Practice exam 2 (full length, 60 questions) · CCAR-F

This exam mirrors the real format: 60 items, four scenarios of 15 items each, 120 minutes. The scenarios are Customer Support Resolution Agent, Developer Productivity with Claude, Claude Code for Continuous Integration and Structured Data Extraction, the same four-scenario structure you will see on exam day. Items are original and not drawn from the real exam. Answers follow the last question.

Take it in one sitting with a timer, after finishing Practice exam 1 and the drills. Scenario A: Customer Support Resolution Agent (questions 1 to 15)

Your agent uses the Claude Agent SDK and MCP tools get_customer, lookup_order, process_refund and escalate_to_human. The target is 80%+ first-contact resolution.

1. lookup_order returns 43 fields per order. After several lookups in one conversation, answers become less accurate and the agent misquotes delivery dates. What is the most effective fix?

A. Increase max_tokens so the agent has room to reason.
B. Trim each order result to the fields the agent needs before it enters the conversation.
C. Instruct the agent to re-read earlier tool results before answering.
D. Start a new conversation after every three lookups.

2. A customer writes: "Third time asking. My order is late AGAIN." The order is simply delayed in transit, which policy covers with a shipping refund. What should the agent do?

A. Escalate, because the customer is clearly frustrated.
B. Escalate, because this is the third contact.
C. Ask the customer to rate their frustration before deciding.
D. Acknowledge the frustration, apply the shipping refund, and escalate only if the customer then asks for a human.

3. After a timeout, the agent retried process_refund and the customer was refunded twice. What should change?

A. Make the refund tool idempotent per order and return structured errors stating whether the refund was applied and whether a retry is safe.
B. Tell the agent in the system prompt never to retry refunds.
C. Remove retries from the agent entirely.
D. Raise the refund service timeout to 60 seconds.

4. Human agents receiving escalations say they must ask customers to repeat everything. What should the escalation include?

A. A link to the full transcript.
B. The customer's last message.
C. A structured summary: customer ID, issues, root cause, amounts, actions taken and the recommended action.
D. The agent's confidence score.

5. The agent asks itself for a confidence score from 1 to 10 and escalates below 7. Escalation is still poorly targeted. What is the best next step?

A. Raise the threshold to 8.
B. Replace it with explicit escalation criteria and few-shot examples of escalate versus resolve.
C. Add sentiment analysis as a second signal.
D. Train a classifier on last year's tickets.

6. Two MCP servers report amounts in cents and one in dollars. You want all amounts in dollars before the agent reasons about them. Where is the best place to do this?

A. A PostToolUse hook that normalises amounts in tool results.
B. A PreToolUse hook on each tool.
C. A note in the system prompt listing which tools use cents.
D. Few-shot examples of unit conversion.

7. A customer raises a damaged item, a missing loyalty credit and an address change in one message. The agent resolves only the damaged item. What should you change?

A. Ask customers to send one issue per message.
B. Escalate all multi-issue messages.
C. Increase the model's maximum output length.
D. Instruct the agent to list each distinct issue, investigate each, and reply once addressing all of them.

8. Policy says refunds for damaged items require a photo. The agent sometimes refunds without one. The rule must never be broken. What is the most effective fix?

A. Bold the photo requirement in the system prompt.
B. Add three few-shot examples asking for photos.
C. Block process_refund for damage claims unless the case record shows a photo was received.
D. Ask the agent to confirm it has a photo before each refund.

9. A customer asks to transfer loyalty points to a family member. No tool supports it and the policy does not mention it. What should the agent do?

A. Decline, since there is no tool.
B. Escalate, because it cannot make progress and the policy is silent.
C. Promise the transfer and log a note.
D. Suggest the customer try again later.

10. Fifty turns into a complex case, the agent forgets that it promised a replacement by 15 November. What prevents this?

A. A persistent case facts block, outside the summarised history, included in every request.
B. A larger context window.
C. Summarising the conversation more often.
D. Asking the customer to restate commitments at each turn.

11. The payment service returns card_declined. The agent retries five times, then tells the customer "our system is down". What is the fix?

A. Limit retries to two.
B. Add a pause between retries.
C. Remove the payment tool.
D. Return a structured business error with isRetryable: false and a customer-friendly explanation.

