20 free CCAR-F (CCA-F) practice questions with explanations

1. A

   Pairs the right terminal value with the wrong iteration value. "function_call" is what other tool-calling APIs name this event, so the refund branch never fires and every tool request falls through to the reply.

2. B

   Reads naturally and matches the vocabulary of several other vendors' loops, but neither string is emitted here, so the dispatcher matches nothing and the loop stalls on its first tool request.

3. C

   "tool_result" is a real name in this API, for the block carrying a tool's output back in the next user message. It is not a stop_reason, and "stop" is not one either.

4. DCorrect

   Correct. "tool_use" is the value Claude returns when it wants a tool executed and "end_turn" the value it returns when the turn is finished, which is exactly the two-branch split the dispatcher needs.

67% of 390 answers on this site got it right on the first try.

1. A

   Removes the information gap, but shifts the reading cost onto the human: a transcript still has to be mined for the customer ID, the finding, and the proposed remedy. It restores raw data where a conclusion is needed.

2. BCorrect

   Correct. A structured handoff carries exactly what the receiving human needs to act — who the customer is, what was found, and what the agent recommends — without them replaying the conversation.

3. C

   Preserves everything the agent saw, and is useful for later debugging, but raw tool output is even denser than a transcript and buries the analysis Tier-2 needs at the top.

4. D

   Improves the customer's experience of the handoff, and is worth doing, but the apology text is addressed to the customer and carries none of the case facts Tier-2 is missing.

78% of 285 answers on this site got it right on the first try.

1. A

   PreToolUse can rewrite a call before it executes, which is the right instrument when the defect is in the arguments being sent. The epoch integer is chosen by each backend as it answers, so nothing in the outgoing call decides it.

2. BCorrect

   Correct. PostToolUse runs on the result after the call returns and before the model reads it, which is the point at which a tool's own output can be reshaped in code rather than left to the model's judgment.

3. C

   A tool's input schema is a genuine constraint and does reject malformed arguments before a call goes out. It governs the arguments Claude sends rather than the payload the tool hands back, so it never sees the value at issue.

4. D

   Rejecting a non-conforming result is deterministic and would keep the 1970 date away from the customer, but the backend returns the same representation every time, so the call is re-issued until the loop abandons a lookup that succeeded.

73% of 282 answers on this site got it right on the first try.

1. ACorrect

   Correct. A subagent handed a defined slice stops re-deriving what the coordinator already holds; scope, not memory, is what removes the duplicate calls.

2. B

   Cuts the cost of the second lookup and is worth having on a hot path, but both agents still spend a turn deciding to make it and the duplicated reasoning remains.

3. CCorrect

   Correct. Partitioning by record type is the other axis of the same decomposition: each source is read by exactly one agent, so coverage stays complete without overlap.

4. D

   Sets a sensible expectation and costs nothing to try, yet the subagent cannot see which lookups the coordinator ran, so it is guessing about state it has no access to.

5. E

   Produces one consistent answer and catches contradictions between the two agents, but it accepts the duplicated work as the price rather than removing it.

38% of 111 answers on this site got it right on the first try.

1. ACorrect

   Correct. Keyword-sensitive system prompt language creates unintended tool associations: "look up" and "customer" read as a standing directive to invoke get_customer, overriding otherwise sound descriptions.

2. B

   Would be the usual first suspect and is worth ruling out, but the stem states both descriptions are detailed and accurate, and the symptom is repetition of one tool rather than misrouting between two.

3. C

   Would explain re-querying, and context loss does cause duplicate calls in other systems, but tool results are appended to the conversation and remain visible for the rest of the turn.

4. D

   Would guarantee a tool call rather than prose, which is a real effect of that setting, but it does not explain why the same tool is chosen a second time in a single turn.

59% of 150 answers on this site got it right on the first try.

1. A

   Gives the agent more to go on than one generic string, but it puts the retry decision back on string matching, which breaks the moment someone rewords a message.

