CCAR-F Practice Questions (with explanations)

~30 scenario questions in the style of the exam. Cover the answer, then reveal. For every one, force yourself to say why the wrong options are wrong — on this exam the distractors are “works but worse,” and spotting that is the skill. Answers and reasoning follow each block.

Domain 1 — Agentic Architecture & Orchestration

Q1. A pipeline always runs the same three steps: extract entities → validate against a DB → format a report. Best pattern?

• A) Autonomous agent with dynamic decomposition
• B) Prompt chaining with a validation gate
• C) Orchestrator-workers
• D) Voting with best-of-5

Q2. You must fix a bug that may touch an unknown number of files; the coordinator should decide what to delegate at runtime. Pattern?

• A) Parallelization (sectioning)
• B) Routing
• C) Orchestrator-workers
• D) Prompt chaining

Q3. You want to embed an agent loop with file/command tools and context management inside your Python backend without writing the loop yourself. Choose:

• A) Client SDK (Anthropic API SDK)
• B) Claude Agent SDK
• C) Managed Agents
• D) Claude Code CLI, interactive

Q4. Which is the strongest reason to give a subtask its own subagent?

• A) It always reduces token cost
• B) It isolates context and can restrict the tool set, keeping the main thread clean
• C) The model can’t call tools otherwise
• D) It disables the agent loop

Q5. Incoming messages are clearly “billing,” “technical,” or “account.” Cheapest reliable design?

• A) One giant prompt handling all three
• B) Routing to three specialized handlers (and cheaper models for easy classes)
• C) Orchestrator-workers spawning subagents per message
• D) Evaluator-optimizer loop

Domain 2 — Claude Code Configuration & Workflows

Q6. Coding conventions that must apply to everyone who clones the repo go in:

• A) ~/.claude/CLAUDE.md
• B) CLAUDE.local.md
• C) Root CLAUDE.md, committed
• D) settings.local.json

Q7. A rule that must always hold (“never run terraform destroy without confirmation”). Most reliable enforcement?

• A) A sentence in the system prompt
• B) A PreToolUse hook that blocks it
• C) A note in CLAUDE.md
• D) A slash command

Q8. You want Claude Code to research and propose a big refactor but make no edits until you approve. Use:

• A) tool_choice: none
• B) Plan mode
• C) A subagent
• D) Voting

Q9. CLAUDE.md vs settings.json:

• A) Both store secrets
• B) CLAUDE.md = instructions/context; settings.json = behavior/permissions
• C) settings.json = coding style; CLAUDE.md = tool permissions
• D) They are interchangeable

Q10. The team should get two MCP servers automatically on clone. Configure them at which scope?

• A) local
• B) user
• C) project (committed .mcp.json)
• D) enterprise only

Domain 3 — Prompt Engineering & Structured Output

Q11. After 40 turns, Claude starts ignoring a rule that is still in the system prompt every request. Best explanation + fix?

• A) The API dropped the system prompt; resend it — it isn’t being sent (it is)
• B) Attention to it weakens as history grows; reinforce the rule at a breakpoint
• C) The model has server-side memory that expired
• D) Lower max_tokens

Q12. You need every response to be a call to your log_decision tool. Set:

• A) tool_choice: "auto"
• B) tool_choice: {"type": "any"}
• C) tool_choice: {"type": "tool", "name": "log_decision"}
• D) tool_choice: "none"

Q13. Most reliable way to guarantee machine-parseable output?

• A) Ask “reply only in JSON” in the system prompt
• B) Provide a JSON Schema via output_config.format (constrained decoding)
• C) Add “no prose” to every user turn
• D) Post-process with regex

Q14. Your extractor sometimes invents a phone number when the document has none. Best fix?

• A) Raise temperature
• B) Make phone nullable/optional in the schema so “absent” is representable
• C) Add “don’t make things up” to the prompt
• D) Switch to a bigger model

Q15. A structured extraction passes JSON-Schema validation but the due_date is wrong. What does this teach?

• A) Schemas are useless
• B) Syntactic validity ≠ semantic correctness; add semantic checks + provenance
• C) Always retry blindly
• D) Use prose instead of schema

Q16. To control the format of answers, which is usually most effective?

• A) A long prose description of the format
• B) Two or three few-shot examples of the exact output
• C) Raising max_tokens
• D) Setting tool_choice: none

Domain 4 — Tool Design & MCP

Q17. The three MCP primitives are controlled by, respectively:

• A) Tools—user, Resources—model, Prompts—app
• B) Tools—model, Resources—application, Prompts—user
• C) All three—the model
• D) Tools—app, Resources—user, Prompts—model

Q18. A refund-policy PDF the agent should consult before deciding eligibility should be exposed as a:

• A) Tool
• B) Resource
• C) Prompt
• D) Hook

Q19. An MCP tool is annotated readOnlyHint: true. Can you auto-approve it as safe?

• A) Yes — annotations are authoritative
• B) No — annotations are untrusted hints; enforce trust on the client (permissions/sandbox/approval)
• C) Only for local servers
• D) Only if idempotentHint is also true

Q20. A search_orders tool would return 50,000 rows. Best output design?

• A) Return all rows so the model has everything
• B) Return a page + cursor; let the model request more on demand
• C) Return a prose summary of all rows
• D) Refuse if > 1000 rows

Q21. Difference between a JSON-RPC protocol error and a tool result with isError: true?

