CCDV-F Practice Questions (with explanations)

~33 questions weighted toward the heavy domains (Applications & Integration, Model Selection). Answer first, explain every distractor, then check the key. This is a developer exam — expect code-shaped scenarios.

Applications & Integration (heaviest)

Q1. Why does a multi-turn conversation retain earlier context?

• A) Anthropic stores the session server-side
• B) Your code resends the full messages history each call (stateless API)
• C) The model caches it internally for 24h
• D) max_tokens preserves it

Q2. max_tokens=100 and the reply is cut off. stop_reason is:

• A) end_turn
• B) tool_use
• C) max_tokens
• D) stop_sequence

Q3. In the tool-use loop, you continue looping while:

• A) stop_reason == "end_turn"
• B) stop_reason == "tool_use"
• C) usage.output_tokens > 0
• D) the content has text

Q4. Which errors should you retry with exponential backoff?

• A) 400, 401, 403
• B) 429, 500, 529
• C) All of them
• D) None — never retry

Q5. A request that charges a card times out; you can’t tell if it succeeded. Best handling?

• A) Retry immediately
• B) Idempotency key or check state before retrying
• C) Assume success
• D) Assume failure and re-charge

Q6. You want to reduce perceived latency in a chat UI. Use:

• A) Message Batches
• B) Streaming (SSE)
• C) A bigger max_tokens
• D) Prompt caching only

Q7. 200,000 documents need offline field extraction, not user-facing. Cheapest correct approach?

• A) 200k synchronous calls on Opus
• B) Message Batches (~50% off) + Haiku + a JSON Schema + caching
• C) One request containing all documents
• D) Streaming each document

Q8. Prompt caching stopped helping after you added the current time to the system prompt. Why?

• A) Caching expires hourly regardless
• B) The cached prefix changed, invalidating it; keep it stable and put volatile content last
• C) Timestamps can’t be cached
• D) You must call a cache API first

Q9. You attach an image and text; you should place the instruction:

• A) Before the image
• B) After the image
• C) In the system prompt only
• D) It doesn’t matter and images are free

Q10. Which field lets you confirm a cache hit and monitor cost?

• A) stop_reason
• B) usage (incl. cache_read_input_tokens)
• C) id
• D) role

Q11. content[0] in a response — safe to assume it’s text?

• A) Yes, always
• B) No — it may be a tool_use (or thinking) block; inspect block.type
• C) Only for Haiku
• D) Only when streaming

Model Selection & Optimization

Q12. A high-volume, latency-critical, well-scoped classification task. Best model?

• A) Claude Fable 5
• B) Claude Opus 5
• C) Claude Haiku 4.5
• D) Whichever is newest

Q13. Complex agentic coding across an enterprise codebase. Default choice?

• A) Haiku 4.5
• B) Opus 5
• C) A non-Claude model
• D) Sonnet 3

Q14. Production app needs reproducible behavior and stable evals. Reference the model by:

• A) The alias claude-sonnet-5
• B) A pinned dated ID (e.g. claude-haiku-4-5-20251001)
• C) “latest”
• D) No model field

Q15. Structured extractor gives different output each run. First fix?

• A) Raise temperature
• B) Lower temperature toward 0
• C) Increase max_tokens
• D) Switch to streaming

Q16. Two levers that cut cost on a bulk offline job (pick the best combined answer)?

• A) Streaming + bigger model
• B) Batches + prompt caching (+ a cheaper model)
• C) Higher temperature + top_k
• D) More max_tokens

Q17. When is spending extra thinking budget justified?

• A) Simple key lookups
• B) Hard multi-step reasoning / planning / tricky debugging
• C) To reduce cost
• D) Always, on every call

Agents & Workflows

Q18. Steps are fixed and known; you need predictability and low cost. Build a:

• A) Fully autonomous agent
• B) Workflow (code-orchestrated)
• C) Multi-agent swarm
• D) Voting ensemble

Q19. Agent SDK vs Client SDK — the difference?

• A) Agent SDK is slower
• B) Agent SDK runs the agent loop for you; Client SDK means you implement the loop
• C) Client SDK can’t call tools
• D) They’re identical

Q20. Orchestrator-workers vs parallel sectioning — the distinguishing question?

• A) Which model is used
• B) Whether subtasks are decided at runtime (orchestrator) or known in advance (sectioning)
• C) Number of tools
• D) Streaming or not

Q21. Best reason to run a subtask in a subagent?

• A) It’s the only way to call tools
• B) Isolated context + restricted tools (and parallelism)
• C) It’s always cheaper
• D) It disables hooks

Prompt & Context Engineering

Q22. Most reliable way to control output format?

