Claude Certification Blog
Ten CCAR-F practice questions, worked option by option
These ten CCAR-F practice questions are original scenario items written to the five-domain official Exam Guide v1.0, allocated by the published weights so every domain appears, each with the key and a stated reason why every option holds or fails.
The ten CCAR-F practice questions below are not exam items and were not written from any exam item. Each is an engineering scenario with a measured symptom and a constraint, then four options, or five for a select-two. Seven are single response and three are select-two, the two formats the guide describes. Answer every item cold, record the misses in the map below, and open the reason for every option afterwards. Three items sit in Agentic Architecture and Orchestration because it is the heaviest domain at 27%.
The official CCAR-F certification page and exam guide are the authority on format and weights; the Claude Certified Architect - Foundations page carries the blueprint and six more open items. The reading method these items reward is in the Claude certification sample questions hub, and what official material exists is in the guide to Claude certification practice exams.
How are these CCAR-F practice questions weighted?
By the official blueprint: 27% Agentic Architecture and Orchestration, 20% Claude Code Configuration and Workflows, 20% Prompt Engineering and Structured Output, 18% Tool Design and MCP Integration, 15% Context Management and Reliability. The flattest blueprint of the four exams lets ten items reach every domain: three, two, two, two and one. The live exam frames its 60 items with four scenarios drawn from a bank of six; these ten are standalone scenarios, and the free 20-item CCAR-F set is not built as scenarios either.
Agentic Architecture and Orchestration: who decides, and in code or in prose?
Three items on the heaviest domain. The first asks what a coordinator should do when most queries need one subagent and every query gets four; the second pairs a financial step that must never be skipped with a handoff that arrives as a one-line note; the third is a fixed pipeline applied to an open-ended task. The shared trap is a stronger instruction where the guide asks for a structural change: a gate in code, a coordinator that selects, a plan that adapts. Parallel spawning and prompt chaining are real techniques, offered here for problems they do not fix.
The Riverdale Transport Authority, a public agency, runs a multi-agent research system whose coordinator always invokes four subagents in a fixed order: a web searcher, then a document analyst, then a synthesiser, then a report writer. A usage review shows 70 percent of staff queries are single-fact lookups such as a route's opening date, yet every query costs the same as a full literature review, and the budget holder has frozen further spend. Latency is acceptable. Which change best fits the constraint?
Show the answer and the reason for every option
Answer: C
Dynamic subagent selection, not the full pipeline. Task statement O02: Orchestrate multi-agent systems with coordinator-subagent patterns.
A. Parallel Task calls shorten the pipeline but still run every subagent on every query, so cost is unchanged. The stem says latency is acceptable and cost is the frozen constraint, which makes this the wrong lever.
B. Direct subagent-to-subagent handoff bypasses the coordinator, which the guide says must route all inter-subagent communication for observability, error handling and controlled information flow.
C (correct). Routing by query complexity is the guide's named remedy for a coordinator that always runs every stage. Single-fact lookups consume one subagent, broad topics consume four, and spend falls with the mix.
D. A report cache only pays off when an identical question recurs. Most of Riverdale's queries are distinct lookups, each of which would still trigger the full four-subagent pipeline on first ask.
Marlowe and Finch, an online furniture retailer, runs a returns agent on the Claude Agent SDK with the MCP tools get_customer, lookup_order, process_refund and escalate_to_human. The system prompt states that identity must be verified through get_customer before any refund. Last month the agent skipped that step in 9 percent of refunds and paid two of them to the wrong account. Separately, the human agents who receive escalations say each handoff arrives as a one-line note asking them to assist, and they have no access to the transcript. Which two changes address both problems? (Select two.)
Show the answer and the reason for every option
Answer: B and C
Gate the refund in code, hand off in structure. Task statement O04: Implement multi-step workflows with enforcement and handoff patterns.
A. Repositioning and repeating the rule keeps enforcement in the prompt, and the guide says prompt instructions alone carry a non-zero failure rate where compliance must be deterministic. Some refunds would still be paid to unverified accounts.
B (correct). A prerequisite gate that refuses process_refund until get_customer has returned a verified customer ID is the programmatic enforcement the guide prescribes when a tool sequence protects a financial operation. The skip rate goes to zero by construction.
