Practice exam 3 (full length, 60 questions) · CCAR-F

A second full-length exam using a different scenario draw: Code Generation with Claude Code, Multi-Agent Research System, Customer Support Resolution Agent and Structured Data Extraction. Together with Practice exam 2, it covers all six scenarios at full depth. Take it last, under exam conditions: 120 minutes, no notes.

Scenario A: Code Generation with Claude Code (questions 1 to 15)

Your team uses Claude Code for generation, refactoring, debugging and documentation, with shared configuration in the repository.

1. The team wants every session to follow the same commit-message format. Where should the rule go?

A. In each developer's ~/.claude/CLAUDE.md.
B. In the project CLAUDE.md, committed to the repository.
C. In a skill that developers invoke before committing.
D. In a .claude/rules/ file scoped to **/*.md.

2. React conventions apply to components in src/components/ and also in packages/*/ components/. What is the most maintainable way to apply them automatically?

A. A CLAUDE.md in src/components/ only.
B. Copy the conventions into every package's CLAUDE.md.
C. A skill named react-conventions.
D. A .claude/rules/ file whose paths globs match both locations.

3. A developer wants a stricter personal version of the team's /deploy-check command without affecting anyone else. What should they do?

A. Copy it to ~/.claude/commands/ under a different name and customise the copy.
B. Edit .claude/commands/deploy-check.md and not commit it.
C. Add their checks to the project CLAUDE.md.
D. Ask Claude to remember the extra checks.

4. A brainstorming skill generates ten alternative designs, which clutter the main conversation. Which frontmatter setting helps?

A. allowed-tools: Read
B. argument-hint: <topic>
C. context: fork
D. description: brief

5. A skill needs a ticket ID, and developers often invoke it without one. Which frontmatter setting helps?

A. context: fork
B. argument-hint: <ticket-id>
C. allowed-tools: Bash
D. required: ticket-id

6. A teammate built a /standup command, but nobody else can run it. What is the most likely cause?

A. Commands need a restart to load.
B. The command file has no frontmatter.
C. The command name is reserved.
D. It lives in the teammate's ~/.claude/commands/.

7. A package's maintainers want its CLAUDE.md to include only the API standards file from a shared docs/standards/ folder. What should they use?

A. An @import of that one standards file in the package's CLAUDE.md.
B. A copy of the file pasted into the package's CLAUDE.md.
C. A symbolic link to the whole standards folder.
D. A skill that loads all standards.

8. You need to add one field to one DTO, with a clear specification. Which approach fits?

A. Plan mode.
B. The Explore subagent first.
C. Direct execution.
D. A forked session.

9. You must choose between event sourcing and simple CRUD for a new module that other services will depend on. Which approach fits?

A. Direct execution with a detailed prompt.
B. Plan mode, to explore and compare approaches before changing code.
C. Implement both and see which works.
D. Direct execution, switching to plan mode if trouble appears.

10. You are implementing payment idempotency, a domain you have not worked in. What technique surfaces considerations you might miss?

A. Ask for the implementation immediately and review it.
B. Provide a long specification.
C. Use plan mode with no further input.
D. Have Claude interview you with questions about failure modes and edge cases before implementing.

11. A date-parsing bug affects sorting, which affects pagination. How should you report these to Claude?

A. In one detailed message, because the fixes interact.
B. One at a time, starting with pagination.
C. One at a time, in random order.
D. Only the pagination bug, since it is visible.

12. You have two unrelated bugs, in logging and in a CSV export. How should you report them?

A. In one message, to save time.
B. Only the more severe one.
C. One at a time, since they are independent.
D. As a single combined test case.

13. A migration script mishandles null values in one column. What is the most effective way to get it fixed?

A. Ask Claude to "handle nulls properly".
B. Give a specific test case with example input containing nulls and the expected output.
C. Rewrite the script yourself.
D. Switch to plan mode.

14. Claude behaves differently for two developers in the same repository. What should you check first?

A. Their Claude Code versions only.
B. Their network connections.
C. Their operating systems.
D. Which memory files /memory shows as loaded in each session.

15. A rules file with paths: ["terraform/*"] does not load when Claude edits terraform/ modules/vpc/main.tf. What is the fix?

A. Change the glob to terraform/**/*.
B. Move the rules into the root CLAUDE.md.
C. Add a CLAUDE.md in terraform/modules/vpc/.
D. Rename the file to terraform.md.

Scenario B: Multi-Agent Research System (questions 16 to 30)

A coordinator delegates to search, document analysis, synthesis and report subagents, producing cited reports.

