Free CCAR-F mock exam, timed, with a reason for every option

Answer as you read

The CCAR-F paper: every question

20 items, each tagged with its official domain. Work through the paper one question at a time: pick an option to check it, or ask for the key, and read why each option holds or fails. Nothing is saved. For a scored sitting against the clock, use the start modes above.

1. Question 1Domain 1: Agentic Architecture & OrchestrationSingle answer

   Ostrander Freight runs a customs-clearance agent, with an agentic loop the team wrote directly against the Anthropic API, that calls tariff_lookup and shipment_status tools before drafting a declaration. Traces show the same tariff_lookup call, with identical arguments, repeated three to five times per run, and declarations that ignore the tariff the tool actually returned. Reading the loop, the engineer finds that each iteration sends the original user message plus a one-line note saying the tool ran, rather than the tool's output. Only the loop code is within the team's remit; the tools are owned elsewhere. What should the engineer change?

   Select the best answer.

   AAdd a system prompt rule stating that the model must never call a tool twice with the same arguments, and have the loop reject any repeated call by returning an error string instead of running it.

   BAfter executing each requested tool, append the tool_use block and a matching tool_result carrying the tool's full output to the conversation, and send that complete history on the next iteration.

   CCache each tool's output in the loop keyed by its arguments, so a repeated call is served from the cache without a network request and the run completes faster.

   DReplace model-driven tool selection with a fixed sequence coded in the loop, calling tariff_lookup once and shipment_status once before a single final model request.
   Show the answer
2. Question 2Domain 1: Agentic Architecture & OrchestrationSelect 2

   Tallis University's research service uses a coordinator with search, document-analysis and synthesis subagents. Faculty reviewers found that 30 percent of reports miss an aspect the request named, and in the coordinator's logs the analysis subagent's output goes straight to synthesis without the coordinator seeing it, so a failed analysis pass produces a report with a silent gap. The service must reach full coverage of the requested aspects before it is offered to doctoral students. Which two changes should the service team make? (Select two.)

   Select exactly 2. 0/2 selected.

   ARaise the synthesis subagent's max_tokens so a longer report has room for every aspect the request named, and instruct it to write a section for each one.

   BHave the coordinator evaluate each synthesis output against the requested aspects, re-delegate targeted searches and analysis for any gap, and re-run synthesis until coverage is sufficient.

   CRun two search subagents in parallel on the same query and merge their results so that the analysis subagent receives a wider set of sources for every request.

   DRoute the analysis subagent's output back through the coordinator instead of directly to synthesis, so the coordinator can detect a failed or partial pass before synthesis begins.

   EGive the synthesis subagent a longer, numbered procedure listing every research step in order, so a single pass covers all aspects without needing a second round.
   Check answerShow the answer
3. Question 3Domain 1: Agentic Architecture & OrchestrationSelect 2

   Marlowe and Vance LLP designed a due-diligence coordinator on the Claude Agent SDK with AgentDefinitions for a registry subagent and a data-room subagent. In production, where allowedTools lists only the firm's MCP tools and Read, the coordinator never spawns either; it answers every matter itself and runs out of context on large data rooms. When an engineer tested a configuration that permits spawning, the coordinator launched the registry subagent, waited for it to finish, and only then launched the data-room subagent, doubling the run time. The two investigations need nothing from each other. Which two changes fix both problems? (Select two.)

   Select exactly 2. 0/2 selected.

   AAdd Task to the coordinator's allowedTools, since a coordinator cannot spawn any subagent, however complete its AgentDefinitions are, unless the Task tool is available to it.

   BDescribe the two subagents in prose inside the coordinator's system prompt so the model knows they exist and can decide when to hand work to them.

   CGive the coordinator every tool the two subagents hold so it can complete both investigations itself without spawning, avoiding the coordination overhead entirely.

   DRewrite each AgentDefinition system prompt as a step-by-step procedure so that the subagents finish faster and the coordinator waits less between launches.

