CClaude Cert Prep
Practice · 51 questions

CCA-F mock exam

51 scenario-based questions built from the exam's most-confused concepts — the ones where two mechanisms both look right. Every option is plausible and roughly the same length, so you have to reason, not pattern-match. The set ends with an architect trap section of cross-domain deep cuts. Answer as many as you like (unanswered count as incorrect), then get the explanation for every choice.

Architect-level reasoning Distractors = real misconceptions Per-domain scorecard
How to use it · Commit to an answer before revealing the key. When you review, read the explanation for the options you didn't pick too — the distractors encode the exact misconceptions the real exam exploits. This is a study quiz, not the graded platform exam.
0 of 51 answered
  1. D1 · Q1Coordinator vs subagent
    A customer-support agent handles billing questions end-to-end. Product asks it to also run a monthly 40-page contract audit that reads every clause, cross-checks a pricing table, and produces a findings report. You worry the audit's long, noisy tool output will crowd the support conversation's context. What is the soundest orchestration change?
  2. D1 · Q2Sequential vs parallel Task calls
    A research coordinator must gather evidence from four independent sources — a docs index, a web search, an internal wiki, and a metrics API — before it synthesizes an answer. None of the four depends on another's output. How should it dispatch the four retrieval tool calls to be both fast and well-formed?
  3. D1 · Q3Handling stop_reason (pause_turn)
    An agent using a server-side web-search tool returns a response with stop_reason "pause_turn" and a trailing server_tool_use block; the answer is clearly unfinished. Your loop currently only branches on end_turn, tool_use, and max_tokens. What is the correct way to let the search complete?
  4. D1 · Q4Hooks vs tool use for enforcement
    A code-review agent has a git_push tool and a run_tests tool. Policy is absolute: it must never push a branch while tests are failing, even if the model is convinced the failure is unrelated. Where should this guarantee live in the architecture?
  5. D1 · Q5Resolve-or-escalate terminal guarantee
    A tier-1 support agent must guarantee that every conversation ends in exactly one terminal state: resolved, or escalated to a human with a handoff. In load testing, some sessions end with the agent giving up mid-task — apologizing and stopping without resolving or escalating. What design most directly restores the terminal guarantee?
  6. D1 · Q6Handoff packages preserving authorization state
    A tier-1 agent verified the caller's identity and confirmed they are authorized to modify the account, then hands the conversation to a tier-2 refund agent via a handoff package. The tier-2 agent immediately tries to re-run identity verification and stalls because the caller is no longer on a live channel. What does a well-formed handoff package most need to carry?
  7. D2 · Q7tool_choice
    You are building a support agent that must ALWAYS return a structured triage object on every turn, but the model may pick from three different triage tools depending on the ticket. During testing you notice that when you force output, the model stops emitting its usual reasoning text before the tool call. Which tool_choice configuration matches this requirement and explains the missing reasoning?
  8. D2 · Q8mcp_resources_vs_tools
    An MCP server exposes a company's product catalog. The team wants Claude to read catalog entries as reference context that a human selects and attaches, and separately to place orders that Claude decides to invoke on its own. How should these two capabilities be modeled on the MCP server?
  9. D2 · Q9structured_tool_errors
    A data-extraction tool sometimes hits a rate limit that clears in a few seconds. Currently the tool throws and the whole request fails. You want Claude to notice the failure and retry sensibly instead of aborting. What is the correct way to return the failure?
  10. D2 · Q10mcp_server_scope
    A team wants a Postgres MCP server available to everyone who clones the Claude Code repo, with each developer supplying their own connection string rather than a shared committed secret. Which setup meets both goals?
  11. D2 · Q11hooks_vs_tool_use
    In a team Claude Code setup, every file the agent writes must be run through the repo's formatter, and this must happen deterministically no matter what the model decides. A developer proposes adding a 'format the file' tool and instructing the model to always call it after Write. Why is a hook the better mechanism?
  12. D2 · Q12sequencing_dependent_calls
    An agent must look up a customer ID from an email, then fetch that customer's open invoices using the returned ID. In one assistant turn the model emitted BOTH tool calls in parallel, and the invoice call arrived with a null ID. What is the correct fix?
  13. D3 · Q13CLAUDE.md vs hook (deterministic enforcement)
    Your team requires that Prettier reformats every file Claude Code edits, with zero tolerance for an unformatted file ever being committed. A senior engineer added 'Always run Prettier after editing any file' to the project CLAUDE.md, but unformatted diffs still slip through occasionally. What is the correct fix to make formatting truly non-negotiable?
  14. D3 · Q14Path Rules via globs vs CLAUDE.md scope
    In a large monorepo, every test file (matching */.test.ts anywhere in the tree) must follow a specific mocking convention, but that guidance is noise for the hundreds of non-test files. You want the convention injected only when Claude touches a test file, not loaded into context for all work. Which mechanism fits best?
  15. D3 · Q15Skills vs slash commands vs hooks (unified commands)
    During onboarding you want a reusable, multi-step 'generate release notes' procedure that new hires can invoke on demand as /release-notes, and that Claude can also load on its own when it recognizes the task. You want one artifact that serves both the explicit command and the automatic-loading behavior. What should you build?
