Free Claude Certification (CCA-F) Practice Test

12 free Claude Architect practice questions

1. Q1Agentic Architecture & Orchestration

   Question 1. A retail company is building a Claude-powered order support agent. In a test run, Claude responds with a request to call the lookup_order tool, and the API response has stop_reason set to tool_use. What should the application’s agentic loop do next?

   1. AExecute lookup_order, append the tool result to the conversation history, and send the updated history back to Claude for the next iteration.
   2. BTreat the tool request as Claude’s final answer because the assistant has already selected the correct tool.
   3. CExecute lookup_order, store the result in an application log only, and ask Claude to answer from the original conversation.
   4. DAsk Claude to summarize why it wanted the tool before executing the tool, then continue only if the summary contains the word "necessary."

   Show the answer

   Correct answer: A. Execute lookup_order, append the tool result to the conversation history, and send the updated history back to Claude for the next iteration.

   tool_use means the loop should execute the requested tool, add the result to context, and call Claude again.

2. Q2Tool Design & MCP Integration

   Question 2. A healthcare operations assistant has find_appointment_slots and book_appointment. Users often ask, "Can you see if Dr. Rao has anything open next Tuesday?" Claude sometimes calls book_appointment with partial information. The book_appointment description says "handles scheduling appointments." Which improvement best reduces accidental booking attempts?

   1. ARemove find_appointment_slots and make book_appointment check availability before booking whenever information is missing.
   2. BAdd the word "book" to the user-facing assistant name so users learn to phrase requests more clearly.
   3. CUpdate book_appointment to state that it is only for confirmed booking requests, list required inputs, show examples, and explicitly direct availability checks to find_appointment_slots.
   4. DKeep the description as-is, but require the book_appointment tool to accept null values for unknown fields.

   Show the answer

   Correct answer: C. Update book_appointment to state that it is only for confirmed booking requests, list required inputs, show examples, and explicitly direct availability checks to find_appointment_slots.

   The booking tool needs clear boundaries, required input formats, and examples that distinguish it from availability search.

3. Q3Claude Code Configuration & Workflows

   Question 3. A platform engineer is investigating why Claude Code gives different migration advice in two sessions started from the same repository. One session appears to load database migration standards, while another does not. Before restructuring the repository, what is the best first diagnostic step?

   1. ARun the /memory command in each session to verify which memory files are loaded.
   2. BDelete all directory-level CLAUDE.md files so only the root file can influence Claude Code.
   3. CImmediately move all database migration instructions into ~/.claude/CLAUDE.md so they load consistently for the platform engineer.
   4. DAssume the inconsistency is caused by Claude Code randomness and add the migration standards twice to the root CLAUDE.md.

   Show the answer

   Correct answer: A. Run the /memory command in each session to verify which memory files are loaded.

   /memory is used to verify loaded memory files and diagnose inconsistent behavior across sessions.

4. Q4Prompt Engineering & Structured Output

   Question 4. A financial technology company uses Claude to classify code review findings as Critical, High, Medium, or Low. Different runs classify the same null-check omission as either Low or High depending on wording in the surrounding code. The team wants consistent severity labels that developers will accept. What should the prompt include?

   1. AA reminder that Claude should assign the highest severity whenever user money or financial data is mentioned anywhere in the file.
   2. BConcrete severity criteria with code-oriented examples for each level, such as Critical for exploitable authorization bypasses, High for data corruption or financial miscalculation paths, Medium for reachable failures with bounded impact, and Low for non-blocking maintainability concerns.
   3. CAn instruction to always downgrade severity by one level unless the model is absolutely certain.
   4. DA requirement that Claude explain why it is confident in the severity label, without changing the criteria for each label.

   Show the answer

   Correct answer: B. Concrete severity criteria with code-oriented examples for each level, such as Critical for exploitable authorization bypasses, High for data corruption or financial miscalculation paths, Medium for reachable failures with bounded impact, and Low for non-blocking maintainability concerns.