12. (Choose two.) Which requirements should be enforced programmatically rather than by prompt?

A. Identity verification before any account change.
B. A friendly, concise tone.
C. No refunds above US$500 without human approval.
D. Addressing the customer by first name.

13. get_customer returns an empty list both when no customer matches and when the CRM is unreachable. What should it do instead?

A. Always return at least one closest match.
B. Distinguish a valid empty result from an access failure, returning a structured error for the latter.
C. Raise an exception that ends the conversation.
D. Retry silently until the CRM responds.

14. Product managers want to add 12 more tools to the support agent, for inventory, marketing and shipping. What is the best response?

A. Add them all; more tools mean more capability.
B. Add them and describe each in detail.
C. Keep the support agent's tools scoped to its role, and route the rare other needs through separate specialised agents or a coordinator.
D. Replace all tools with one do_anything tool.

15. The system prompt says "always check the customer's order history thoroughly". The agent now calls lookup_order for password-reset questions. What is the likely root cause?

A. A keyword-sensitive system prompt instruction is overriding the tool descriptions.
B. The lookup_order description is too long.
C. The model needs tool_choice: "any".
D. The password-reset tool is missing a required field.

Scenario B: Developer Productivity with Claude (questions 16 to 30)

You are building an Agent SDK tool that helps engineers explore unfamiliar codebases, using Read, Write, Edit, Bash, Grep, Glob and MCP servers.

16. Asked how authentication works, the agent reads all 300 source files first and runs out of context. What is the better strategy?

A. Read files in alphabetical order until context is full.
B. Use a model with a larger context window.
C. Grep for entry points such as login and authenticate, then Read along the imports.
D. Ask the user which files to read.

17. Which tool should the agent use to find every call of formatCurrency?

A. Grep
B. Glob
C. Read
D. Write

18. Which tool should the agent use to find every Dockerfile in the repository?

A. Grep for FROM
B. Read each directory
C. Bash cat
D. Glob for **/Dockerfile

19. Mapping a legacy billing module produces pages of grep and file output that crowd out the main investigation. What helps most?

A. Ask the agent to write shorter answers.
B. Delegate the mapping to a subagent that returns a summary.
C. Disable Grep.
D. Restart after the mapping completes, without a summary.

20. An investigation will span three days. How should findings carry over?

A. A findings file updated each session and used to seed new sessions.
B. One session kept open for three days.
C. Rely on the model remembering earlier sessions.
D. Re-run the full exploration each morning.

21. A long multi-agent indexing job crashes at 70% and restarts from zero. What design prevents this?

A. Run it on a larger machine.
B. Reduce the number of subagents.
C. Add more retries to each tool.
D. Have each agent export state to a known location and the coordinator load a manifest on restart.

22. The team needs Jira access in the agent. What is the recommended approach?

A. Write a custom Jira MCP server from scratch.
B. Give the agent Bash and let it call the Jira REST API with curl.
C. Use an established Jira MCP server, reserving custom servers for team-specific workflows.
D. Export Jira to CSV nightly and let the agent Read it.

23. A "test-writer" subagent was given Bash, Write, WebFetch and deployment tools. It once triggered a deployment. What is the fix?

A. Add "never deploy" to its prompt.
B. Restrict its tools in its AgentDefinition to those its role needs.
C. Review its transcripts weekly.
D. Lower its maximum turns.

24. A utility is re-exported from an index file as both default and fmt. Grep for its original name misses most usages. What is the correct method?

A. Identify every name it is exported under, then Grep for each.
B. Read every file that imports from the index.
C. Use Glob for files containing the name.
D. Ask the user where it is used.

25. You want to compare two caching strategies starting from the same codebase analysis. What should you use?

A. Plan mode twice.
B. Two unrelated fresh sessions.
C. --resume and switch strategies halfway.
D. fork_session from the analysis baseline.

26. You resume yesterday's session, but a teammate has since refactored two of the files it analysed. What should you do?

A. Resume without comment; the agent will notice.
B. Start over with no context.
C. Resume and tell the agent which two files changed so it re-reads them.
D. Resume and ask it to re-read the entire codebase.

27. The coordinator delegates "find all usages of the deprecated API" and the subagent replies that it does not know which API is meant. What is the root cause?