   EHave the coordinator emit the Task calls for the registry and data-room subagents in one response so the two independent investigations run in parallel.
   Check answerShow the answer
4. Question 4Domain 1: Agentic Architecture & OrchestrationSingle answer

   Kestrel Savings Bank's support agent issues goodwill credits through a credit_account tool. Policy caps agent-issued credits at 200 dollars; anything larger needs a human approver. The system prompt states the cap and gives six examples, and compliance over the quarter measured 99.4 percent, yet 31 credits above the cap were issued. Internal audit has written that the control must be demonstrably incapable of failing, not merely rare to fail, before the agent stays in production, and the bank wants credits within the cap to remain fully automated. Which change satisfies both requirements?

   Select the best answer.

   AAdd a tool call interception hook that checks every credit_account call before it executes, blocks any amount above 200 dollars, and redirects that case into the human approval workflow.

   BAdd a PostToolUse hook that reads each credit_account result after it executes, flags any amount above 200 dollars, and writes the case with its customer details to a reversal queue that staff review each morning.

   CExtend the prompt's six examples to twenty, covering the request phrasings found in the 31 failures, and add a closing instruction that the cap is absolute.

   DRemove the credit_account tool from the agent and have it draft a credit request for a human to submit, accepting a slower resolution on every credit regardless of size.
   Show the answer
5. Question 5Domain 1: Agentic Architecture & OrchestrationSingle answer

   Lumen Analytics spent a two-day Claude Code session analysing its billing service, with the agent building a detailed map of the module boundaries and call paths. No code has changed since. The team now wants to evaluate two refactoring approaches, an event-sourced ledger and a conventional audit table, and see each one explored to a proposal before choosing. They do not want the second exploration coloured by conclusions from the first, and they cannot spare the two days to re-run the analysis. How should the engineer proceed?

   Select the best answer.

   AResume the analysis session and ask the agent to explore the event-sourced approach to a proposal, then in the same session ask it to explore the audit table approach and compare the two.

   BStart two fresh sessions, paste a summary of the module map into each, and explore one approach per session so that neither is influenced by the earlier conversation.

   CUse fork_session twice from the analysis session, so each fork inherits the full module map as a shared baseline and explores one approach independently of the other.

   DResume the analysis session and ask the agent to weigh both approaches in a single response, so the shared context is used once and the team receives one comparative recommendation.
   Show the answer
6. Question 6Domain 2: Tool Design & MCP IntegrationSingle answer

   Harrow Home Retail's product-content agent has one MCP tool, analyze_document, that takes a file and a free-text instruction and returns whatever text the backend model produces. The agent uses it to pull specifications from supplier sheets, to summarise care guides and to check marketing claims against the supplier's own wording. Downstream code that expects specification fields fails on 18 percent of calls because the output shape varies, and the agent sometimes sends web pages through the same tool. Each use has a different consumer with fixed expectations. What should the tool designer do?

   Select the best answer.

   AExpand the analyze_document description with worked examples of each instruction style, so the model phrases its free-text instruction consistently and the backend returns a steadier shape.

   BAdd few-shot examples to the agent's system prompt showing the exact output expected for specifications, summaries and claim checks, so the model reformats the tool's text before passing it on.

   CRename analyze_document to analyze_supplier_document and state explicitly in its description that web pages are out of scope, so the model stops misrouting pages to it and the remaining calls settle down.

   DSplit analyze_document into extract_specifications, summarize_care_guide and verify_claim_against_source, each with a defined input and output contract and a description of when the model should choose it.
   Show the answer
7. Question 7Domain 2: Tool Design & MCP IntegrationSelect 2

   Wrenfield Hospital Group's scheduling agent books appointments through MCP tools. Two failure modes recur. When the booking backend times out, which happens for about two percent of calls and clears within seconds, the tool returns the text Operation failed and the agent tells the patient to call back later. When the agent's service account lacks access to a particular clinic's calendar, the tool returns the same text and the agent retries six times before giving up. Clinic managers want retries where they help and none where they cannot. Which two changes to the tool's responses should be made? (Select two.)

