10 Practice exam · CCDV-F

This practice exam mirrors the real one: 53 items, 120 minutes, weighted to the official blueprint (18 items on Applications and Integration, 9 on Model Selection, 8 on Agents, 6 each on Prompt and Context and on Tools and MCPs, 4 on Security, 1 each on Claude Code and Eval). Items are written for this book, not taken from the live exam. They are grouped by domain so you can review weak areas; for a realistic run, answer them in random order. Aim for at least 75% correct before booking.

Applications and Integration (items 1 to 18)

1. A retailer must classify 50,000 product reviews by sentiment for a weekly report. Results are needed within a day and cost is the main concern. What is the best approach?

A. Send all requests in parallel through the Messages API with streaming.
B. Submit the reviews through the Message Batches API.
C. Lower max_tokens on each synchronous request to 10.
D. Use Claude Code headless mode in a loop.

2. Users of a support chatbot complain they stare at a blank screen for 15 seconds before the full answer appears. What should the developer change?

A. Switch to the Batch API.
B. Increase max_tokens.
C. Add "respond quickly" to the system prompt.
D. Stream the response and render text deltas as they arrive.

3. An extraction service occasionally fails to parse Claude's JSON because it ends mid-object. Which check would have caught this before parsing?

A. stop_reason equals max_tokens.
B. usage.cache_read_input_tokens is zero.
C. The HTTP status is 529.
D. The response has more than one content block.

4. A developer's chatbot answers each question well but forgets what the user said two turns ago. What is the cause?

A. The context window is too small.
B. Prompt caching expired.
C. The model needs a higher effort level.
D. The Messages API is stateless, and the app is not resending the conversation history.

5. Which two errors should a client retry with exponential backoff? (Choose two.)

A. 400 invalid_request_error
B. 429 rate_limit_error
C. 401 authentication_error
D. 529 overloaded_error
E. 404 not_found_error

6. A job submits 2,000 batch requests and writes each result into a list in the order returned, then joins the list with the input rows by index. Some rows get the wrong label. Why?

A. Batch results can arrive in any order; results must be matched by custom_id.
B. The batch exceeded the size limit.
C. Batch outputs use a different tokenizer.
D. The batch ran on a smaller model.

7. A bank requires that all inference run inside its existing AWS account, authenticated with IAM roles. Which integration best fits?

A. The first-party Claude API with a long-lived API key.
B. claude.ai Team plan.
C. Claude on Amazon Bedrock using its dedicated client class.
D. A browser extension.

8. An app asks many different questions about the same 80-page PDF over a week. The developer wants to avoid re-uploading the base64 file on every request. What should they use?

A. The Files API, referencing the uploaded file_id.
B. A longer system prompt.
C. Message Batches.
D. A higher max_tokens.

9. A legal research tool must show users exactly which passage of a source document supports each claim. Which feature fits best?

A. Extended thinking.
B. Citations enabled on the document blocks.
C. A stop sequence.
D. Tool search.

10. A prompt that produced excellent results in claude.ai gives thinner, less consistent answers when sent through the API. What is the most likely reason?

A. claude.ai applies context and a product system prompt that the API does not; the developer must supply that context.
B. The API uses a weaker model by default.
C. The API ignores system prompts.
D. API requests have a smaller context window.

11. A production app's behavior changed overnight with no deploy. Investigation shows the code references a model alias rather than an exact ID. What is the best practice going forward?

A. Use the alias with a higher effort level.
B. Pin an exact model version and upgrade deliberately after running evals.
C. Add a stop sequence.
D. Move to the Batch API.

12. Which two practices belong to good prompt configuration management? (Choose two.)

A. Store prompts in version control with review.
B. Edit prompts directly in production when quality dips.
C. Link each prompt version to its evaluation results so you can roll back.
D. Keep prompts only in developers' local notes.

13. A Python service must send 500 independent requests as fast as possible without exceeding its rate limit. What is the best approach?

A. A synchronous for loop with no delay.
B. The async client with a bounded concurrency limit, backing off on 429s.
C. One request containing all 500 tasks.
D. Disable SDK retries.

14. An app summarizes documents uploaded by the public. How should the developer structure the request?

A. Paste the document into the system prompt.
B. Put the document before the system prompt.
C. Remove the system prompt so the document is the only instruction.
D. Keep instructions in the system prompt and place the document in the user turn inside labelled tags, telling Claude to treat it as data.

15. An extraction schema includes status, which must be one of "open," "pending," or "closed." How should the field be defined?