2. B

   Handles transient failures close to the source, which is genuinely good practice, but it does nothing for the policy rejections that should never be retried at all.

3. C

   A stable code is a real improvement on prose and is something the agent can match exactly rather than interpret. It exports the backend's internal taxonomy, so the agent must carry its own map of which classes are worth retrying.

4. DCorrect

   Correct. Structured error metadata lets the agent branch on isRetryable rather than infer intent from prose, and errorCategory separates transient failures from business-rule violations.

87% of 128 answers on this site got it right on the first try.

1. ACorrect

   Correct. A constrained replacement makes the account boundary part of the tool contract, so a cross-account row cannot come back however the model phrases its request.

2. B

   Sharpens selection and would reduce casual reaching for it, but the capability is unchanged: a tool that can read any account still will when the description is outweighed.

3. CCorrect

   Correct. The narrow tool only helps if the broad one is gone; leaving both in the set keeps the unconstrained path one selection away.

4. D

   Filtering the result set is deterministic and does keep the foreign rows out of the model's context, which is worth having as a backstop. The query still ran unconstrained, so the read happened and the tool can be asked for anything.

5. E

   Establishes the account early and is a sound use of forced selection, but nothing then stops a later run_query from reaching past that account anyway.

33% of 60 answers on this site got it right on the first try.

1. A

   Guaranteeing a tool call rather than conversational text is the right instrument when an agent answers from memory, and it would fix that failure. It does not say which tool, so the eighth dispute still opens on lookup_order.

2. BCorrect

   Correct. Forced tool selection names the one tool that must be called, which is how a specific call is guaranteed to run first; the rest of the dispute then proceeds on auto in the follow-up turns.

3. C

   Prompt instructions steer selection, cost nothing to add, and would carry most disputes. A stated preference competes with the model's own reading of the tools rather than deciding the call, so a fraction still goes the other way.

4. D

   Varying the tool set by turn does make the wrong opening impossible and is a real technique. It asks the caller to track dispute state and rebuild the tool set each turn, where a single tool_choice value does the same job.

71% of 122 answers on this site got it right on the first try.

1. ACorrect

   Correct. User-level configuration applies only to the user who wrote it; it is never shared through the repository, so a teammate who was never told to create it simply does not have the rule.

2. B

   Directory-level CLAUDE.md files do add scoped instructions, and diagnosing an omission there is a real exercise — but that would affect everyone editing those directories, not one new clone.

3. C

   A paths mismatch is a genuine cause of rules that never fire, and worth checking. It would not explain a convention that works for every other developer in the same files.

4. D

   Context pressure does degrade adherence in very long sessions, and shortening them helps. The pattern here is per-person rather than per-session, which points at configuration scope.

59% of 116 answers on this site got it right on the first try.

1. ACorrect

   Correct. A command file under the repository's .claude/commands/ is version-controlled, so it arrives with the code and needs no install, and its body is read when someone invokes it rather than loaded into every session.

2. B

   This is a real commands directory and the command runs perfectly for whoever placed a copy there, but the user-level path sits outside version control, so each maintainer installs it by hand and the copies drift apart.

3. C

   Path-scoped rules do load conditionally and would keep scaffold guidance out of unrelated sessions, which is half of what is being asked. A rule file is context that activates on a matching edit, not an artefact anyone invokes by name.

4. D

   The root file travels with the repository and reaches every maintainer, so the distribution half is satisfied. It is always-loaded context rather than an invocable command, so it is carried by every session and nothing answers to /atlas-scaffold.

69% of 104 answers on this site got it right on the first try.

1. A

   Restricting allowed-tools is a sound safety measure for a skill that only needs to read, and it prevents destructive edits. It does nothing about the volume of output landing in the conversation.

2. BCorrect

   Correct. context: fork runs the skill in an isolated sub-agent context, so its verbose output never pollutes the main conversation and the refactor in progress stays intact.

3. C

   An argument-hint advertises the parameter a skill takes, and an audit scoped to one package would indeed print less. It relies on the developer narrowing scope rather than isolating the output structurally.