   Select exactly 2. 0/2 selected.

   AReturn every failure as the same isError response and add a system prompt rule allowing the agent to retry any failed call at most twice before it escalates to a scheduler.

   BCatch all exceptions inside the MCP server and return an empty result set with no error flag, so the agent moves on and no patient is told to call back.

   CReturn timeouts with isError set, an errorCategory of transient and isRetryable true, so the agent retries the booking a moment later instead of sending the patient away.

   DReturn calendar access failures with isError set, an errorCategory of permission, isRetryable false and a description naming the clinic, so the agent escalates once rather than retrying.

   EHave the MCP server retry every failed call three times with backoff before responding, and return a generic unavailable status to the agent once its own retries are exhausted.
   Check answerShow the answer
8. Question 8Domain 2: Tool Design & MCP IntegrationSingle answer

   Calder Valley Council processes planning applications with an agent that has four tools: extract_metadata, classify_application, enrich_with_gis and summarise_objections. Metadata must be extracted first because the other three tools take its application_id as input, and when the model skips that step the later calls fail on a missing identifier, which happened on 11 percent of applications last month. The council cannot change the downstream tools' input contracts before the next release. What is the most reliable fix?

   Select the best answer.

   ASet tool_choice to any on every request so the model must call a tool rather than reply in text, which ensures that the extraction stage is never skipped.

   BAdd a system prompt paragraph explaining the dependency between the four tools and place extract_metadata first in the tools array so the model encounters it before the others on every request.

   CMerge the four tools into one process_application tool whose description lists the four internal stages, so a single call always performs metadata extraction before the rest.

   DForce extract_metadata on the first turn with tool_choice set to that tool by name, then continue with tool_choice auto so the model chooses the remaining tools itself.
   Show the answer
9. Question 9Domain 2: Tool Design & MCP IntegrationSingle answer

   Veltrix, a SaaS company, added an internal MCP server to its project .mcp.json that searches code semantically across all 40 of its repositories and returns file paths with snippets. Engineers report that Claude Code still answers questions such as where a function is called by running Grep on the local checkout, so callers in other repositories are missed. The MCP tool's description reads: Searches code. Grep remains the right choice for the local repository. Which change should the team make first?

   Select the best answer.

   ARemove Grep from the allowed tools in the project settings so that every code search has to go through the MCP server, whatever repository the answer lives in.

   BExpand the MCP tool's description to explain that it searches all 40 repositories semantically and returns paths and snippets, and the cases where it beats Grep.

   CMove the server definition from the project .mcp.json to each engineer's user-level ~/.claude.json, since user-scoped servers are discovered before project-scoped ones.

   DAdd a CLAUDE.md instruction stating that Claude must always use the MCP search tool for every search, so it never falls back to Grep on the checkout.
   Show the answer
10. Question 10Domain 3: Claude Code Configuration & WorkflowsSingle answer

    Bexley Aerospace keeps six packages in one repository with a single 900-line root CLAUDE.md that mixes the standards for all of them. Engineers working in the telemetry package see Claude apply the ground-station UI rules, sessions start with thousands of tokens of irrelevant guidance, and each package's standards are maintained by a different lead who does not want to edit a shared file. The team wants each package to load only its own standards while the leads keep ownership of their own files. What should the team do?

    Select the best answer.

    AGive each package a CLAUDE.md that uses @import to include only the standards files relevant to that package, with each file maintained by the lead who owns it.

    BSplit the root file into six topic files under .claude/rules/, one per package, so each lead maintains a file of their own and every standard is easier to find and review.

    CTurn each package's standards into a project skill engineers run when they begin a task in that package, so nothing loads until someone asks for it.