C (correct). A structured handoff with customer ID, root cause, the amount in dispute and a recommended action is the guide's protocol for escalating to human agents who cannot read the transcript. It gives them what the one-line note omits.
D. Few-shot examples demonstrate the pattern but still rely on probabilistic compliance. The guide treats them, like stronger wording, as insufficient when a skipped step has financial consequences, so the wrong-account refunds would continue at a lower rate.
E. A full reasoning trace and every raw tool result bury the facts the human needs in verbose output, the opposite of the compiled summary the guide describes. The receiving agent still has to reconstruct customer ID, root cause and recommended action.
Quillbrook Software asked a Claude Agent SDK workflow to add comprehensive tests to a 900-module legacy codebase. The workflow follows a fixed pipeline: list modules alphabetically, generate tests for each, then run the suite. After three days it has covered 210 modules, almost all small utility helpers, while the payment reconciliation and permissions modules that caused last quarter's incidents remain untested. The team has budget for two more days of agent time. How should the decomposition change?
Show the answer and the reason for every option
Answer: B
Open-ended tasks need adaptive decomposition. Task statement O06: Design task decomposition strategies for complex workflows.
A. Sharding the alphabetical pipeline is faster but not smarter: the incident-prone modules still sit wherever the alphabet puts them, and with 690 modules left there is no guarantee they are reached in two days.
B (correct). Map structure, identify high-impact areas, build a prioritised plan that adapts as dependencies are found: this is the guide's own decomposition for open-ended legacy testing tasks, and it spends the remaining budget where incidents came from.
C. Prompt chaining into fixed passes is the guide's pattern for predictable multi-aspect reviews. It adds an integration pass but keeps a fixed order, so prioritisation by risk is still absent.
D. Writing more tests per module increases depth on the utility helpers already covered. The stated gap is which modules are tested at all, not how thoroughly the covered ones are tested.
Tool Design and MCP Integration: what does the model select on, and where do secrets live?
One select-two and one single. The first is a misrouting case with two causes at once: two near-identical one-line descriptions, and a system-prompt phrase that ties a whole class of questions to the wrong tool. Both fixes are needed; the classifier, the forced tool call and the merged tool each miss one. The second is scope and secrets: a shared server belongs in the project's committed configuration, and the token reaches it through an environment variable, never through the repository.
Fernhill Outfitters, a retailer, gives its support agent two MCP tools, search_orders and search_shipments. Their descriptions are one sentence each and nearly identical. The system prompt tells the agent to check the order first before answering any delivery question. Logs show 31 percent of delivery questions are routed to search_orders, which cannot return carrier status, and the agent then apologises for missing information. Which two changes most directly improve tool selection? (Select two.)
Show the answer and the reason for every option
Answer: B and E
Descriptions select tools; prompts can override them. Task statement O08: Design effective tool interfaces with clear descriptions and boundaries.
A. A pre-classification routing layer bypasses the model's natural language understanding and adds infrastructure to solve a problem the guide attributes to inadequate descriptions and prompt wording.
B (correct). Expanding each description with purpose, accepted identifier formats, example queries and boundary explanations is the guide's primary fix for misrouting between similar tools, and renaming removes any residual overlap.
C. The agent is already calling a tool; it is calling the wrong one. The any setting guarantees some tool call but has no effect on which of two similar tools is chosen.
D. Merging the tools behind a first-match rule discards the distinction the agent needs and can return an order record when carrier status was wanted. The guide favours differentiating tools, not blurring them.
E (correct). The guide lists looking through the system prompt for instructions that are keyword-sensitive as a skill in this objective. The order-first phrase is an association that steers delivery questions to the wrong tool regardless of descriptions.
Caldmere Bank's platform team wants every developer's Claude Code session to reach the internal ticketing system through the same MCP server. A developer added the server to their personal ~/.claude.json with an access token pasted in, and it works only on that laptop. Security has since found the same token committed in a teammate's copy of the configuration. The team needs a shared setup that never places credentials in the repository. How should the server be configured?
Show the answer and the reason for every option
Answer: D
Shared servers in .mcp.json, secrets in env vars. Task statement O11: Integrate MCP servers into Claude Code and agent workflows.