• A) A long prose description
• B) Few-shot examples of the exact output (+ a schema)
• C) Raising temperature
• D) Adding “be precise”

Q23. Your extractor invents a value when the source lacks it. Fix?

• A) Add “don’t hallucinate” to the prompt only
• B) Make the field nullable/optional in the schema
• C) Raise temperature
• D) Use a bigger model

Q24. Even with a 1M-token window, why not include the whole knowledge base every call?

• A) The API rejects it
• B) Cost, latency, and diluted attention — relevance beats volume
• C) It breaks streaming
• D) No reason; always include everything

Q25. First structural defense when a prompt includes untrusted web content?

• A) Trust the content
• B) Separate instructions from data — tag the content and treat it as data, not commands
• C) Raise max_tokens
• D) Use tool_choice: any

Tools & MCPs

Q26. What does the model primarily use to decide when to call a tool?

• A) The parameter names
• B) The tool description
• C) The return type
• D) The order in the list

Q27. A capability multiple internal apps all need. Best implementation?

• A) Duplicate inline tools in each app
• B) One MCP server, reused across apps
• C) A slash command
• D) A bigger prompt

Q28. Signal to the model that a tool failed so it can adapt?

• A) Return empty string
• B) is_error: true in the tool_result
• C) Raise an HTTP 500
• D) Change tool_choice

Security & Safety

Q29. An agent can send email and reads arbitrary web pages. Preventing exfiltration via a malicious page relies most on:

• A) Telling the model to ignore malicious instructions
• B) Least privilege + separating data from instructions + human approval before sending
• C) A bigger model
• D) Higher temperature

Q30. Where do API keys belong?

• A) In CLAUDE.md
• B) In a secrets manager / env vars — never in code or version control
• C) Hardcoded for convenience
• D) In the system prompt

Q31. Guarantee “never run a destructive command without confirmation.” Use:

• A) A system-prompt sentence
• B) A PreToolUse hook that blocks it
• C) A comment in CLAUDE.md
• D) Lower temperature

Claude Code + Eval/Debugging

Q32. Run Claude Code in CI and parse the result programmatically. Use:

• A) Interactive mode
• B) Headless -p with --output-format json
• C) A slash command
• D) Streaming to a terminal

Q33. Claude “forgot” an instruction after a long chat. Most likely:

• A) Model defect
• B) Integration/context issue (history grew or was truncated); reinforce/manage context
• C) Wrong model
• D) API outage

Answer key & reasoning

1. B. Stateless API; you resend history. No server-side session (A/C wrong).
2. C. Hitting the output cap → max_tokens.
3. B. Loop continues while the model keeps asking for tools.
4. B. 429/5xx/529 are transient → backoff. 4xx are your bug → fix, don’t retry.
5. B. Uncertain write → idempotency/check-state; never blind-retry a side effect.
6. B. Streaming cuts perceived latency. Batches cut cost, not latency.
7. B. Bulk offline → Batches + Haiku + schema + caching.
8. B. Volatile content invalidated the cached prefix; keep stable-first.
9. B. Put the instruction after the image; images cost tokens (not free).
10. B. usage, including cache_read_input_tokens.
11. B. A response block may be tool_use/thinking; check block.type.
12. C. Haiku — cheap, fast, high-volume, well-scoped.
13. B. Opus 5 is the default for hard agentic coding.
14. B. Pin a dated ID for reproducibility.
15. B. Lower temperature toward 0 for deterministic structured output.
16. B. Batches + caching (+ cheaper model) cut cost.
17. B. Thinking budget pays off on hard multi-step reasoning, not simple lookups (and it adds cost).
18. B. Fixed known steps → workflow.
19. B. SDK runs the loop; Client SDK = you write it.
20. B. Runtime-decided subtasks (orchestrator) vs pre-known (sectioning).
21. B. Context isolation + restricted tools + parallelism (not always cheaper).
22. B. Few-shot examples + schema beat prose for format.
23. B. Nullable field lets the model report “absent” instead of fabricating.
24. B. Relevance beats volume — cost, latency, diluted attention.
25. B. Separate instructions from data; treat tagged content as data.
26. B. The description is the interface.
27. B. One reusable MCP server, not duplicated inline tools.
28. B. is_error: true in the tool_result.
29. B. Least privilege + data/instruction separation + human approval. Prompt-pleading alone (A) is weak.
30. B. Secrets manager / env; never in code, CLAUDE.md, or VCS.
31. B. A hook enforces deterministically; prompt text only guides.
32. B. Headless -p + --output-format json.
33. B. Usually an integration/context problem, not a model defect.

Scoring guide

• 29–33: exam-ready.
• 23–28: solid; re-read the deep-dive behind each miss (weight Applications & Integration).
• < 23: another full pass, focusing on Domains 1–2 (~50% of the exam), then retake.