    DKeep the single root CLAUDE.md but add a heading for each package and an instruction telling Claude to apply only the section matching the current directory.
    Show the answer
11. Question 11Domain 3: Claude Code Configuration & WorkflowsSingle answer

    Halcyon Assurance's engineering platform team maintains a project skill, /release-notes, in .claude/skills/, used by 30 engineers to draft customer-facing notes from merged pull requests. An engineer on the internal tooling team wants the same workflow but with a terse, bullet-only format for an internal channel. She edited the shared SKILL.md to produce her format, and the next day customer-facing notes from every other team came out as bullets. The team wants her to keep her format without changing anyone else's output. What should she do?

    Select the best answer.

    ARevert the shared skill and add context: fork to its frontmatter, so each engineer's invocation runs in an isolated context and formatting changes cannot leak between teams.

    BRevert the shared skill and create a personal skill in ~/.claude/skills/ that keeps the name /release-notes, so her copy applies in her sessions while teammates continue to load the shared one.

    CRevert the shared skill and create a personal skill in ~/.claude/skills/ under a different name, such as /internal-notes, containing her bullet-only variant of the workflow.

    DRevert the shared skill and move her format into a section of the project CLAUDE.md that applies whenever release notes are being drafted for an internal channel.
    Show the answer
12. Question 12Domain 3: Claude Code Configuration & WorkflowsSingle answer

    Greyfield Manufacturing must replace a deprecated date library used in 64 files across its production scheduling application. Two candidate replacements handle time zones differently, and the plant's release window closes in five working days. On a previous migration the team began editing on day one and threw away three days of work when a hidden dependency surfaced. The engineer running the task in Claude Code wants to avoid a repeat without missing the window. How should the task be run?

    Select the best answer.

    AUse direct execution from the first prompt, naming the chosen replacement and the 64 files, because the scope is already known, the edits are mechanical, and five days leaves no time for a planning phase.

    BEnter plan mode to map the dependency graph and settle the choice between the two libraries, then switch to direct execution to carry out the agreed plan across the affected files.

    CStay in plan mode for the whole migration, requesting a fresh plan before each file, so that every change is reviewed in advance and no dependency can surface unannounced.

    DRun the migration as direct-execution sessions of about ten files each, so every change stays small and well scoped and a problem found in one batch cannot spread rework into the others.
    Show the answer
13. Question 13Domain 3: Claude Code Configuration & WorkflowsSelect 2

    Orrington Bank generates tests for each pull request from a CI job that runs Claude Code with -p. Reviewers reject 45 percent of the generated tests: about half duplicate scenarios the existing suite already covers, and the rest exercise trivial getters or set up fixtures by hand that the shared test harness already provides. The prompt currently contains only the diff and the instruction to write thorough tests. The job must stay within its current run time. Which two changes best fit these constraints? (Select two.)

    Select exactly 2. 0/2 selected.

    AAdd an instruction to the prompt to skip duplicate and trivial tests and to use the shared harness, so the model applies the team's bar without adding any files to the context.

    BProvide the existing test files for the changed modules in the prompt context so generation can see which scenarios are already covered and avoid proposing them again.

    CFeed the tests reviewers rejected on earlier pull requests, from other modules, into the prompt as examples of what not to produce, so the model learns the bar from past output.

    DAdd a post-generation step that deletes any generated test whose function name matches the name of a test already present in the suite before the tests are posted.

    EDocument the testing standards, the criteria for a valuable test and the fixtures the shared harness provides in CLAUDE.md so CI-invoked Claude Code loads them on every run.
    Check answerShow the answer
14. Question 14Domain 4: Prompt Engineering & Structured OutputSingle answer

    Meridian Mutual, an insurer, extracts claim details from adjuster reports with a tool_use call whose JSON schema marks every field required. A sample of 1,200 extractions shows that in 9 percent of cases the policy_end_date value looks plausible but appears nowhere in the report, and those reports never state an end date at all. Downstream billing consumes the field and cannot tolerate an invented date, and the team must keep schema enforcement because it removed the JSON syntax failures they had before. Which change most directly removes the fabricated dates?

    Select the best answer.