16. A simple factual question ("When did the regulation take effect?") still runs the full fouragent pipeline. What should change?

A. Add a cache in front of the pipeline.
B. Use a faster model for every agent.
C. Have the coordinator assess each query and invoke only the subagents it needs.
D. Remove the report subagent.

17. Two search subagents return heavily overlapping results. What is the fix?

A. Deduplicate results after synthesis.
B. Partition the search scope between them, for example by subtopic or source type.
C. Use only one search subagent.
D. Run them sequentially.

18. The coordinator gives subagents a step-by-step procedure. When an expected source is missing, they stall. What would help?

A. A longer procedure with more cases.
B. Fewer subagents.
C. A higher maximum number of turns.
D. Briefs that state the research goal and quality criteria instead of fixed steps.

19. After synthesis, nobody can tell which claim came from which source. What should subagents produce?

A. Structured claim-source mappings that synthesis must preserve.
B. Longer summaries.
C. A list of URLs at the end.
D. Their full transcripts.

20. The synthesis subagent's prompt says "use the previous results". Its output is generic. What is wrong?

A. The synthesis model is too small.
B. Synthesis needs more tools.
C. The coordinator must include the complete findings in the prompt; subagents do not inherit context.
D. Results must be stored in a database.

21. A report treats a 2015 recycling rate and a 2023 rate as contradictory. What prevents this?

A. Discard older sources.
B. Require subagents to include publication or collection dates with each claim.
C. Average the figures.
D. Ask synthesis to choose the most credible one.

22. When search fails, the coordinator receives only "search unavailable". What should it receive instead?

A. Nothing; the subagent should keep retrying.
B. An empty result set.
C. An exception that stops the workflow.
D. Structured context: failure type, attempted query, partial results and alternatives.

23. The synthesis subagent was given the full web search toolset and now runs its own searches, duplicating work. What is the better design?

A. Remove its search tools; give it a scoped verify_fact tool and route deeper research through the coordinator.
B. Let it search but limit it to five searches.
C. Remove the search subagent instead.
D. Add "do not search" to its prompt.

24. Financial comparisons are buried in long paragraphs. What should synthesis do?

A. Shorten the paragraphs.
B. Move numbers to an appendix.
C. Render financial data as tables, news as prose and technical findings as structured lists.
D. Convert everything to bullet points.

25. You want the coordinator to keep researching until coverage is sufficient. What pattern is this?

A. A fixed pipeline.
B. An iterative refinement loop: check for gaps, re-delegate targeted queries, re-synthesise.
C. A batch job.
D. Self-review in the same session.

26. The synthesis agent has a small context budget and receives verbose subagent reasoning. What should change upstream?

A. Nothing; truncate the input.
B. Use a larger model for synthesis.
C. Summarise everything into one paragraph.
D. Have upstream agents return structured key facts, citations and relevance scores.

27. One of five search subagents fails after local retries. What should the coordinator do?

A. Proceed with the available results and annotate the coverage gap in the report.
B. Abort the whole report.
C. Present the report as complete.
D. Restart all five subagents.

28. (Choose two.) What belongs in a subagent's structured error report?

A. The failure type.
B. The query that was attempted.
C. The model's internal stack trace.
D. The subagent's full conversation transcript.

29. Synthesis receives results from ten subagents and tends to omit findings from those in the middle. What helps?

A. Give synthesis more subagents.
B. Shuffle the order each time.
C. Put a key-findings summary first and organise detailed results under clear headers.
D. Remove the middle results.

30. You want to compare two report structures built from the same completed research. What should you use?

A. Two fresh sessions that re-run the research.
B. fork_session from the research baseline.
C. --resume twice in sequence.
D. Plan mode.

Scenario C: Customer Support Resolution Agent (questions 31 to 45)

Your Agent SDK support agent uses MCP tools get_customer, lookup_order, process_refund and escalate_to_human.

31. After a long conversation is summarised, the agent quotes the refund amount as US45insteadofUS54. What prevents this?

A. Summarise less often.
B. Ask the customer to confirm amounts each turn.
C. Use a larger context window.
D. Keep amounts in a case facts block outside the summarised history.

32. Midway through investigating, the customer writes "I want a manager, now." What should the agent do?

A. Escalate immediately with a structured hand-off.
B. Finish the investigation first, then escalate.
C. Offer a discount to keep the customer engaged.
D. Ask why they want a manager.