A. As a free-text string with no description.
B. As a string with an enum of the three values and marked required.
C. As an integer.
D. Left out; Claude will infer it.

16. A team will use Claude Code to migrate a large codebase to a new framework. Which approach is best?

A. Ask for the whole migration in one prompt and merge the result.
B. Disable all permission prompts and let it run unattended on main.
C. Have Claude produce a plan first, then work in small steps, running tests after each.
D. Rewrite the code by hand and use Claude only for documentation.

17. Which requirement is an infrastructure requirement rather than a functional one?

A. "The assistant drafts replies to customer emails."
B. "The assistant classifies tickets into five categories."
C. "Responses must return within 2 seconds at the 95th percentile, in an EU region."
D. "The assistant summarizes call transcripts."

18. A developer sends an image and a question about it. Which ordering of content blocks is recommended?

A. Question first, then image.
B. Image first, then the question.
C. Image in the system prompt.
D. Order does not matter, so send two separate requests.

Model Selection and Optimization (items 19 to 27)

19. A help-desk bot includes the same 30,000-token policy manual on every request, followed by the user's question. What reduces cost most with no quality loss?

A. Prompt caching with a breakpoint after the manual.
B. Switch to the smallest model.
C. Lower max_tokens.
D. Remove the manual.

20. Caching is enabled, but cache_read_input_tokens is always 0. The system prompt begins with "Current time: {now}." What is wrong?

A. Cache breakpoints are not allowed in system prompts.
B. The changing timestamp at the start of the prefix invalidates the cache on every request.
C. The model does not support caching.
D. The 1-hour TTL must be set.

21. Which model family best suits a router that must classify 2 million short messages a day at low latency and cost?

A. Haiku
B. Opus
C. Fable
D. Any model with maximum effort

22. A team finds its Opus-tier agent too expensive. Before moving to a smaller model, what should they test first?

A. Raising effort to max.
B. Removing the system prompt.
C. Turning off streaming.
D. The same model at a lower effort level, measured against their eval.

23. The same prompt produces slightly different answers on repeated calls. What is the appropriate engineering response?

A. File a bug; LLM output should be identical.
B. Design for variability: validate outputs and measure quality with evals.
C. Retry until two outputs match.
D. Send the request twice and pick the shorter one.

24. A developer needs the exact token count of a large prompt before sending it. What should they use?

A. A generic open-source tokenizer.
B. Word count multiplied by 1.3.
C. The token counting endpoint (/v1/messages/count_tokens).
D. The max_tokens parameter.

25. Users of an analytics assistant send follow-up questions about the same cached dataset roughly every 20 minutes. Which cache setting fits?

A. Default 5-minute TTL.
B. No caching.
C. 1-hour TTL.
D. Four breakpoints at 5 minutes each.

26. A prompt includes three worked input-output examples before the real input. What is this technique called?

A. Zero-shot prompting
B. Multi-shot (few-shot) prompting
C. Prefill
D. Extended thinking

27. After upgrading to a newer model, requests that end with a partial assistant message (prefill) start returning 400 errors. What does this illustrate, and what is the fix?

A. A breaking behavior change across model releases; use structured outputs or system instructions instead, and re-run evals.
B. A temporary outage; retry.
C. A rate limit; slow down.
D. A billing problem; contact support.

Agents and Workflows (items 28 to 35)

28. An accounts-payable process always runs the same steps: extract invoice fields, check totals against the purchase order, then post to the ERP. Which design fits best?

A. An autonomous agent with all ERP tools.
B. A multi-agent system with a supervisor.
C. A workflow using prompt chaining with validation gates between steps.
D. A single prompt asking Claude to do everything.

29. A coding assistant must apply a change across an unknown number of files that it discovers as it goes. Which pattern fits?

A. Orchestrator-workers: a lead model plans and delegates file-level work dynamically.
B. Routing.
C. Prompt chaining with fixed steps.
D. Voting.

30. A team wants an agent with built-in file read, edit, bash, and search tools, running entirely on their own servers. Which option fits?

A. The SDK Tool Runner.
B. Claude Managed Agents.
C. A plain Messages API call.
D. The Claude Agent SDK, self-hosted.

31. A team wants Anthropic to run both the agent loop and a sandboxed execution environment, with versioned agent configurations. Which option fits?