    AAdd a prompt instruction telling the model to report a date only when it is certain of it, while keeping policy_end_date required so that downstream billing code never has to handle a missing value.

    BMake policy_end_date nullable in the schema and instruct the model to return null when the report does not state an end date, so the requirement to fill the field no longer forces a plausible-looking value.

    CDrop the tool_use schema and ask for JSON in the response text with a prompt that permits omitting unknown fields, since a required field in the tool definition is what pushes the model to fill it.

    DAdd a retry step that re-sends the report with a validation error whenever the extracted date fails a plausibility check, giving the model a second chance to correct the value it returned.
    Show the answer
15. Question 15Domain 4: Prompt Engineering & Structured OutputSingle answer

    Halden Logistics extracts bill-of-lading fields with a retry-with-error-feedback loop capped at three attempts. In a run of 2,000 documents, 140 failed first-pass validation. The logs show 110 of them failed on a missing consignee_tax_id and were still failing after the third attempt; an analyst confirms the identifier appears only in a separate customs form that is never sent with the request. The other 30 failed on date-format mismatches and were corrected on the second attempt. Every retry re-reads the whole document, and the finance team has capped monthly extraction spend. Which change should the architect make?

    Select the best answer.

    ARaise the retry cap from three to six for every failing document, since the second attempt already resolves the date mismatches and further attempts should eventually work through the remaining failures as well.

    BReplace the retry loop with one stricter prompt instructing the model to search the entire document more thoroughly for the tax identifier, including headers and footers, before it returns any extraction at all.

    CRelax the validator so a missing consignee_tax_id no longer fails a document, letting all 2,000 pass on the first attempt, removing the retry cost from the monthly bill and leaving the field to be filled later.

    DStop retrying when the error is a missing consignee_tax_id, attaching the customs form or recording the field as null instead, and keep retry with error feedback for date-format and other structural errors.
    Show the answer
16. Question 16Domain 4: Prompt Engineering & Structured OutputSelect 2

    Corvane University audits research-grant documents for compliance using the Message Batches API. Its policy promises a result within 30 hours of a document arriving, and the finance office insists on keeping the 50 percent batch discount. Last month a submission of 30,000 documents saw 2 percent of requests fail because the documents exceeded the context limit, and the team resubmitted the whole batch, doubling that day's cost. The team currently submits one batch each night at midnight. Which two changes should the architect introduce? (Select two.)

    Select exactly 2. 0/2 selected.

    AMove the whole audit to the synchronous Messages API with high concurrency so context-limit failures surface within minutes and can be retried immediately instead of a day later.

    BUse the custom_id on each failed request to resubmit only the failed documents, chunking the ones that exceeded the context limit before they are sent again.

    CKeep the single midnight submission, because the 24-hour processing window already guarantees that every document is complete before the next night's batch is assembled and sent.

    DSubmit a batch at least every four hours, so a document arriving just after one submission still completes within 30 hours even when processing takes the full 24-hour window.

    EEnable multi-turn tool calling inside each batch request so the model can fetch the remaining sections of an oversized document by itself while the batch is still processing.
    Check answerShow the answer
17. Question 17Domain 4: Prompt Engineering & Structured OutputSingle answer

    Pellwater Bank runs a Claude Code review on every pull request. Over the last quarter developers dismissed 71 percent of findings in the comment-accuracy category but only 12 percent in the security category, and an internal survey shows most of them now skip the review output altogether. Last sprint the team told the review to be conservative and to report only findings it was highly confident in; dismissal rates did not move. A regulatory release freeze in six weeks leaves no room for a long prompt rewrite. What should the architect do first?

    Select the best answer.

    ATemporarily disable the comment-accuracy category so developers see the findings they trust, and rewrite its rule as a specific criterion: flag a comment only when its claimed behaviour contradicts the code.

    BAdd a stricter instruction that the review may report only findings it is at least 90 percent confident in, and require the model to print the confidence figure and a one-line justification alongside each finding it reports.