33. A customer asks for a refund on a digital gift card. Policy covers physical goods only. What should the agent do?

A. Refuse, because gift cards are not covered.
B. Refund, because it is similar to a physical product.
C. Escalate, because the policy does not address this request.
D. Offer store credit instead.

34. get_customer returns the same response for "no such customer" and "service timeout". What should change?

A. Return a default customer.
B. Return a valid empty result for no match and a structured, retryable error for the timeout.
C. Retry every call three times.
D. Remove the tool.

35. The agent sometimes calls process_refund before lookup_order has confirmed eligibility. What is the most effective fix?

A. Add examples of the correct order.
B. Mention the order in the tool description.
C. Use a larger model.
D. Block process_refund programmatically until eligibility has been confirmed in this session.

36. Finance needs every refund capped at US$500 without exception. Where should this be enforced?

A. A pre-tool hook on process_refund.
B. The system prompt.
C. A daily report reviewed by finance.
D. Few-shot examples.

37. lookup_order's description is "Gets order". The agent sometimes uses get_customer for order questions. What is the first fix?

A. Merge both tools.
B. Add a keyword router.
C. Expand both descriptions with inputs, examples and when to use each.
D. Remove get_customer.

38. What should an escalation to a human include?

A. Only the customer's name.
B. Customer ID, issue, root cause, amounts, actions taken and the recommended action.
C. The agent's apology text.
D. A sentiment score.

39. When customers mention several orders, the agent mixes up their details. What helps most?

A. Ask customers to contact support once per order.
B. Use a bigger context window.
C. Escalate all multi-order cases.
D. Keep a structured record per issue (order ID, amount, status) separate from the conversation history.

40. Order results include internal warehouse notes that confuse the agent. What should you do?

A. Filter results to the fields the agent needs before they reach it.
B. Tell the agent to ignore warehouse notes.
C. Escalate orders with notes.
D. Add examples of ignoring notes.

41. The agent escalates whenever sentiment is negative. Escalation rates are high and many cases were simple. What is the fix?

A. Raise the sentiment threshold.
B. Add a second sentiment model.
C. Replace sentiment triggers with explicit escalation criteria and few-shot examples.
D. Escalate only after three negative messages.

42. A customer gives an order number but no other identity details. What should happen?

A. Look up and discuss the order immediately.
B. Verify the customer with get_customer first; the gate blocks order tools until then.
C. Escalate to a human.
D. Ask for the customer's password.

43. get_customer returns two customers with the same name and city. What should the agent do?

A. Choose the customer with more orders.
B. Choose the one with the most recent order.
C. Escalate.
D. Ask for another identifier, such as email or an order number.

44. process_refund returns refund_window_expired. What should the tool's response enable the agent to do?

A. Recognise a non-retryable business error and explain the policy to the customer.
B. Retry with a later date.
C. Retry until it succeeds.
D. Tell the customer the system is down.

45. (Choose two.) Which are appropriate escalation triggers?

A. The customer explicitly asks for a human.
B. The policy does not cover the request.
C. The customer's tone is negative.
D. The model reports low confidence.

Scenario D: Structured Data Extraction (questions 46 to 60)

Your system extracts data from unstructured documents, validates it with JSON schemas and feeds downstream systems.

46. warranty_period is a required string. Many documents have no warranty, and the model invents one. What is the fix?

A. Add "do not invent" to the prompt.
B. Delete the field.
C. Make the field nullable.
D. Default it to "12 months".

47. There is exactly one extraction schema, and you must guarantee it is used. Which tool_choice?

A. {"type": "auto"}
B. {"type": "tool", "name": "extract_record"}
C. {"type": "none"}
D. No tool_choice; rely on the prompt.

48. There are two schemas, for purchase orders and quotes, and the document type is unknown. Which tool_choice?

A. {"type": "auto"}
B. Force the purchase-order tool.
C. Force the quote tool.
D. {"type": "any"}

49. What does a strict JSON schema via tool use not prevent?

A. Semantic errors such as totals that do not add up.
B. Invalid JSON syntax.
C. Missing required keys.
D. Values of the wrong type.

50. Validation rejects "13/25/2026" as a date. What should happen?

A. Drop the record.
B. Store it as-is.
C. Retry with the document, the failed extraction and the specific error.
D. Replace it with today's date.

51. A value is "per the attached schedule", and the schedule was not provided. What is true?

A. Retrying with feedback will fix it.
B. Retries will not help, because the information is absent from the input.
C. A larger model will find it.
D. The model should estimate it.

52. Documents describe sizes informally ("about three feet", "roughly 1m"), and extraction is inconsistent. What helps most?