A. The subagent's tools are too limited.
B. The coordinator did not include the necessary context in the subagent's prompt.
C. The subagent's model is too small.
D. Grep is not configured.

28. (Choose two.) Which configurations are correct?

A. A team-wide MCP server in .mcp.json using ${GITHUB_TOKEN}.
B. A personal experimental MCP server in ~/.claude.json.
C. A personal experimental server in .mcp.json with a comment saying it is optional.
D. The shared server's token written directly in .mcp.json.

29. The agent makes many exploratory calls to discover what database tables exist. What reduces this?

A. A tool that lists tables one at a time.
B. A longer system prompt naming every table.
C. Forcing a describe_table call first.
D. Exposing the schema catalogue as an MCP resource.

30. An engineer asks the agent to "add retries to every HTTP client across our services". What decomposition fits?

A. Map where HTTP clients are created, identify the highest-impact ones, and follow a prioritised plan that adapts as dependencies appear.
B. A fixed pipeline that edits every file in alphabetical order.
C. Edit files as they are found, without a plan.
D. One prompt containing every file in the repository.

Scenario C: Claude Code for Continuous Integration (questions 31 to 45)

Your pipeline runs Claude Code to review pull requests, generate tests and post feedback. The goal is actionable feedback with few false positives.

31. The pipeline asks the same Claude Code session that generated a feature to review it. It approves code a human later rejects. What helps most?

A. Ask it to "think harder" during review.
B. Run the review twice in that session.
C. Review with an independent instance that has no access to the generation session.
D. Lower the temperature for review.

32. You need a JSON array of findings, each with file, line, severity and message, validated against a schema. Which approach fits?

A. Ask for JSON in the prompt and fix errors with a script.
B. Use --output-format json with --json-schema.
C. Parse the markdown output.
D. Post the raw text as one comment.

33. The same kind of issue is labelled "critical" on one PR and "minor" on another. What improves consistency?

A. Ask the model to be consistent.
B. Remove severity from the output.
C. Run three times and take the most common severity.
D. Define each severity level explicitly with a concrete code example.

34. Security findings are 50% false positives while bug findings are accurate. Developers have started ignoring all comments. What should you do first?

A. Temporarily disable the security category while you improve its prompt.
B. Disable all categories.
C. Add "only flag high-confidence security issues".
D. Double the number of security checks.

35. The review prompt says "be thorough but only flag real issues". Output is still noisy. What is the best change?

A. Add "be very conservative".
B. Use a larger model.
C. List exactly which issue types to report and which to skip, with examples.
D. Limit output to three findings.

36. You want to learn which kinds of findings developers dismiss most often. What should you add?

A. A thumbs-up button with no other data.
B. A detected_pattern field on each finding, analysed against dismissals.
C. A confidence score only.
D. A weekly survey.

37. Which workload should move to the Message Batches API?

A. The blocking pre-merge review.
B. The check that gates deployment.
C. Interactive developer questions in the IDE.
D. Nightly test generation for modules changed that day.

38. The CI reviewer does not know your testing conventions or review criteria. Where should they go?

A. The project CLAUDE.md, which CI-invoked Claude Code reads.
B. Each developer's ~/.claude/CLAUDE.md.
C. A comment at the top of every PR.
D. Environment variables.

39. Which invocation runs non-interactively and returns machine-readable output?

A. claude "Review this PR"
B. claude --batch "Review this PR"
C. claude -p "Review this PR" --output-format json
D. CLAUDE_HEADLESS=1 claude "Review this PR"

40. After a developer pushes fixes, the bot re-flags issues that were already resolved. What should the next run receive?

A. Only the latest commit's diff, without context.
B. The prior findings, with an instruction to report only new or still-unaddressed issues.
C. Nothing different; run it again.
D. A lower temperature.

41. A 30-file PR gets thorough review on early files and almost nothing on later ones. What should you change?

A. Put the most important files first.
B. Ask developers to submit fewer files.
C. Use a larger context window.
D. Review each file in its own pass, then run a separate cross-file integration pass.

42. Findings arrive sometimes as prose, sometimes as bullets, sometimes without line numbers. What is the most effective fix?

A. Add few-shot examples showing the exact finding format: location, issue, severity, suggested fix.
B. Ask for "a consistent format".
C. Post-process with regular expressions.
D. Remove the suggested-fix field.