A. A custom agent loop.
B. Claude Managed Agents.
C. The Claude Agent SDK.
D. LangGraph on a laptop.

32. An agent must run the project's formatter after every file edit, without exception. Where should this live?

A. A PostToolUse hook.
B. A sentence in the system prompt.
C. A sentence in CLAUDE.md.
D. A few-shot example.

33. Which two are genuine benefits of delegating work to subagents? (Choose two.)

A. Context isolation: the lead agent receives summaries instead of raw material.
B. Lower total token usage in every case.
C. Parallel execution of independent subtasks.
D. Guaranteed deterministic outputs.

34. A marketing team wants product copy translated into Japanese, then checked against a clear style rubric and revised until it passes. Which pattern fits?

A. Routing.
B. Sectioning.
C. A single zero-shot call.
D. Evaluator-optimizer.

35. A developer needs an agent framework that models the workflow as an explicit state graph with checkpoints and human-in-the-loop interrupts. Which fits best?

A. PydanticAI
B. Strands Agents
C. LangGraph
D. Claude Code headless mode

Prompt and Context Engineering (items 36 to 41)

36. Where should a support assistant's persona, tone, and refund policy be placed?

A. The system prompt.
B. Each user message.
C. A tool description.
D. An assistant prefill.

37. A prompt includes three long contracts and one question about them. Where should the question go for best results?

A. Before the contracts.
B. In the middle, between contracts.
C. At the end, after the contracts.
D. In a separate request.

38. Four hours into a session, an agent starts ignoring constraints it was given at the start. Which response best addresses this?

A. Increase max_tokens.
B. Raise the temperature.
C. Switch to the Batch API.
D. Treat it as context drift: keep goals in a task or memory file, compact history while preserving key decisions, and delegate isolated work to subagents.

39. Which feature guarantees that a response conforms to a supplied JSON schema?

A. A system-prompt instruction to "return valid JSON."
B. Structured outputs (output_config.format).
C. A stop sequence of "}".
D. Higher effort.

40. A finance summary from Claude states totals with complete confidence. What should the application do?

A. Display the totals as-is because the model is confident.
B. Ask Claude if it is sure.
C. Recompute the totals in code from source data and flag mismatches.
D. Lower the effort level.

41. Which instruction is most likely to get plain prose without markdown formatting?

A. "Write your answer in flowing prose paragraphs."
B. "NEVER USE MARKDOWN."
C. "Don't format."
D. "Markdown is bad."

Tools and MCPs (items 42 to 47)

42. Claude keeps calling search when it should call lookup_customer. What should the developer fix first?

A. Raise max_tokens.
B. Switch to a larger model.
C. Remove search.
D. Rewrite both tool names and descriptions so their purposes and when-to-use guidance are distinct.

43. Claude requests three tools in one turn. How should the results be returned?

A. Three separate user messages, one per result.
B. One user message containing all three tool_result blocks.
C. Only the first result.
D. Inside the system prompt.

44. A tool's downstream API times out. What should the tool loop send back?

A. Nothing; skip the result.
B. A fabricated default value.
C. A tool_result with is_error: true and a helpful error message.
D. A new system prompt.

45. Several internal Claude applications need the same CRM lookups, maintained by one platform team. What is the best approach?

A. Build an MCP server exposing CRM operations as tools.
B. Paste CRM exports into every prompt.
C. Rely on built-in web search.
D. Duplicate a custom tool in every app.

46. A team wants its brand writing guidelines and a review checklist applied whenever Claude drafts marketing copy. Which mechanism fits best?

A. A custom tool.
B. An MCP server.
C. A built-in tool.
D. A Skill.

47. Which MCP transport suits a server hosted centrally for hundreds of remote users?

A. stdio
B. Streamable HTTP
C. Shared clipboard
D. A local subprocess on each laptop

Security and Safety (items 48 to 51)

48. An agent summarizes web pages and also has a tool that can send email. A page contains hidden text telling it to email the user's data to an outside address. Which is the most effective mitigation?

A. Raise the temperature.
B. Add "please ignore malicious instructions" to the system prompt.
C. Treat page content as untrusted, remove or restrict the email tool for this task, and require approval for outbound messages.
D. Switch to a larger model.

49. A mobile app calls the Claude API directly with an API key embedded in the app. What is the correct fix?

A. Move calls to a backend service that holds the key in a secrets manager, and authenticate users to that backend.
B. Obfuscate the key in the app binary.
C. Rotate the key monthly but keep it in the app.
D. Put the key in the system prompt.