A. Make the field optional.
B. Use a regular expression.
C. Increase max_tokens.
D. Add normalisation rules and few-shot examples covering informal measurements.

53. Which statement about the Message Batches API is true?

A. It cannot run multi-turn tool calls within a single request.
B. It guarantees results within one hour.
C. It costs the same as the synchronous API.
D. It returns results in submission order.

54. Batches take up to 24 hours, and you submit every 4 hours. What is the worst-case time from a document's arrival to its result?

A. 4 hours.
B. 24 hours.
C. About 28 hours.
D. 48 hours.

55. You are about to process a million documents for the first time. How do you reduce costly resubmissions?

A. Submit in one batch.
B. Refine the prompt on a representative sample first.
C. Use the synchronous API.
D. Skip validation.

56. Human reviewers can check only 5% of documents. Which should they see first?

A. A random 5%.
B. The shortest documents.
C. The newest documents.
D. Low-confidence extractions and documents with ambiguous or contradictory sources.

57. How do you keep measuring the error rate of extractions you no longer review routinely?

A. Stratified random sampling of high-confidence extractions.
B. Wait for complaints.
C. Ask the model to self-audit.
D. Review only failed validations.

58. Field-level analysis shows tax_id is 99% accurate on invoices but 74% on receipts. What should you do?

A. Accept the 95% average.
B. Remove tax_id.
C. Keep human review for tax_id on receipts until it improves.
D. Retrain the model.

59. Some documents genuinely could fit two categories. What schema feature helps?

A. Making category free text.
B. An unclear enum value for ambiguous cases.
C. Forcing the first matching category.
D. Removing the category.

60. (Choose two.) What should a retry request after a validation failure include?

A. The original document.
B. The specific validation errors.
C. A higher temperature.
D. A different model.

Answer key for Practice exam 3

# Answer Domain Why

1 B 3 Project CLAUDE.md is shared through version control.

2 D 3 Glob-scoped rules cover files across directories.

3 A 3 Personal variants go in ~/.claude/ under a new name.

4 C 3 context: fork isolates verbose output.

5 B 3 argument-hint prompts for missing arguments.

6 D 3 User-scoped commands are not shared.

7 A 3 @import includes exactly the files a package needs.

8 C 3 Small, clear changes suit direct execution.

9 B 3 Architectural choices with alternatives suit plan mode.

10 D 3 The interview pattern surfaces unknowns first.

11 A 3 Interacting issues go in one message.

12 C 3 Independent issues are fixed one at a time.

13 B 3 Specific input and expected output fix edge cases.

14 D 3 /memory shows which memory files are loaded.

15 A 3 * matches one level; **/* matches all levels.

16 C 1 Coordinators choose subagents per query.

17 B 1 Partition scope to avoid duplication.

18 D 1 Goals and quality criteria let subagents adapt.

19 A 5 Claim-source mappings preserve provenance.

20 C 1 Subagent context must be passed explicitly.

21 B 5 Dates stop temporal differences looking like contradictions.

22 D 5 Structured error context enables recovery.

23 A 2 Scope tools by role; add narrow cross-role tools.

24 C 5 Fit format to content type.

25 B 1 Iterative refinement until coverage is sufficient.

26 D 5 Structured upstream output saves downstream context.

27 A 5 Proceed with partial results and annotate gaps.

28 A, B 5 Failure type and attempted query, plus partial results and alternatives. 29 C 5 Mitigate position effects with ordering and headers.