43. You want humans to look first at the findings the reviewer is least sure of. What does the guide suggest?

A. Show all findings in random order.
B. Only post findings that three runs agree on.
C. Have the reviewer report a confidence with each finding, to route human attention.
D. Hide low-severity findings.

44. Generated tests often duplicate tests that already exist. What helps most?

A. Ask for "novel tests".
B. Include the existing test files in context.
C. Generate more tests and deduplicate later.
D. Use a different model for test generation.

45. (Choose two.) What belongs in the project CLAUDE.md for CI use?

A. Testing standards, available fixtures and what makes a test valuable.
B. Review criteria: what to report and what to skip.
C. One developer's editor preferences.
D. API keys for the review job.

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

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

46. In research-paper extraction, references are captured when they are inline but come back empty when the paper uses a bibliography at the end. What helps most?

A. Make the references field required.
B. Retry until references are found.
C. Increase max_tokens.
D. Add few-shot examples showing correct extraction from both citation styles.

47. Expense receipts sometimes do not fit any category in your enum, and the model forces them into the nearest one. What should you change?

A. Remove the enum and accept free text.
B. Add an other value with a detail field.
C. Add 40 more categories.
D. Tell the model to pick the closest category.

48. With tool_choice: {"type": "auto"}, the model sometimes answers "I couldn't find all the fields" in text instead of calling the extraction tool. What fixes this?

A. Set tool_choice so a tool call is required, either any or the specific extraction tool.
B. Add "always call the tool" to the prompt.
C. Remove the text response from the output.
D. Retry until a tool call appears.

49. Metadata extraction must happen before any classification tool runs. What should the first request use?

A. tool_choice: {"type": "auto"}
B. tool_choice: {"type": "any"}
C. tool_choice: {"type": "tool", "name": "extract_metadata"}
D. A system prompt instruction to extract metadata first.

50. Validation fails on a date format. What should the retry request contain?

A. Only the instruction "try again".
B. The failed extraction only.
C. The document only.
D. The original document, the failed extraction and the specific validation error.

51. Some extractions fail because amounts are "as per Appendix C", and Appendix C was not provided. Retries keep failing. What should you do?

A. Retry with a stronger model.
B. Stop retrying these; return null with a reason, and supply the appendix or route to review.
C. Ask the model to estimate the amount.
D. Retry with a higher temperature.

52. Documents write dates as "3rd March 2026", "03/04/26" and "2026-03-03". What helps most?

A. Put explicit normalisation rules in the prompt alongside an ISO 8601 schema description, and flag dates that are genuinely ambiguous.
B. Accept whatever format appears.
C. Use a regex after extraction only.
D. Remove the date field.

53. Documents must be processed within 36 hours of arrival. Batches can take up to 24 hours. How often must you submit batches, at most?

A. Once every 24 hours.
B. At least every 12 hours, and more often to leave a margin.
C. Once a week.
D. Every 30 hours.

54. Batch results do not come back in submission order. How do you match results to documents?

A. By position in the results file.
B. By timestamp.
C. By the custom_id you set on each request.
D. By document length.

55. In a 10,000-document batch, 180 requests fail because the documents exceed the context limit. What should you do?

A. Resubmit the full batch.
B. Drop the 180 documents.
C. Switch the whole workload to the synchronous API.
D. Resubmit only those 180, by custom_id, after chunking them.

56. Before batch-processing 500,000 documents for the first time, what should you do?

A. Refine the prompt on a representative sample until first-pass success is high.
B. Submit everything and fix failures afterwards.
C. Submit in one batch to save time.
D. Use the synchronous API for the first 100,000.

57. Overall accuracy is 98%, but invoices from one large supplier extract correctly only 70% of the time. What does this show?

A. The model needs fine-tuning.
B. The 98% is wrong.
C. Aggregate accuracy can hide weak segments; measure by document type and source before automating.
D. The supplier's invoices should be excluded.

58. When the model reports 0.95 confidence, the actual accuracy on a labelled set is 85%. What should you do?

A. Trust the model's confidence anyway.
B. Calibrate review thresholds against the labelled validation set.
C. Remove confidence scores.
D. Ask the model to be more honest.