    CSwitch the review off entirely until the prompt has been rewritten and re-evaluated against a labelled set of past pull requests, since a review that developers ignore returns nothing for its API spend and erodes trust further.

    DKeep every category running and require developers to record a reason with each dismissal, so the prompt can be revised on evidence once a full quarter of dismissal data has been collected and analysed by category.
    Show the answer
18. Question 18Domain 5: Context Management & ReliabilitySingle answer

    Brightline Retail runs a support agent that handles returns across sessions of 30 or more turns, summarising the conversation history once it passes turn 20. In 6 percent of long sessions the refund the agent issues differs from the amount it quoted earlier, for example 84.50 quoted and 85.00 refunded, and each lookup_order result adds more than 40 fields to the context when only a handful concern the return. The context budget per session is fixed and cannot grow. Which change most directly protects the quoted amounts?

    Select the best answer.

    AInstruct the summariser to produce longer, more detailed summaries that quote every numerical value verbatim, and move the summarisation trigger from turn 20 to turn 30 so the condensing happens later in the session.

    BTurn summarisation off and send the complete conversation history in every request, because the API is stateless and coherent behaviour depends on the full transcript being present each time.

    CKeep amounts, dates, order numbers and statuses in a case-facts block sent with every prompt outside the summarised history, and trim each lookup_order result to its return-relevant fields.

    DHave the agent restate the quoted refund figure and order number to the customer on every turn so they always sit in the most recent part of the context, where the model reads them most reliably.
    Show the answer
19. Question 19Domain 5: Context Management & ReliabilitySelect 2

    Ashgrove Health extracts clinical fields from referral letters, discharge summaries and laboratory reports. On the validation set the model reaches 97 percent field accuracy overall, and the operations lead wants to remove human review for every extraction with a model confidence above 0.9. Reviewers can check about 300 of the 4,000 documents that arrive each day. Before any extraction bypasses review, which two steps should the architect require? (Select two.)

    Select exactly 2. 0/2 selected.

    AAsk the model to attach a short written justification to every confidence score so reviewers can check the reasoning instead of the extraction itself, roughly halving the time each document takes.

    BSplit the 97 percent figure into per-document-type and per-field accuracy, and confirm every segment clears the bar, since the aggregate can hide a weak segment such as laboratory reports or a dosage field.

    CRaise the automation threshold from 0.9 to 0.95 before going live, so that the model's own uncertainty screens out marginal extractions and fewer documents bypass review than the operations lead proposed.

    DStart by automating the two highest-volume document types, where the model has seen the most examples and the reviewer time saved is greatest, and extend automation to the remaining types once those run cleanly.

    ECalibrate the 0.9 threshold against a labelled validation set, then keep a stratified random sample of high-confidence extractions flowing to reviewers to measure error rates and detect new patterns.
    Check answerShow the answer
20. Question 20Domain 5: Context Management & ReliabilitySingle answer

    Fenwick Doyle LLP runs a multi-agent research system that drafts market briefs for clients. A partner found a brief stating the sector grew 4.1 percent when the two documents the analysis subagent read gave 4.1 percent in a 2024 annual report and 6.3 percent in a 2025 mid-year update; the synthesis agent had picked one figure, dropped the other and attached no date or source to either. The firm requires that disagreements between sources are visible in the brief, never resolved silently. Which pipeline change is most defensible?

    Select the best answer.

    AInstruct the synthesis agent to prefer the most recent figure whenever two sources disagree, since newer data supersedes older data in market research and a client brief should carry one headline number per metric.

    BAdd a verification pass in which a separate Claude instance, without the synthesis agent's context, re-reads the finished brief and removes any statistic that does not appear in at least two of the source documents.

    CHave the coordinator require subagents to return their full reasoning chains with each finding, so the synthesis agent can judge which figure is better supported before it writes the brief and cite the stronger one.

    DRequire the analysis subagent to attach a publication date and a source mapping to every figure, and have the synthesis agent present conflicting values side by side with their sources instead of choosing one.
    Show the answer