   Explicit severity criteria with examples produce more consistent classification than abstract labels.

5. Q5Context Management & Reliability

   Question 5. A legal research product aggregates 25 case analyses from subagents into one synthesis prompt. The synthesis model consistently captures the first and last few cases but misses important holdings from cases listed in the middle. Which prompt organization best mitigates this failure mode?

   1. APlace all methodological notes first, put key findings in the middle, and reserve the end for citations.
   2. BSort the case analyses randomly on each run so no case is always placed in the middle.
   3. CRemove section headers so the model treats the input as one continuous narrative without artificial boundaries.
   4. DPut a concise key-findings summary at the beginning of the aggregated input, then organize the detailed case analyses under explicit section headers.

   Show the answer

   Correct answer: D. Put a concise key-findings summary at the beginning of the aggregated input, then organize the detailed case analyses under explicit section headers.

   Placing key findings early and using clear section headers helps mitigate the lost-in-the-middle effect.

6. Q6Agentic Architecture & Orchestration

   Question 6. A renewable energy consultancy uses subagents to research policy incentives, grid interconnection constraints, financing models, and equipment supply chains. After the first synthesis, the coordinator detects that the answer is strong on incentives and financing but weak on supply chain risks for battery components. What should the coordinator do next?

   1. AAsk all subagents to repeat their original research from scratch with a higher-level prompt asking for more detail.
   2. BRe-delegate targeted follow-up questions to the relevant search and analysis subagents, then re-run synthesis after the gap is addressed.
   3. CHave the supply chain subagent directly edit the final report section without routing changes through the coordinator.
   4. DAccept the first synthesis because subagents have already completed their initial assignments, and further delegation may introduce inconsistencies.

   Show the answer

   Correct answer: B. Re-delegate targeted follow-up questions to the relevant search and analysis subagents, then re-run synthesis after the gap is addressed.

   This is an iterative refinement loop: evaluate synthesis for gaps, re-delegate targeted work, and synthesize again.

7. Q7Tool Design & MCP Integration

   Question 7. A subscription billing assistant uses an MCP tool apply_plan_credit(accountId, creditCode). If the code has expired, company policy says the credit cannot be applied, but the assistant should politely explain why and suggest contacting support for exceptions. Which error response best supports that behavior?

   1. A{ "isError": true, "errorCategory": "transient", "isRetryable": true, "description": "Credit was not applied." }
   2. B{ "isError": true, "errorCategory": "validation", "isRetryable": true, "description": "The credit code format is invalid; ask the customer to re-enter it." }
   3. C{ "isError": false, "applied": false, "description": "Expired code" }
   4. D{ "isError": true, "errorCategory": "business", "isRetryable": false, "description": "This credit code expired on 2026-01-31 and cannot be applied automatically. The customer may contact support to request an exception review." }

   Show the answer

   Correct answer: D. { "isError": true, "errorCategory": "business", "isRetryable": false, "description": "This credit code expired on 2026-01-31 and cannot be applied automatically. The customer may contact support to request an exception review." }

   This identifies a non-retryable business-rule violation and gives the agent a customer-friendly explanation.

8. Q8Claude Code Configuration & Workflows

   Question 8. A healthcare software team wants Claude Code to follow privacy-safe logging conventions on every task, but it also has an occasional workflow that generates a migration-risk checklist for a specific database table. Which configuration design best separates these needs?

   1. APut the privacy-safe logging conventions in CLAUDE.md and package the table-specific migration checklist as an on-demand skill.
   2. BPut both the privacy logging rules and the migration checklist in a single project skill so Claude loads them only when explicitly invoked.
   3. CCreate a user-scoped slash command in ~/.claude/commands/ for the logging rules and a project CLAUDE.md entry for the migration checklist.
   4. DPut both items in CLAUDE.md so developers never need to invoke anything manually.

   Show the answer

   Correct answer: A. Put the privacy-safe logging conventions in CLAUDE.md and package the table-specific migration checklist as an on-demand skill.