  16. D3 · Q16Permission precedence (deny > ask > allow)
    Your settings allow Bash(npm:*) so routine npm commands run without prompting. Security now mandates that npm publish must never execute from Claude Code under any circumstance, not even with a confirmation prompt, while the rest of npm stays frictionless. What single change enforces this correctly?
  17. D3 · Q17Plan mode vs direct execution vs explore
    You are handed a vaguely specified refactor that spans an unfamiliar codebase. Before any file is touched, you want Claude to investigate the relevant code, then produce a concrete step-by-step approach for you to approve, guaranteeing it makes no edits until you sign off. Which working mode matches this need?
  18. D3 · Q18@-reference vs CLAUDE.md; context hygiene
    A 2,000-line API style guide is relevant to maybe one task in ten. A teammate pasted its entire contents into the project CLAUDE.md so it is 'never forgotten,' but sessions now start with bloated context and higher token costs. What is the cleaner way to keep the guide available without loading it every session?
  19. D4 · Q19structured-output reliability ladder
    Your invoice-extraction service parses supplier PDFs into a fixed set of fields and writes them straight into an accounting ledger, so a single malformed response corrupts a downstream posting. You currently ask the model in the prompt to 'reply with only JSON' and roughly 3% of responses include prose or a trailing comma. Which change gives the strongest structural guarantee that every response conforms to your field shape?
  20. D4 · Q20tool_choice to guarantee invocation
    You expose a single classify_ticket tool that returns a structured label, but in about 8% of turns the model answers in plain prose and skips the tool entirely, breaking your pipeline that expects a tool_use block every time. Which adjustment most directly forces the tool to be invoked on every request?
  21. D4 · Q21schema design to avoid fabrication
    Your contract-parsing schema marks termination_date and governing_law as required strings, and reviewers notice the model invents plausible-looking dates and jurisdictions when a contract simply omits them. You need extraction to admit uncertainty instead of guessing. Which schema change best prevents fabrication?
  22. D4 · Q22few-shot for messy real-world formats
    A vendor sends dates as '3/4/25', '04 Mar 2025', and 'Mar-4', and your extractor must always emit ISO 8601. Your schema already constrains the field to a string, but the normalized values are still inconsistent because the schema cannot express the transformation rule. What is the most effective addition?
  23. D4 · Q23Messages API vs Message Batches
    A nightly job re-summarizes 400,000 archived support threads with no user waiting on the results, and finance wants the cost cut. Your current code loops over the synchronous Messages API, which is expensive and takes hours of blocking calls. Which approach best fits this offline, cost-sensitive workload?
  24. D4 · Q24structured error feedback loops
    Your extraction service validates each tool response against a Zod schema and, on failure, sends the model back a retry. Right now the retry just says 'that was invalid, try again', and the second attempt fails almost as often. How should you design the feedback to make retries converge?
  25. D5 · Q25stateless API state
    You built a customer-support assistant directly on the Claude Messages API. In turn 6 the model no longer 'remembers' a preference the user stated in turn 2, even though your logs show turn 2 was handled correctly. Your backend stores each request and response in a database keyed by conversation ID. What is the most likely root cause?
  26. D5 · Q26subagent isolation vs bigger context window
    An agent explores a large monorepo to answer one architecture question. A single main thread that reads dozens of files fills its context with file dumps and starts losing track of the original question. A teammate proposes 'just use a bigger context window.' Why is delegating the exploration to subagents usually the better design here?
  27. D5 · Q27scratchpad files vs targeted reads
    During a long multi-step data-migration task, an agent must track roughly 40 intermediate decisions and revisit a few of them much later. Loading every prior file and note back into the conversation each step is bloating context and slowing the agent. What is the soundest context-management approach?
  28. D5 · Q28resume vs compact vs new session
    An engineer paused a coding session yesterday mid-refactor. Today they want to keep going with all of yesterday's file edits, reasoning, and decisions intact, and the session's context is not yet near its limit. Which action best fits this goal?
  29. D5 · Q29retry vs escalation
    An automated invoice-extraction agent hits a supplier's API and receives a 400 response whose body reads 'invalid tax_id format.' Its default policy is to retry failed calls up to three times with backoff. What should the agent actually do, and why?
  30. D5 · Q30confidence-based review routing
    An extraction pipeline processes 10,000 documents a day and self-reports a calibrated confidence score per document. A human review team can only inspect about 500 per day. The team currently reviews a random 5% sample. How should review be routed to catch the most real errors within the same budget?
  31. Architect trap section
    D1 · Q31Subagent state
    A coordinator agent spawns a subagent to reconcile an invoice against a purchase order, but the subagent replies that it cannot find the PO number or the approver's authorization. The coordinator holds both in its own conversation. In production this fails intermittently. What is the most appropriate fix?
  32. D3 · Q32PostToolUse normalization
    A pricing tool returns timestamps in a mix of Unix epoch, RFC 822, and localized strings, plus pages of debug logging. The model keeps mis-ordering events and miscomputing durations from this output. What is the most appropriate fix?