A. User-scoped configuration is for personal or experimental servers. A shared server kept there is set up by hand on each laptop, and a token written into an onboarding guide is a credential in a document, not in an environment.
B. Once the token is committed it is in history regardless of a later .gitignore entry. Rotation reduces the window of exposure but leaves in place the anti-pattern that security has already flagged.
C. CLAUDE.md provides project context and instructions; it does not configure MCP servers, and writing a token into it commits the credential to the repository that the requirement forbids.
D (correct). Project-scoped .mcp.json is the guide's location for shared team servers, and ${VAR} expansion keeps the token out of version control while letting every developer's session connect with their own credential.
Claude Code Configuration and Workflows: what stays out of the main context?
Two items. The first is a team skill whose findings are worth keeping and whose 60,000 tokens of scan noise are not; the frontmatter option that runs it in a forked sub-agent isolates the noise without making the skill personal or narrowing what it scans. The second is a pull request pipeline with two defects, unparseable output and duplicated findings, and the two fixes the guide names for exactly those: a JSON schema on the output, and the earlier findings fed back into context. Batch processing fails because the check is blocking; session memory fails because the guide prefers explicit context.
Dunmore Kettering LLP's engineering team built a project skill, /licence-scan, that reads every dependency manifest and prints licence findings. It works, but after one run the main Claude Code conversation holds about 60,000 tokens of scan output, and developers report the session forgetting the feature they were implementing before the scan. The skill must remain available to the whole team from the repository. Which frontmatter change addresses the context problem?
Show the answer and the reason for every option
Answer: A
Isolate verbose skills with context: fork. Task statement O14: Create and configure custom slash commands and skills.
A (correct). context: fork executes the skill inside a forked sub-agent, so the scan output stays out of the main thread and only the findings return. The guide names verbose analysis as the case this option exists for.
B. allowed-tools limits the tool set available while the skill runs. It is a safety control against destructive actions and has no effect on how much output enters the main thread.
C. A personal variant in ~/.claude/skills/ is how the guide suggests tailoring a skill without touching what teammates run. It removes the skill from the repository, contradicting the stem, and the output still floods the session.
D. argument-hint prompts for missing parameters. Limiting each run to one manifest shrinks output by changing what the scan covers, which is a workaround rather than the isolation the guide provides.
Greyfriars Insurance drives Claude Code from its pull request workflow using the -p flag. Two problems persist: the free-text output breaks the script that posts inline comments in about one build in five, and after each new commit the review repeats every finding it already posted, so a typical pull request ends up carrying 40 duplicate comments. Reviews must stay blocking and complete within the existing 15-minute job. Which two changes fix these problems? (Select two.)
Show the answer and the reason for every option
Answer: A and C
Structured CI output and deduplicated re-reviews. Task statement O18: Integrate Claude Code into CI/CD pipelines.
A (correct). The guide names --output-format json and --json-schema as the CLI flags for enforcing structured output in CI. A schema-conformant result replaces the fragile free-text parsing behind the failing builds.
B. The guide's mechanism for avoiding duplicates is to supply the earlier findings explicitly, not to depend on a persisted session. A reviewer carrying its own earlier reasoning also inherits its earlier blind spots, which is the isolation point the guide makes.
C (correct). Supplying the earlier findings and asking only for what is new or still open is the guide's own instruction for re-running reviews after new commits without duplicate comments.
D. Batch requests may take as long as a full day to complete and carry no latency guarantee, so they cannot serve a blocking check that must finish within the 15-minute job. Cost savings do not change that.
E. The -p flag is what keeps Claude Code from waiting for interactive input in a pipeline. Removing it reintroduces the hang the flag exists to prevent, and stdin redirection does not fix formatting.
Prompt Engineering and Structured Output: what fixes both symptoms with one change?
Two single-response items. The first is a self-review that declared its own defect intentional; the guide's answer is an independent instance with no access to the generation conversation, not more thinking, more passes or a per-table split inside the same session. The second is a review whose two pages of instructions still produce six layouts and a 45 percent false-positive rate; a strict schema fixes only the layout, a conservative instruction fixes neither, and a few targeted examples fix both.
