Claude Certification Blog

CCAR-P practice questions: ten worked architecture scenarios across the seven Professional domains

CCAR-P practice questions earn their place only when every option is explained. Here are ten original worked architecture scenarios across all seven Professional domains, weighted like the blueprint, each with its key and a reason every distractor fails.

10 worked itemsSingle and select-two7 official domains

30 min read

These ten CCAR-P practice questions are original Cred Farmer items written to the CCAR-P Exam Guide v1.0 (July 2026): a named organisation, a measured symptom, a stated constraint, then four options, or five where the item asks you to select two. They cover all seven official domains in rough proportion to weight, every wrong option carries a reason it fails, and none is drawn from a live exam form or from the question bank behind the login. Answer each one cold, then score yourself on the miss map.

The official CCAR-P exam guide publishes three CCAR-P sample questions of its own in Section 8 and says they are not drawn from the live item bank. The Claude certification sample questions hub works one item per code; this post is its Architect Professional spoke. The blueprint and fee are on the Claude Certified Architect - Professional page, and the guide to Claude certification practice exams covers what official practice material exists.

The ten are spread by official weight, so integration, solution design and evaluation get two items each and developer productivity at 7 percent gets one; three ask for two answers. Read the stated constraint before the options.

Seven CCAR-P domains by official weight, from 7 to 19 percent, with one or two items from this set in eachOfficial weight · items hereD1 Solution Design & Architecture17% · 2 itemsD2 Models, Prompting & Context Engineering13% · 1 itemD3 Integration19% · 2 itemsD4 Evaluation, Testing & Optimization16% · 2 itemsD5 Governance, Safety & Risk Management14% · 1 itemD6 Stakeholder & Lifecycle Management14% · 1 itemD7 Developer Productivity & Enablement7% · 1 item
Weights from the CCAR-P Exam Guide v1.0, Section 6; the item counts are this post's ten.

How do you scope a solution to the sponsor's metric, and when do you split one agent?

Domain 1, Solution Design & Architecture, is 17 percent of the blueprint, and two of the ten sit here. The first turns on aligning scope to a business value pillar inside a governance constraint: a named metric, a deadline, and a committee that has put part of the problem out of reach. The second is an orchestration decision about one agent showing three symptoms at once; the stem hands you the fact that decides whether routing can be deterministic.

Question 1 · Written for this post · Domain 1 · single response

Ridgeway Savings Bank asks for a Claude solution to cut contact-centre volume by 30 percent within one quarter. An analysis of 60,000 calls shows 42 percent are card-status queries whose answer already exists in the card-tracking system, 25 percent are transaction disputes that need a judgment, and the rest are mixed. The risk committee has not approved autonomous decisions on disputes and will not meet again for two months. Which scope best aligns the solution with the sponsor's goal?

  • A. An agent that handles every call type with tools for card tracking, dispute case management and account changes, so the 30 percent target can be reached from the widest possible base of calls the centre receives.
  • B. A dispute-resolution agent that settles the 25 percent of calls needing a judgment without a human agent, since those calls are the longest and removing them delivers the largest saving per call handled.
  • C. A card-status assistant that answers the 42 percent of calls by calling the card-tracking system at request time, measured on deflection of those calls, leaving disputes with human agents.
  • D. A summarisation tool that prepares dispute case notes for human agents, cutting handling time on the 25 percent that need judgment without making any autonomous decision on a dispute.
Reveal the answer and the reason for every option

Answer: C. Scope to the metric inside the governance constraint.

A. An agent across every call type includes dispute decisions the committee has not approved, and its breadth makes the 30 percent measurement harder to attribute. The widest base is not the fastest route to the target.

B. Settling disputes end to end is the transformation play, but it is exactly the autonomous decision the committee has not cleared for two months, and 25 percent of calls cannot deliver a 30 percent deflection even if every one were removed.

C. Correct. The largest slice of volume has a live source of truth, a low-stakes answer and no governance blocker, so a tool-backed assistant scoped to it can hit and prove the deflection target within the quarter.

D. Summarising case notes shortens handling time for human agents, which is a productivity gain, not a reduction in call volume. The sponsor's success metric is deflection, and this scope does not move it.

Aligning to a business value pillar means picking the scope that delivers the sponsor's measure inside the constraints the stem gives: a quarter, and no autonomous dispute decisions until the committee approves. Card-status queries are 42 percent of volume, answered by a live system, and carry low risk, so an assistant that calls that system at request time can plausibly hit a 30 percent deflection target on its own and be measured directly. Disputes are the higher-value transformation, but the governance constraint rules out autonomous handling for the whole quarter. The all-call agent widens scope past what is approved, and the note-summarisation tool improves productivity without reducing volume, which is not the metric the sponsor set.