30 B 1 Fork from a shared baseline.

31 D 5 Case facts survive summarisation.

32 A 5 Honour explicit requests for a human immediately.

33 C 5 Policy gaps trigger escalation.

34 B 2 Distinguish empty results from access failures.

35 D 1 Programmatic prerequisites enforce order.

36 A 1 Hooks guarantee hard limits.

37 C 2 Better descriptions are the first fix.

38 B 1 Structured hand-offs for humans without the transcript.

39 D 5 Per-issue structured records for multi-issue sessions.

40 A 5 Trim tool output to relevant fields.

41 C 5 Sentiment is not complexity; use explicit criteria.

42 B 1 Identity verification gates order tools.

43 D 5 Ask for identifiers instead of guessing.

44 A 2 Non-retryable business errors with explanations.

45 A, B 5 Explicit requests and policy gaps are triggers.

46 C 4 Nullable fields prevent fabrication.

47 B 4 Forced tool selection guarantees that tool.

48 D 4 any guarantees a call when the type is unknown.

49 A 4 Schemas do not catch semantic errors.

50 C 4 Retry with error feedback.

51 B 4 Absent information cannot be retried into existence.

52 D 4 Normalisation rules and few-shot examples.

53 A 4 No mid-request tool calls in batches.

54 C 4 Up to 4 hours waiting plus up to 24 hours processing.

55 B 4 Refine on a sample before large batches.

56 D 5 Prioritise low-confidence and ambiguous documents.

57 A 5 Stratified sampling measures ongoing error rates.

58 C 5 Segment-level weaknesses keep human review.

59 B 4 unclear handles genuine ambiguity.

60 A, B 4 Document, failed extraction and specific errors.

Scoring. 48 or more correct (80%) suggests readiness. Domain counts in this exam: Domain 1 has 10 items, Domain 2 has 4, Domain 3 has 15, Domain 4 has 12 and Domain 5 has 19. Together, Practice exams 2 and 3 give 120 items across all five domains and all six scenarios. Tempting wrong answers in Practice exam 3

# Tempting answer Why it fails

1 C A skill must be invoked; the rule must always apply (4.2).

2 A A directory CLAUDE.md misses packages/*/components/ (4.3).

3 B Uncommitted edits to a shared file are fragile and confusing (4.2).

4 A Restricting tools does not isolate output (4.2).

5 D required is not a skill frontmatter field (4.2).

6 A Restarts do not share a user-scoped file (4.2).

7 B Pasted copies drift from the source (4.1).

8 A Plan mode adds overhead to a clearly specified change (4.4).

9 D The complexity is known up front; plan from the start (4.4).

10 B A long spec reflects only what you already know to ask (4.5).

11 B Fixing interacting bugs one by one causes churn (4.5).

12 A Mixing independent fixes makes each harder to verify (4.5).

13 A "Handle nulls properly" is vague; examples are precise (4.5).

14 A Version differences are possible, but loaded memory is the likelier cause (4.1). 15 C A directory file fixes one folder, not the pattern (4.3).

16 A Caching hides the cost; the routing is still wrong (2.2).

17 A Deduplicating afterwards still wastes the search effort (2.2).

18 A More procedure means less adaptability (2.3).

19 C A URL list at the end loses which claim came from where (6.6).

20 A Model size cannot compensate for missing input (2.3).

21 A Discarding older data loses valid history (6.6).

22 A Endless retries delay without informing the coordinator (6.3).

23 D A prompt rule is weaker than scoping the tools (3.3).

24 D Uniform bullets suit no content type well (6.6).

25 A A fixed pipeline cannot respond to gaps (2.2).

26 B A larger model does not remove the noise in its input (6.1).

27 B Aborting wastes the four successful searches (6.3).

28 D A full transcript buries the useful context (6.3).

29 B Shuffling moves the blind spot around without removing it (6.1).

30 A Re-running research is costly and changes the baseline (2.7).

31 C A larger window delays but does not protect the fact (6.1).

32 B The customer has already asked; investigating first ignores them (6.2).

33 A Refusing treats a policy gap as a "no" (6.2).

34 C Retrying "no such customer" wastes calls (3.2).

35 A Examples improve but do not guarantee ordering (2.4).

36 B The prompt cannot guarantee a cap (2.5).

37 B A keyword router is over-engineered as a first step (3.1).

38 D Sentiment does not help a human resolve the case (2.4).

39 B A larger window does not separate the issues (6.1).

40 B Telling the model to ignore data it still sees is unreliable (6.1).

41 A A higher threshold on the wrong signal is still the wrong signal (6.2).

42 A Skipping verification is exactly what the gate prevents (2.4).

43 A Order count is a heuristic (6.2).

44 B The error is a policy outcome, not a timing problem (3.2).

45 C Tone is not a trigger (6.2).

46 A The required field still pressures the model to fill it (5.3).

47 A auto allows a text reply (5.3).

48 A auto allows a text reply; forcing one tool guesses the type (5.3).

49 D Types are enforced by the schema (5.3).

50 B Storing invalid data pushes the error downstream (5.4).

51 A Feedback cannot create missing information (5.4).

52 B A regex cannot interpret "about three feet" (5.2).

53 D Results are matched by custom_id, not order (5.5).

54 B Forgets the wait for the next submission window (5.5).

55 C The synchronous API costs twice as much at this volume (5.5).

56 A Random review wastes scarce reviewer time (6.5).

57 C Self-audits are not independent measurement (6.5).

58 A The average hides the weak segment (6.5).

59 C Forcing the first match hides ambiguity (5.3).

60 C Temperature does not tell the model what was wrong (5.4).