4. D

   A personal variant in the user-level skills directory avoids affecting teammates and can be tuned per developer. The output still lands in whichever session invokes it.

64% of 108 answers on this site got it right on the first try.

1. A

   Three designs make the tradeoffs concrete and can be compared quickly. It builds the thing three times to elicit requirements a conversation would settle, and the comparison has no agreed criteria to judge against.

2. BCorrect

   Correct. The interview pattern surfaces the decisions nobody has made — invalidation, deploy-time misses, tolerable staleness — before any code exists that quietly assumes an answer to each of them.

3. C

   Test-first is the right driver when the expected behaviour is known and can be written down, as it is for a normalizer with fixed output shapes. Here the tests would encode guesses, and passing them would prove nothing.

4. D

   Shipping something conventional is fast and puts a real artefact in front of reviewers. Review comments arrive after the design is chosen, which is the expensive moment to learn the team wanted different staleness rules.

59% of 108 answers on this site got it right on the first try.

1. A

   This does dial down flag volume, and lower volume is part of what reviewers want. But "be conservative" and "high-confidence" are the same vague instruction in a quieter register: the model still has no definition of what counts as questionable, so the flags that survive are an arbitrary subset rather than a more accurate one.

2. B

   Certainty ratings give reviewers an ordering, which is a real improvement to queue triage. They do not improve the underlying judgment, though - a model that cannot tell a genuine problem from a normal value will rate its bad flags confidently too, so the top of the sorted queue stays polluted.

3. CCorrect

   Correct. Explicit categorical criteria with concrete examples define what a flag means, so the model is classifying against a rule rather than guessing at "questionable." That is what moves precision, and it is what restores reviewer trust in the flags that remain.

4. D

   A second pass and a disagreement filter would catch genuine instability in the extraction, and it is a reasonable consistency check. It costs a full extra call per invoice and still measures agreement rather than correctness: two passes reading the same vague instruction can agree on the same wrong flag.

79% of 100 answers on this site got it right on the first try.

1. A

   Labelling the weak category is honest and keeps some payment_terms coverage alive, which has value. It also leaves the noise in the queue and asks reviewers to do the discounting themselves - the behaviour they have already shown they will not do, since they are skimming past all three today.

2. B

   A uniform threshold rise would reduce total queue volume, which addresses the reviewers' workload complaint. It suppresses the accurate categories at the same rate as the inaccurate one, so you lose real findings from vendor_name and stated_total to fix a problem that lives in only one category.

3. CCorrect

   Correct. The damage from a high false-positive category is not confined to that category - it teaches reviewers that flags in general are noise. Disabling the one bad category preserves the trust the two accurate ones have earned, and costs only the payment_terms coverage you were not getting value from anyway.

4. D

   Stopping everything guarantees no further false positives reach reviewers, which is a defensible way to protect trust. It also discards two categories that are working, leaves invoices unchecked for a sprint, and treats a localised precision problem as though the whole rule set were suspect.

70% of 104 answers on this site got it right on the first try.

1. A

   Resubmitting failed custom_ids is the documented way to handle batch failures and is worth planning for anyway. A layout the prompt has never met fails as a class rather than at random, so most of the 200 come back for a second paid run.

2. BCorrect

   Correct. Refining the prompt on a sample before batch-processing a large volume is what raises the first-pass success rate, and a dozen sample manifests cost a fraction of the resubmission a bad first night would force.

3. C

   A tighter schema does keep malformed extractions out of the AP ledger and is good hygiene on any new document type. Rejection is not extraction: the batch is paid for in full and every rejected manifest still has to be worked afterwards.

4. D

   Synchronous calls surface each failure immediately and would certainly expose the new layout fast. It forgoes the batch saving for a week on volume that has no latency requirement, to learn what a sample of a dozen answers in an hour.

63% of 87 answers on this site got it right on the first try.