Ravenhurst Hospitals uses Claude to generate database migration scripts for its scheduling system. Each script is generated and then reviewed in the same session with the instruction to check your work carefully and list any risks. In the last quarter, 6 of 48 scripts reached staging with defects the self-review had explicitly declared safe, including a null handling error the review described as intentional. Each script must still be reviewed before staging within the current hour-long window. What change most improves defect detection?
Show the answer and the reason for every option
Answer: C
A fresh instance has no reasons to defend. Task statement O24: Design multi-instance and multi-pass review architectures.
A. The guide says independent instances are more effective than either self-review instructions or extended thinking. More reasoning inside the same context still starts from the belief that the null handling was intentional.
B. Three passes in the same session share the same reasoning history, so each inherits the original rationale for the defect. Consensus among them measures consistency with the generator, not correctness.
C (correct). A second instance without the generation conversation has no prior decisions to justify, which the guide identifies as the reason independent review outperforms self-review. It fits within the existing review window.
D. Per-table and cross-table passes address attention dilution in large reviews. The reported defects were declared safe on purpose, which is a shared-context problem that splitting the same session's review does not remove.
Tidewater Logistics runs an automated Claude review on pull requests. Its prompt contains two pages of instructions on output format and on which patterns to flag, yet findings arrive in six different layouts and 45 percent of them flag repository conventions the team considers acceptable, such as a shared retry wrapper. Developers dismiss most comments unread. The team wants one change that fixes both the layout variance and the false positives. Which change should they make?
Show the answer and the reason for every option
Answer: B
Examples teach format and judgement together. Task statement O20: Apply few-shot prompting to improve output consistency and quality.
A. A JSON schema through tool use guarantees the shape of each finding, which solves the layout variance. It cannot teach the model that the shared retry wrapper is acceptable, so the false positives continue.
B (correct). The guide names few-shot examples as the remedy when detailed instructions still yield inconsistent output, and as the way to show acceptable patterns versus genuine issues. One set of examples addresses both defects.
C. The guide states that general caution instructions and confidence-based filtering fail to improve precision compared with specific criteria or examples. Shorter prose does not add the judgement that is missing.
D. Temperature controls sampling randomness. A deterministic sample of a prompt that has not conveyed the team's conventions produces the same misjudgements every time, in one of the same six layouts.
Context Management and Reliability: is an empty list a result or a failure?
One item on the lightest domain, and the sharpest distinction in the set. A search subagent that returns the same empty list for no matches and for a timeout leaves its coordinator unable to tell a gap in the literature from a broken API. The closest miss adds a fixed status that separates the two but strips the context a retry decision needs; terminating on any empty list is an anti-pattern the guide names outright.
Brackenridge University's multi-agent research system has a search subagent that returns an empty list both when a query genuinely matches nothing and when the search API times out. The coordinator treats every empty list as no information available and moves on. Reviewers found that 22 percent of reports list a topic as having no published work when the search had simply failed. Retries cost roughly 30 seconds each and are affordable only where they can succeed. How should the subagent report outcomes?
Show the answer and the reason for every option
Answer: A
Failures and empty results are different outcomes. Task statement O27: Implement error propagation strategies across multi-agent systems.
A (correct). Distinguishing access failures from valid empty results, and attaching failure type, attempted query and partial results, is exactly what this objective requires. The coordinator retries failures and records genuine gaps as gaps.
B. A blanket retry spends 30 seconds on every legitimate no-match result and, when it too returns empty, the coordinator still cannot tell whether the topic is unpublished or the API failed twice.
C. Ending the whole research run on any empty list is the workflow-termination anti-pattern the guide names, and it treats a legitimate empty result, which is a successful query, as a failure.
D. A generic status separates failure from no-match but hides the failure type, attempted query and partial results. The coordinator cannot tell a retryable timeout from a malformed query or a revoked key, so it still cannot spend its 30-second retries only where they can succeed.