Question 2 · Written for this post · Domain 1 · single response

Westmark Freight built a single agent to resolve shipment exceptions. It holds 38 tools spanning customs, damage and delay handling. In production, tool-selection errors run at 14 percent, the context passes 90,000 tokens by the twelfth turn, and p95 resolution time is over four minutes against a target of ninety seconds. The exception type is known from the ticket before the agent starts. Which orchestration design should the architect propose?

  • A. Keep the single agent but switch it to a tier whose context window is larger, and add a worked example for each of the 38 tools so selection improves without changing the architecture.
  • B. An orchestrator that reads the exception type, dispatches to one specialist subagent holding only that type's tools, and receives a structured result rather than the subagent's transcript.
  • C. Three specialist subagents that all run in parallel on every exception, with the orchestrator merging their proposed resolutions, so no exception waits on a single agent's judgment.
  • D. Three specialists chained in a fixed order, each receiving the previous specialist's full transcript so that context is never lost between the customs, damage and delay stages.
Reveal the answer and the reason for every option

Answer: B. Route once, scope the specialist, return structure.

A. A bigger window absorbs the growth without stopping it, and 38 worked examples add tokens to every turn. The selection error comes from the size of the choice set, which this design leaves unchanged.

B. Correct. Deterministic routing on a known exception type, a specialist with only the relevant tools, and a structured handback address selection error, context growth and latency together.

C. Fan-out suits cases where the right path is unknown until several specialists have looked. Here the type is known upfront, so parallel runs triple spend and latency to produce two results that will be discarded.

D. Passing full transcripts between chained specialists carries the whole context forward, so the token growth returns, and three sequential stages for a single-type exception make the ninety-second target harder.

The stem says the exception type is known before the agent starts, which is the signal that routing can be deterministic and each case needs only one specialist. An orchestrator that dispatches to a single subagent with a scoped tool set attacks all three symptoms at once: fewer tools cuts selection error, a fresh context per specialist stops the 90,000-token growth, and a structured result keeps the orchestrator's own context small. Running every specialist on every case triples cost and latency to solve a routing problem that the ticket already solved. Chaining specialists on full transcripts recreates the context bloat and serialises latency. A larger window and more examples treat the symptoms while leaving 38 tools in one agent's choice set.

What do you cut when the context window has become the cost?

Claude Models, Prompting & Context Engineering is 13 percent, and its context objective asks what in the window is doing work on this turn. The select-two below describes a session whose per-turn input has grown until cost has tripled, with a hard requirement that exact text can still be quoted. Two options each remove one named source of growth; the other three move the ceiling instead of the spend.

Question 3 · Written for this post · Domain 2 · select two

Oakhaven Insurance runs an underwriting assistant for commercial property. A typical session lasts thirty turns. On every turn the application resends the full 60-page policy wording and the complete conversation history, so input tokens per turn pass 120,000 by the end of a session and cost per session has tripled since launch. Underwriters say the assistant must still quote the exact clause when asked. Which two changes reduce the cost while keeping that ability? (Select two.)

  • A. Move to the model tier with the largest available context window so that the growing history and the policy wording fit comfortably for the whole session.
  • B. Compact earlier turns into a structured running summary once the history passes a set size, keeping the system prompt and any stable instructions at the front of every request.
  • C. Cut each underwriter message to a fixed 500-token limit before it is sent, so the history grows more slowly across the thirty turns of a session.
  • D. Lower the maximum output tokens on every turn so that each of the assistant's replies adds less to the history the next turn carries.
  • E. Retrieve only the clauses relevant to the current question from an indexed copy of the policy wording, instead of resending all 60 pages on every turn.
Reveal the answer and the reason for every option

Answer: B and E. Retrieve the clause, summarise the history.

A. A larger window changes the ceiling, not the spend. Every turn still resends 60 pages and the full history, so the tokens billed per session are unchanged and the cost problem the stem describes is untouched.

B. Correct. A structured summary of earlier turns bounds history growth while preserving the decisions made, and keeping the stable prefix first lets prompt caching serve the unchanging part of the request.

C. Underwriter messages are a small share of the 120,000 tokens. Capping them at 500 tokens loses detail the underwriter chose to give and barely moves the total.

D. Shorter replies trim the history slightly, but the assistant must quote exact clauses when asked, and a low output cap forces truncated answers on the turns where completeness matters most.

E. Correct. The policy wording is the dominant fixed cost on every turn. Indexed retrieval of the relevant clauses removes most of it while still giving the model the exact text to quote.