50. Which control reliably blocks an agent from running rm -rf even if a prompt injection asks for it?

A. A system-prompt warning.
B. A CLAUDE.md note.
C. A lower effort level.
D. A PreToolUse hook or deny permission rule.

51. Which two practices improve API key management? (Choose two.)

A. One shared key for dev and prod.
B. Separate keys and workspaces per environment, with spend limits.
C. Committing keys to a private repository.
D. Regular rotation and immediate revocation of exposed keys.

Claude Code (item 52)

52. A developer wants Claude Code to run npm test without asking, on their machine only, without changing the team's shared configuration. Which file should hold the rule?

A. The project CLAUDE.md.
B. .claude/settings.json.
C. .claude/settings.local.json.
D. A Skill.

Eval, Testing, and Debugging (item 53)

53. An agent sometimes replies "I couldn't find that order," yet the order API logs show it returned the order. Where should the developer look first?

A. The agent's trace, to check whether the tool_result was appended to the conversation correctly and was not buried in noise.
B. The model's pricing page.
C. The temperature setting.
D. The Batch API dashboard.

Answer key and rationales

# Answer Rationale

1 B Latency-tolerant, high-volume, cost-sensitive work is what the Batch API (50% off) is for. 2 D Streaming fixes perceived latency; batch and bigger caps make it worse.

3 A A max_tokens stop means truncated output; always check stop_reason first. 4 D The API keeps no state; the app must resend history each call.

5 B, D 429 and 529 are transient. 400, 401, 404 need the request fixed.

6 A Batch results are unordered; join on custom_id.

7 C Bedrock runs inside the customer's AWS account with IAM auth.

8 A Upload once to the Files API and reference the file_id.

9 B Citations tie claims to exact passages in document blocks.

10 A The API has no product system prompt; supply the context yourself.

11 B Pin exact versions; upgrade behind evals.

12 A, C Version control plus eval-linked versions enable review and rollback. 13 B Async with bounded concurrency maximizes throughput within limits.

14 D Keep trusted instructions separate from untrusted content, tagged as data. 15 B Closed sets belong in an enum, marked required.

16 C Plan first, small verified steps, tests after each.

17 C Latency and region are infrastructure requirements.

18 B Put images and documents before the question.

19 A Caching a large, repeated prefix cuts cost with no quality change.

20 B Caching is a prefix match; a changing timestamp at the top invalidates it. 21 A Haiku is built for high-volume, low-latency, low-cost work.

22 D Lower effort on the same model often holds quality and keeps one cache. 23 B LLM output is non-deterministic; validate and measure with evals.

24 C The count_tokens endpoint gives exact counts for the target model.

25 C Gaps of 5 to 60 minutes are where the 1-hour TTL pays off.

26 B Several examples before the input is multi-shot (few-shot) prompting.

27 A Prefill removal is a breaking change; use structured outputs and re-test. 28 C Fixed, predictable steps call for a workflow, not an agent.

29 A Unpredictable subtasks fit orchestrator-workers.

30 D The Agent SDK is the Claude Code harness with built-in tools, self-hosted. 31 B Managed Agents is the only option where Anthropic hosts loop and sandbox. 32 A Must-always actions belong in deterministic hooks.

33 A, C Subagents isolate context and run in parallel; they usually raise total tokens. 34 D Clear criteria plus iterative improvement is evaluator-optimizer.

35 C LangGraph models agents as state graphs with checkpoints.

36 A Stable role and policy go in the system prompt.

37 C For long documents, put the question at the end.

38 D Drift is solved by context management, not sampling or limits.

39 B Structured outputs constrain the response to the schema.

40 C Be skeptical of confident output; verify numbers in code.

41 A Positive instructions work better than shouted prohibitions.

42 D Tool names and descriptions drive tool selection.

43 B Return all parallel results in one user message.

44 C Return errors as is_error results so Claude can adapt.

45 A MCP gives a reusable, independently maintained integration.

46 D Skills package know-how and procedures loaded when relevant.

47 B Streamable HTTP serves many remote clients; stdio is local.

48 C Isolation, least privilege, and approval are enforceable; prompts are not. 49 A Never ship keys to clients; proxy through an authenticated backend.

50 D Hooks and deny rules are deterministic and injection-proof.

51 B, D Separate per-environment keys and rotation limit blast radius.

52 C settings.local.json is personal and git-ignored.

53 A Isolate integration versus model problems by reading the trace.

Scoring guide. 45 or more correct: book the exam. 38 to 44: review the chapters for your weakest domains, then retake. Under 38: work through Chapters 2 to 9 again and build a small project before retesting.