross-component dependencies before implementation. Its quality still depends on evidence, and the eventual changes require validation.

Question 13

Context Management
An MCP tool reports the same generic failure for a missing credential and invalid generated syntax. Which improvement gives the caller useful recovery information without exposing the credential?

Show the answer and explanation

Correct answer: B. Return an error indication with a safe failure code and relevant diagnostic details.

Credential configuration and syntax repair require different responses. MCP can indicate tool execution failure; application-specific codes can describe the distinction without revealing secrets.

Question 14

Prompt Engineering
An incident extractor must report a fix only when the source specifies one; otherwise fix must be null. A ticket describes a known defect but says nothing about a remedy. Which example output follows that contract?

Show the answer and explanation

Correct answer: B. Set fix to null because this ticket supplies no remedy.

A missing remedy is represented by the contract’s null value. Neither a likely vendor action nor a familiar workaround is an extracted fact.

Question 15

Tools & Development
After reading the relevant code, an agent confirms that oldApiCall() occurs once in the current file and that newApiCall() has the required semantics. The available editor supports exact-text replacement. Which change is appropriately scoped?

Show the answer and explanation

Correct answer: B. Replace the confirmed old call with the new call, then inspect the diff and relevant checks.

A targeted edit limits the intended change to the inspected call. Textual uniqueness is not a substitute for checking semantics, so verify the diff and relevant behavior after applying it. Respect the editor’s prerequisites and use current file contents.

Question 16

AI & Anthropic Fundamentals
A code-review agent has reproduced a failure but has not established its root cause. It must hand off to an engineer without the conversation transcript. How should the summary describe the cause?

Show the answer and explanation

Correct answer: B. Label observations and hypotheses separately, with accessible evidence and the next verification step.

The recipient needs to know what is established and what remains a hypothesis. A concise evidence-based handoff can support the next investigation without overstating certainty. Structured fields help preserve that distinction, but they do not make a speculative cause factual.

Question 17

Claude Capabilities
A feature can use either a compatible API extension or a breaking replacement. Both are feasible, but clients have different upgrade constraints. The team asks to compare the approaches before coding. What should the agent produce first?

Show the answer and explanation

Correct answer: B. Compare client compatibility, migration effort and verification needs, then propose a justified approach.

The useful planning output connects alternatives to the decision’s constraints. The team can then resolve the consequential choice with concrete tradeoffs.

Question 18

Prompt Engineering
A structured-output request returns HTTP 200 with stop_reason set to refusal. The response body is explanatory text rather than the expected extraction object. What should the application do?

Show the answer and explanation

Correct answer: C. Handle the refusal outcome separately and avoid recording a completed extraction.

Transport success is not extraction success. A refusal must follow its own handling path instead of becoming fabricated structured data.

Question 19

Tools & Development
A team shares a stdio MCP server definition in one Claude Code repository. The server reads GITHUB_TOKEN from its process environment, and each developer supplies an individual token outside version control. Which configuration fits?

Show the answer and explanation

Correct answer: B. Commit a project .mcp.json definition whose env.GITHUB_TOKEN references ${GITHUB_TOKEN}.

For this stdio server, the environment field matches how the process expects its credential. The shared file contains a reference; each launch environment supplies the value. Keep the resolved credential out of logs and use the intended permissions. Sharing a definition does not distribute credentials.

Question 20

AI & Anthropic Fundamentals
A troubleshooting subagent needs specific errors from a very large log archive. The full archive will not fit comfortably in its context. Which handoff best preserves relevant evidence?

Show the answer and explanation

Correct answer: B. Pass key errors and findings with accessible source references and scoped read access.

A useful handoff combines focused context with verifiable access to additional evidence. It need not paste every raw byte into the prompt. References must actually be readable by the worker, and omitted or summarized material should not be misrepresented as a complete source.

Keep studying CCAR-F

These are original practice questions, not official exam questions. Practice results do not predict an exam outcome.

Sign in to report an error