• A) None
• B) Protocol error = plumbing/transport failure; isError = the tool ran but its task failed (feed to the model to adapt)
• C) isError means the server crashed
• D) Protocol errors are shown to the user only

Q22. You have 150 tools; the model keeps mis-selecting. Best remedy?

• A) Put all 150 in every request with longer names
• B) Progressive availability — expose a relevant subset per phase; use tool search / list_changed
• C) Remove tool descriptions to save tokens
• D) Force tool_choice: any

Domain 5 — Context Management & Reliability

Q23. Complete the principle: “The model sees a request, not your ___.”

• A) prompt
• B) database
• C) tools
• D) schema

Q24. A 3-day support conversation must remember the agreed resolution and preferences. Best mechanism?

• A) Resend the full 3-day transcript each turn
• B) A structured state object (decisions/preferences/facts) updated and re-injected
• C) Rely on the model’s memory
• D) A sliding window of the last 2 turns

Q25. A charge-card tool call times out; you don’t know if it succeeded. Correct handling?

• A) Immediately retry the charge
• B) Use an idempotency key or check state before retrying; treat “uncertain” as its own state
• C) Assume success and continue
• D) Assume failure and refund

Q26. You have a 1M-token window, so you decide to include the entire knowledge base in every request. Problem?

• A) None — bigger context is always better
• B) Higher cost, higher latency, and diluted attention; relevance beats volume
• C) The API will reject it
• D) It disables caching only

Q27. Your system prompt embeds the current timestamp each call, and prompt-cache hit rate is near zero. Why?

• A) Caching is disabled by default
• B) The cached prefix changes every request, invalidating the cache; move volatile content to the end
• C) Timestamps are not cacheable data types
• D) max_tokens is too low

Q28. 50,000 historical invoices need field extraction; it’s not user-facing. Cheapest correct approach?

• A) 50,000 synchronous calls
• B) Message Batches API (async, ~50% cheaper) + prompt caching + a JSON Schema
• C) One request with all 50,000 invoices
• D) Streaming responses

Q29. An extraction pipeline keeps making the same mistake on a field. Best long-term response?

• A) Retry each failure a second time
• B) A feedback loop: fix upstream (schema, tool description, validation) so the class of error stops
• C) Raise temperature
• D) Ignore it below a threshold

Q30. For a high-stakes single answer where correctness matters more than cost, which raises reliability?

• A) Lower max_tokens
• B) Voting / best-of-N with aggregation, plus verification in the loop
• C) Remove the system prompt
• D) Disable tools

Answer key & reasoning

1. B. Fixed known steps → prompt chaining; a gate validates before formatting. An agent (A) adds needless nondeterminism; C/D are overkill for a fixed sequence.
2. C. Subtasks decided at runtime by a coordinator = orchestrator-workers. Sectioning (A) is for pre-known independent partitions.
3. B. Agent SDK runs the loop in your process. Client SDK (A) means you write the loop; Managed Agents (C) is hosted/async; CLI (D) is interactive.
4. B. Isolation + restricted tools + clean main context. It does not always cut cost (A) — coordination has overhead.
5. B. Routing with cheaper models for easy classes is the classic cost-efficient answer.
6. C. Team-wide, shared → committed root CLAUDE.md. A/B are personal; D is personal behavior settings.
7. B. Must-always-hold → deterministic hook, not probabilistic prompt text.
8. B. Plan mode = research/propose, no mutations until approval.
9. B. Instructions/context vs behavior/permissions. Never store secrets in CLAUDE.md.
10. C. Project scope (committed .mcp.json) ships to everyone on clone. local/user are per-machine.
11. B. It is still sent; attention decays with length. Reinforce at a breakpoint. There is no server-side memory (C is wrong on the API model).
12. C. Force a specific named tool. any (B) forces some tool, not that one.
13. B. Constrained decoding via schema. Prose requests (A/C) can be violated; regex (D) is a band-aid.
14. B. Nullability lets the model say “absent” instead of fabricating. Prompt pleading (C) is weaker.
15. B. Valid-shape ≠ correct-value; add semantic validation + provenance.
16. B. Few-shot examples control format better than prose.
17. B. Tools—model, Resources—application, Prompts—user. Memorize this.
18. B. Reference material to consult = resource.
19. B. Annotations are advisory; enforce trust on the client.
20. B. Paginate on demand; never dump 50k rows into context.
21. B. Plumbing failure vs task failure; the latter is fed to the model to adapt.
22. B. Progressive availability / tool search; the visible tool set is a design choice.
23. B. “…not your database. You decide what to include.”
24. B. Structured state object beats replaying transcript or trusting model memory.
25. B. Idempotency / check-then-act; never blind-retry an uncertain write.
26. B. Relevance beats volume; big windows still cost, slow, and dilute attention.
27. B. Volatile content in the cached prefix invalidates the cache; keep stable-first, volatile-last.
28. B. Batches (async, ~50% off) + caching + schema. Bulk, non-interactive → Batches.
29. B. Feedback loop fixes the error class upstream; retrying (A) just re-hits it.
30. B. Voting/best-of-N + in-loop verification trade cost for reliability.

Scoring guide

• 27–30: exam-ready. Do a light review of any missed domain.
• 21–26: solid; re-read the deep-dive files behind your misses.
• < 21: do another full read pass, focusing on Domains 1 and 4 (the heavy, concept-dense ones), then retake.