59. After automating high-confidence extractions, how do you keep measuring quality?

A. Wait for downstream complaints.
B. Re-run all documents monthly.
C. Review only low-confidence items.
D. Review a stratified random sample of high-confidence extractions.

60. (Choose two.) Which checks catch semantic errors that a strict schema cannot?

A. Comparing a calculated_total with the stated_total.
B. A conflict_detected flag for contradictory source data.
C. JSON syntax validation.
D. Checking that required keys are present.

Answer key for Practice exam 2

# Answer Domain Why

1 B 5 Trim verbose tool output before it accumulates.

2 D 5 Frustration alone is not an escalation trigger when the fix is within policy. 3 A 2 Structured errors say whether a retry is safe; idempotency prevents double effects. 4 C 1 Humans lack the transcript; hand off structured facts.

5 B 5 Self-reported confidence is poorly calibrated; use explicit criteria with examples. 6 A 1 PostToolUse hooks normalise results before the model sees them.

7 D 1 Decompose multi-concern requests, then give one unified reply.

8 C 1 A must-never rule needs a programmatic prerequisite.

9 B 5 Policy gap plus no way to progress means escalate.

10 A 5 Case facts outside summarised history survive long conversations.

11 D 2 Business errors are not retryable; say so and explain.

12 A, C 1 Identity and financial limits need guarantees; tone does not.

13 B 2 An access failure is not an empty result.

14 C 2 Too many tools degrade selection; scope by role.

15 A 2 Keyword-sensitive prompt wording can override good descriptions.

16 C 2 Explore incrementally from entry points.

17 A 2 Grep searches file contents.

18 D 2 Glob matches path patterns.

19 B 5 Subagents isolate verbose exploration.

20 A 5 Scratchpad files persist findings across sessions.

21 D 5 State exports and a manifest enable crash recovery.

22 C 2 Use existing servers for standard integrations.

23 B 2 Restrict each subagent's tools to its role.

24 A 2 Trace all exported names through wrappers.

25 D 1 fork_session creates independent branches from a baseline.

26 C 1 Resume and name the changed files for targeted re-analysis.

27 B 1 Subagents only know what their prompt tells them.

28 A, B 2 Shared servers in .mcp.json with env expansion; personal ones in ~/.claude.json. 29 D 2 Resources expose catalogues without exploratory calls.

30 A 1 Open-ended tasks need adaptive decomposition.

31 C 4 Independent instances review better than the author session.

32 B 3 The documented CLI flags for structured CI output.

33 D 4 Explicit severity criteria with examples give consistency.

34 A 4 Temporarily disable a noisy category to restore trust.

35 C 4 Specific categorical criteria beat general caution.

36 B 4 detected_pattern enables analysis of dismissals.

37 D 4 Batch latency-tolerant work only.

38 A 3 CLAUDE.md supplies project context to CI runs.

39 C 3 -p with --output-format json; the other flags do not exist.

40 B 3 Prior findings in context prevent duplicate comments.

41 D 4 Per-file plus integration passes fix attention dilution.

42 A 4 Few-shot examples are the most effective way to fix format.

43 C 4 Per-finding confidence routes reviewer attention.

44 B 3 Existing tests in context prevent duplicates.

45 A, B 3 Project context and criteria; never secrets or personal settings.

46 D 4 Few-shot examples for varied document structures.

47 B 4 other plus detail for extensible categories.

48 A 4 auto lets the model answer in text; require a tool call.

49 C 2 Forced tool selection ensures the step runs first.

50 D 4 Retry with document, failed output and specific errors.

51 B 4 Retries cannot recover absent information.

52 A 4 Normalisation rules in the prompt alongside the schema.

53 B 4 N + 24 ≤ 36, so N ≤ 12; submit more often for margin.

54 C 4 custom_id correlates requests and results.

55 D 4 Resubmit only failures, modified as needed.

56 A 4 Refine on a sample before large batches.

57 C 5 Segment accuracy by type and source.

58 B 5 Calibrate confidence with labelled data.

59 D 5 Stratified sampling keeps measuring error rates.

60 A, B 4 Semantic checks catch what schemas cannot.

Scoring. Count multiple-response items correct only if both choices are right. 48 or more correct (80%) is a strong signal of readiness. Then use the Domain column to total your score per domain and compare with the blueprint: Domain 1 has 9 items here, Domain 2 has 14, Domain 3 has 6, Domain 4 has 20 and Domain 5 has 11. Because this exam's scenario mix leans on Domains 2 and 4, pair it with Practice exam 1 to cover Domain 3 fully.