1. A

   Covering observed phrasings guarantees those exact cases are handled, and for a fixed vocabulary that would be enough. Manifest phrasing is open-ended, so pure pattern coverage produces a lookup table: the model matches the demonstrated strings and is back to guessing on the fourth phrasing that shows up next month.

2. B

   Placement near the document is a sound prompt-construction habit and keeps the examples salient. It changes where the model reads the examples, not what they teach - three examples that only show inputs and outputs teach the same thing wherever they sit in the prompt.

3. C

   Showing the full schema output does reinforce the overall shape of a correct extraction, which is useful when several fields are unstable. Here only quantity is unstable, and padding each example with the other fields dilutes the signal about unit choice rather than sharpening it.

4. DCorrect

   Correct. Examples that show the reasoning - why cases rather than pallets was the right unit here - teach the judgment rather than the string. That is what lets the model generalise to novel phrasings instead of matching pre-specified cases.

47% of 107 answers on this site got it right on the first try.

1. ACorrect

   Correct. Long inputs are processed reliably at the beginning and the end, so a summary of all three blocks placed first puts the billing findings where they will be read.

2. B

   Keeps the freshest work near the front and is a sensible ordering rule in a long session, but it only moves which block occupies the middle; whichever one lands there is omitted next.

3. C

   Cuts the prompt down and would help if the failure were volumetric. It is positional, so re-fetching returns the same findings to the same weak position, at the cost of an extra call.

4. DCorrect

   Correct. Explicit section headers organising the detailed results are the guide's second mitigation for position effects, giving the middle material structure the model can locate.

5. E

   Prevents a long resolution message being cut off mid-sentence, a real failure elsewhere. Here the reply completes; it simply never contains the billing material in the first place.

85% of 46 answers on this site got it right on the first try.

1. A

   Is the right instinct when policy is silent on a request, and it does prevent a wrong-account refund, but handing off a question the customer can answer in one turn spends human capacity needlessly.

2. B

   Involves the customer rather than guessing silently, which is better than the current behavior, but it anchors on a guess and invites a distracted customer to confirm someone else's order.

3. C

   Replaces one heuristic with a more defensible one, and completeness scoring is useful for ranking search results, but any heuristic still resolves identity without evidence from the customer.

4. DCorrect

   Correct. Multiple matches call for clarification, not selection. One additional identifier resolves the ambiguity from authoritative input and keeps the case in first-contact resolution.

84% of 98 answers on this site got it right on the first try.

1. A

   Keeps the workflow alive, which is the improvement being sought, but it disguises a failure as a valid empty result, so the coordinator reports "no billing issues found" for an issue it never saw.

2. B

   Ensures a human eventually sees the case and is appropriate once options are exhausted, but it escalates the whole ticket over one unavailable subquery instead of resolving what is already resolvable.

3. CCorrect

   Correct. Structured error context lets the coordinator decide intelligently: keep the substantiated claim, annotate the billing gap, and choose whether an alternative path is worth attempting.

4. D

   Delegates recovery upward, and backoff is the right response to a transient failure, but the subagent already retried twice, and the coordinator has no information to decide differently.

71% of 85 answers on this site got it right on the first try.

1. A

   /memory is a real command, and its name makes it sound like a context-pressure tool. What it actually does is report which memory files are loaded, for diagnosing inconsistent behaviour across sessions.

2. BCorrect

   Correct. /compact is named for exactly this state: reducing context usage during an extended exploration session once the window has filled with verbose discovery output.

3. C

   --resume does continue a named investigation across working sessions, which is useful here on other days. Resumption restores prior context rather than shrinking it, so it does not relieve a full window.

4. DCorrect

   Correct. A scratchpad file of key findings, referenced when answering later questions, is the named counter to context degradation, because the findings live outside the window.

5. E

   context: fork is real, but it is SKILL.md frontmatter that runs a skill in an isolated sub-agent context. It is set on a skill, not on a running session, so it cannot be applied to this one mid-migration.

49% of 39 answers on this site got it right on the first try.