  33. D1 · Q33Context fork
    A code-audit skill greps a large monorepo and pulls hundreds of matching lines into the main conversation, after which the assistant loses track of the user's original request. You only need the skill's final summary of findings, not its raw search. What is the most appropriate fix?
  34. D3 · Q34Headless CI flags
    A CI job must run Claude Code non-interactively, emit a machine-parseable result your pipeline can validate against a fixed shape, and never exceed a per-run spend limit or block on a prompt. Which invocation is most appropriate?
  35. D3 · Q35Structural enforcement
    A support agent is told in its system prompt never to issue refunds above $500, yet audits keep finding $700 and $900 refunds it approved when a customer argued convincingly. Compliance needs this to stop entirely. What is the most appropriate fix?
  36. D1 · Q36Plan mode
    You ask the agent to rename a core module referenced across roughly 80 files and migrate its database column, and a senior engineer must approve the approach before anything is touched. What is the most appropriate way to run this?
  37. D1 · Q37Task orchestration
    An agent must fetch and summarize three unrelated vendor pages, then draft one comparison report that cites all three summaries. It currently runs every step one after another and is slow. How should the Task calls be structured?
  38. D2 · Q38Implementation leakage
    A routing agent exposes a customer_lookup tool whose description reads 'Queries the Postgres RDS customer table through the internal GraphQL gateway.' Support requests that should hit this tool are increasingly misrouted to a generic web_search tool instead. What is the best fix to the tool definition?
  39. D4 · Q39isError semantics
    A coordinator agent calls a search tool that returns isError:false with an empty results array, meaning the query ran fine but matched no records. The coordinator treats this as a failure, retries the identical call five times, and then aborts the entire workflow. What is the best fix?
  40. D2 · Q40Syntax valid vs semantically correct
    A resume-parsing tool uses strict structured output, so every response conforms to the JSON schema. On resumes with no phone number listed, the model still emits a plausible-looking phone number. The shape is always valid, but values are sometimes fabricated. What best addresses the root cause?
  41. D4 · Q41Batch API myth
    A team runs a high-volume nightly enrichment job that needs multi-turn tool calls. They rule out the Message Batches API 'because it can't do tool use' and keep everything on the synchronous API, paying full price and hand-building a rate-limit queue. A separate interactive chat feature also runs on the sync API. What is the best correction?
  42. D4 · Q42tool_choice auto vs any vs forced
    A workflow must first fetch a customer's plan tier, then call a generate_quote tool that requires that tier as an input. To guarantee structured output, an engineer sets tool_choice to force generate_quote on the very first turn. The call now fails or hallucinates the tier. What is the best design?
  43. D2 · Q43Interview / clarification pattern
    A user tells an operations agent 'clean up the old stuff in the bucket.' The request is highly ambiguous: which bucket, what age counts as old, and whether to archive or permanently delete. The agent currently guesses and deletes files. What pattern best prevents costly mistakes?
  44. D4 · Q44Tool-set focus and built-in vs MCP
    An agent is configured with 40 tools, including a built-in web_search and a specialized MCP tool internal_kb_search for the company knowledge base. For internal-policy questions the model keeps defaulting to web_search and returning generic public answers. What is the best fix?
  45. D5 · Q45Case Facts persistence
    A long-running support session compresses its history into a rolling summary that is rewritten each turn. Over time the entry "$32.40 refund issued on 2026-03-14" degrades to "about thirty dollars refunded recently," and later replies quote the wrong amount. What is the most robust fix?
  46. D5 · Q46Lost-in-the-middle positioning
    A hard constraint, "never issue store credit for orders older than 90 days," sits in the middle of a 40-turn transcript. The agent repeatedly violates it even though the full transcript fits comfortably within the context window. What best explains and fixes this?
  47. D2 · Q47Escalation triggers vs self-reported confidence
    An autonomous account-management agent is designed to hand off to a human "whenever the model reports it is not confident." In production it escalates erratically on trivial cases yet sails through genuinely risky, irreversible actions. What is the sounder escalation design?
  48. D2 · Q48Human-review routing via calibrated confidence
    A batch document-extraction pipeline must decide which of thousands of extracted fields a human should double-check. Unlike an escalation trigger, this is an offline quality-review decision. Which routing policy makes the best use of limited reviewer time?
  49. D5 · Q49Structured errors vs raw exceptions
    An MCP tool wrapping a payments API catches nothing, so a raw Python stack trace propagates back to the model as the tool result. The agent then either retries blindly in a loop or abandons the task entirely. What is the best remedy?
  50. D2 · Q50Independent review vs self-review
    In a CI workflow, an agent generates a code change and then, in the same session holding all of its generation context, reviews its own diff and nearly always approves it. Subtle bugs slip through. How should the review step be restructured?
  51. D5 · Q51Stateless Messages API
    A team building on the Messages API passes a conversation/session id and assumes the server retains prior turns, so each request sends only the newest user message. Replies keep losing all earlier context and prior tool results. What is the correct mental model?
Back to all study guides