   Always-loaded universal standards belong in CLAUDE.md, while task-specific workflows are better packaged as skills invoked when needed.

9. Q9Prompt Engineering & Structured Output

   Question 9. A payroll software team uses Claude to review pull requests for user-facing error-message defects. The team’s prompt lists required output fields—file, line, issue, severity, and suggested fix—but Claude often changes the order, omits the fix, or writes vague findings. Which prompt change is most likely to improve consistently formatted, actionable output?

   1. AAsk Claude to produce only a one-sentence summary for each defect so there are fewer fields to omit.
   2. BReplace the field list with the instruction: “Be strict, concise, and only report high-confidence issues.”
   3. CAdd a post-processing script that reorders whatever text Claude returns into the desired field order.
   4. DAdd 2–4 few-shot examples that show complete findings in the exact desired format, including location, issue, severity, and suggested fix.

   Show the answer

   Correct answer: D. Add 2–4 few-shot examples that show complete findings in the exact desired format, including location, issue, severity, and suggested fix.

   Few-shot examples are the best fit when detailed instructions alone produce inconsistent formatting. They demonstrate the exact output shape and level of actionability expected.

10. Q10Context Management & Reliability

    Question 10. A research coordinator uses several subagents to gather market data for a quarterly report. The Asia-Pacific pricing subagent queries a paid database, receives a timeout twice, then retrieves partial results from cached regional summaries. What should the subagent return to the coordinator?

    1. ATerminate the entire multi-agent workflow because one required source was unavailable.
    2. BA generic status such as "search unavailable" without details, to avoid overloading the coordinator context.
    3. CAn empty successful result set, because the paid database did not return complete live data.
    4. DA structured error context that identifies the timeout as an access failure, includes the attempted query, retry attempts, partial cached results, and possible alternative sources.

    Show the answer

    Correct answer: D. A structured error context that identifies the timeout as an access failure, includes the attempted query, retry attempts, partial cached results, and possible alternative sources.

    Structured error context lets the coordinator decide whether to retry, use partial evidence, or route to another source.

11. Q11Agentic Architecture & Orchestration

    Question 11. A food-delivery marketplace is configuring a coordinator with two subagent types. menu_data_auditor should inspect restaurant menu feeds and call only internal catalog tools. competitor_price_researcher should use external web search but must not access internal restaurant contracts. What is the best way to configure this system?

    1. AGive both subagents all tools, but tell them in the prompt not to use tools outside their specialty.
    2. BCreate separate AgentDefinition configurations with clear descriptions, role-specific system prompts, and tool restrictions for each subagent type.
    3. CUse a single generic researcher subagent and let the coordinator describe the desired behavior differently each time.
    4. DAllow the menu_data_auditor to spawn the competitor_price_researcher directly when it finds missing prices.

    Show the answer

    Correct answer: B. Create separate AgentDefinition configurations with clear descriptions, role-specific system prompts, and tool restrictions for each subagent type.

    AgentDefinition should define descriptions, system prompts, and tool restrictions for each subagent type.

12. Q12Tool Design & MCP Integration

    Question 12. A subscription billing assistant has an entitlement-check subagent. For every user request, the workflow requires the subagent to call one of its available account tools before it writes any final response, but the model may choose the appropriate tool based on the request. Which tool_choice setting best fits this requirement?

    1. AForce tool_choice to {"type":"tool","name":"lookup_invoice_history"} for every request.
    2. BSet tool_choice to "any" for the entitlement-check subagent.
    3. CRemove all account tools and require the coordinator to summarize account state in the prompt before invoking the subagent.
    4. DSet tool_choice to "auto" so the model can decide whether a tool call is necessary.

    Show the answer

    Correct answer: B. Set tool_choice to "any" for the entitlement-check subagent.

    tool_choice: "any" guarantees the model calls a tool while still allowing it to choose which available tool is appropriate.