Tempting wrong answers in Practice exam 2

For each question, the distractor most candidates pick, and why it fails. If you chose it, reread the chapter section in brackets.

1 C Re-reading old results adds tokens; the noise stays (6.1).

2 A Frustration alone is not an escalation trigger (6.2).

3 B A prompt rule cannot tell the agent whether the refund already went through (3.2). 4 A A transcript link is not a hand-off; humans need the facts and a recommendation (2.4). 5 A A higher threshold on an uncalibrated signal is still uncalibrated (6.2). 6 C A prompt note leaves unit conversion to chance on every call (2.5).

7 A Shifts the burden to customers instead of fixing the agent (2.4).

8 D Self-confirmation is still a prompt-level behaviour (2.4).

9 A Declining treats a policy gap as a "no" (6.2).

10 B A larger window delays the loss; it does not protect the fact (6.1).

11 A Fewer retries of a non-retryable error is still wrong (3.2).

12 B Tone matters but does not need a guarantee (2.5).

13 D Silent retries hide outages and stall the conversation (3.2).

14 B Good descriptions do not cure an overloaded tool set (3.3).

15 C any forces a call; it does not change which tool is chosen (3.1).

16 B A bigger window postpones the same mistake (3.5).

17 B Glob matches paths, not code contents (3.5).

18 A Grepping for FROM hits many files that are not Dockerfiles (3.5).

19 D Restarting without a summary discards the findings (6.4).

20 B A three-day session degrades (6.4).

21 C Retries do not survive a crash; saved state does (6.4).

22 A Rebuilding a standard integration wastes effort (3.4).

23 A A prompt rule is weaker than not having the tool (3.3).

24 B Reading every importer is slow and still misses aliases (3.5).

25 B Fresh sessions lose the shared baseline (2.7).

26 D A full re-read wastes time; name the changed files (2.7).

27 C The model was never given the information (2.3).

28 C Optional comments do not make shared config personal (3.4).

29 B A hard-coded list goes stale and costs tokens on every call (3.4).

30 B A fixed pipeline ignores dependencies discovered on the way (2.6).

31 A "Think harder" in the same context keeps the same blind spots (5.6).

32 A Prompt-requested JSON is not schema-enforced (4.6).

33 C Majority voting stabilises noise, not criteria (5.1).

34 C Confidence wording does not improve precision (5.1).

35 A "More conservative" is the same vague instruction (5.1).

36 C A confidence score does not tell you which pattern was dismissed (5.4).

37 A Developers cannot wait up to 24 hours to merge (5.5).

38 B User-level files do not reach CI or teammates (4.1).

39 B --batch does not exist (4.6).

40 C Without prior findings the model cannot know what was resolved (4.6).

41 C More context does not fix attention dilution (5.6).

42 C Regex clean-up treats the symptom (5.2).

43 B Consensus suppresses real but intermittent findings (5.6).

44 C Generating more creates more duplicates (4.6).

45 C Personal preferences do not belong in shared project config (4.1).

46 A A required field invites invention when data is absent (5.3).

47 C No enum is ever complete; add an escape hatch (5.3).

48 B A prompt instruction does not guarantee a tool call (5.3).

49 B any lets the model pick a classification tool first (3.3).

50 A "Try again" gives the model nothing to correct against (5.4).

51 A A stronger model cannot read a missing appendix (5.4).

52 C A regex cannot resolve "03/04/26" day-month order (5.3).

53 A Submitting daily allows up to 48 hours end to end (5.5).

54 A Batch results are not guaranteed in order (5.5).

55 A Resubmitting everything doubles the cost (5.5).

56 B Skipping the pilot multiplies any prompt flaw by 500,000 (5.5).

57 A Fine-tuning is out of scope; measurement and examples come first (6.5).

58 A Uncalibrated confidence misroutes review (6.5).

59 C Reviewing only low confidence never measures high-confidence errors (6.5). 60 C Syntax and key checks are already guaranteed by the schema (5.3).