How to score these CCAR-F questions: the miss map
Score exactly: a select-two item counts only when both letters match. Then write each miss into its domain row rather than adding up a total, because ten items cannot estimate a scaled score. What they can do is point. One miss in a domain means reread the task statement named in that row. Two or more means the domain link, which opens the free set's untimed review filtered to that domain, and then that domain's study section.
| Domain | Items | Your misses | Task statement to reread | Practise the domain |
|---|---|---|---|---|
| D1 Agentic Architecture and Orchestration | 3 | ___ of 3 | O02 Orchestrate multi-agent systems with coordinator-subagent patterns; O04 Implement multi-step workflows with enforcement and handoff patterns; O06 Design task decomposition strategies for complex workflows | Domain 1 in the free set |
| D2 Tool Design and MCP Integration | 2 | ___ of 2 | O08 Design effective tool interfaces with clear descriptions and boundaries; O11 Integrate MCP servers into Claude Code and agent workflows | Domain 2 in the free set |
| D3 Claude Code Configuration and Workflows | 2 | ___ of 2 | O14 Create and configure custom slash commands and skills; O18 Integrate Claude Code into CI/CD pipelines | Domain 3 in the free set |
| D4 Prompt Engineering and Structured Output | 2 | ___ of 2 | O24 Design multi-instance and multi-pass review architectures; O20 Apply few-shot prompting to improve output consistency and quality | Domain 4 in the free set |
| D5 Context Management and Reliability | 1 | ___ of 1 | O27 Implement error propagation strategies across multi-agent systems | Domain 5 in the free set |
If it must happen every time, it is not a prompt
Several of these ten offer a firmer instruction where the stem describes a rule that must hold on every run. Ask what happens on the run where the model ignores the sentence. If nothing stops it, the option is not a control.
The set pairs with the two-week CCAR-F study plan and the CCAR-F cheat sheet, whose decision rules these items exercise; how the four-of-six scenario framing maps to the domains is in the CCAR-F exam scenarios post. The CCAR-F blueprint sets the weights used here. Checked against the official CCAR-F Exam Guide v1.0 (July 2026) on 17 September 2026.
Key takeaways
- Ten original items, all five domains, by weight. The flattest blueprint of the four means no domain rounds to zero, and none can be skipped.
- Answer cold, then reveal. Reading the reasons before committing turns a diagnostic into a reading exercise.
- Score by domain, not by total. Ten items cannot estimate the scaled score; a miss pattern in one domain can name the task statement to reread.
- Enforce in code what must always happen. A firmer instruction is the most common distractor here, and the guide says it carries a non-zero failure rate.
- The free set is a paper, not a replica. Twenty items on the five-domain guide, no scenario framing, a raw count only.
Twenty more items under a clock, then the study sections
The free CCAR-F set is 20 timed items on the five-domain guide with a reason for every option, no account needed. From there the next step is the CCAR-F study sections and the drills by objective, not the older timed forms, which are a retired seven-domain build. Nothing reports a scaled score or a pass probability.
Start the free CCAR-F setQuestions
Frequently asked
The follow-up questions people search next.
Are these real CCAR-F exam questions?
No. All ten were written fresh for this page from the published CCAR-F Exam Guide v1.0 and its 30 task statements, and none is drawn from, or paraphrased from, any exam form or any gated Cred Farmer form. Anthropic keeps its live items confidential; a page offering CCAR-F exam questions that claims otherwise is selling a breach.
How many questions are on the CCAR-F exam?
The official guide sets 60 items in 120 minutes across five domains, framed by four scenarios drawn from a published bank of six. Items are single-answer and multiple-response, and every multiple-response stem states how many options to select. The pass mark is a scaled 720 on a 100 to 1,000 scale; the fee is $125.
Is there a free CCAR-F mock exam?
Yes, with a caveat. The Cred Farmer set at /practice/ccar-f is 20 timed items written to the five-domain guide, no account needed, with a reason for every option after you submit. It is not built as scenarios and is not a replica of the exam, and it reports a raw count by domain, never a scaled score.
Where can I find CCAR-F sample questions?
The official exam guide publishes its own sample items with rationales, and they are the only official ones. Cred Farmer adds six open items on the certification page and the ten worked items here, all written to the same five domains and none drawn from the live bank.
Do the older Cred Farmer CCAR-F timed forms match this exam?
No. They were built on a retired seven-domain outline and are supplemental study context only, so this page never sends you to them. The next step after these ten is the free 20-item timed set, then the study sections and the drills by objective, which follow the five official domains.
Keep reading
Related posts
Not affiliated with, or endorsed by, Anthropic or Pearson VUE. Details are summarised from publicly published program information and can change. Always confirm against the official exam guide before booking.