Context optimisation starts by asking what in the window is doing work on this turn. The 60-page wording is resent whole even though a single question touches a few clauses, so indexing it and retrieving the relevant clauses cuts the largest fixed cost while still letting the assistant quote exact text. The history is the other growth term, and a structured running summary of earlier turns bounds it while a stable prefix at the front keeps caching effective. A larger window absorbs the growth without reducing a single billed token. Truncating underwriter messages or shrinking replies damages the conversation to trim tokens that are not the main term. The two keys each remove one of the two named sources of growth.

Where should authorisation and integration live when agents share systems?

Integration is the heaviest domain at 19 percent. Item 4 is a security-gap read: a tool authenticates with one broad credential, and a model instruction is all that stands between a user and data they must not see. An authorisation decision never belongs inside a probabilistic step. Item 5 is a select-two on the integration mechanism for several agents sharing many systems; its second key guards against capability bloat.

Question 4 · Written for this post · Domain 3 · single response

Sable & Roe LLP gives associates a Claude assistant that searches the firm's document management system. The search tool authenticates with a single service account that can read every matter, and the system prompt tells the model to answer only from documents the user is permitted to see, using a matter list the user provides. Last week a paralegal received an excerpt from a matter she was walled off from. The firm's ethical walls are enforced by the document system's own permissions. What should the architect change?

  • A. Strengthen the system prompt with an explicit rule that the model must refuse any request touching a matter outside the user's list, and add examples of refusals to the prompt.
  • B. Keep the service account but record every retrieval with the requesting user's identity, and have the compliance team review the log each week for access outside a user's matters.
  • C. Have the model compare each retrieved document's access list with the user's stated matter list before answering, and drop any document the comparison shows the user should not see.
  • D. Have the search tool call the document system with a token delegated from the authenticated user, so the system's own permissions decide what the query can return before the model sees it.
Reveal the answer and the reason for every option

Answer: D. Let the system of record enforce access.

A. A stronger instruction still rests on a matter list the user supplied and on the model following the rule every time. Authorisation decided in the prompt can be talked around, and the service account still returns everything.

B. Weekly log review is a detective control. It would have identified the paralegal's access after the fact, which is what happened anyway; it does nothing to stop the next walled-off excerpt being returned.

C. Asking the model to enforce access lists puts a deterministic decision inside a probabilistic step, and it still relies on the user's stated matters. The document has already left the system by the time the comparison runs.

D. Correct. With a credential delegated from the authenticated user, the document system applies its own ethical walls to the query itself, so a walled-off matter never enters the results the model can read.

The gap is that authorisation has been moved out of the system that owns it. A service account with firm-wide read access turns every query into a privileged one, and the only thing standing between a user and a walled-off matter is a list the user typed and a model instruction. Authorisation must be enforced deterministically by the system of record, which means the tool should act as the user, with a delegated credential, so the document system's ethical walls apply to every result before anything reaches the model. Prompt rules and model-side comparisons put an access decision inside a probabilistic step, which is the security gap the objective asks an architect to spot. Logging finds the breach afterwards; it did not prevent this one.

Question 5 · Written for this post · Domain 3 · select two

Fennwick Manufacturing runs four internal Claude agents: procurement, maintenance, quality and planning. Each team wrote its own wrappers for the same nine plant systems, so there are 36 integrations to maintain, adding a system takes about three weeks per agent, and a security review found three different ways of handling credentials. The platform team wants one integration surface that keeps each agent's capabilities narrow. Which two changes should the architect recommend? (Select two.)

  • A. Expose each of the nine plant systems through a Model Context Protocol (MCP) server that all four agents connect to, with authentication and authorisation handled once at the server rather than in each agent.
  • B. Merge the 36 wrappers into one shared library that every agent imports, loading every tool description into each agent's context so any agent can reach any system.
  • C. Let the agents call one another over an agent-to-agent protocol, so an agent that lacks a wrapper borrows the capability from the agent that has it.
  • D. Register with each agent only the tools from those servers that its workflows need, so the procurement agent cannot see maintenance or planning operations at all.
  • E. Replace the four agents with a single plant agent that holds all 36 integrations, removing the duplication by removing the boundaries between the teams' workloads.
Reveal the answer and the reason for every option

Answer: A and D. A shared protocol surface with narrow registration.

A. Correct. MCP servers give several agents a standardised way to reach the same systems, and centralising authentication at the server replaces three inconsistent credential patterns with one that the security team can review.

B. One library still means one integration copy per system, which fixes duplication, but loading all 36 tool descriptions into every agent widens each agent's capability set, the opposite of what the platform team asked for.

C. Borrowing a capability through another agent adds latency and a second model's judgment to every call, and it obscures which identity performed the action, which the security review would flag again.

D. Correct. Registering only the needed tools keeps each agent's capability set narrow, so the shared servers do not turn four scoped agents into four agents that can each operate every plant system.

E. A single agent with all 36 integrations removes the duplication by removing the boundaries, which is capability bloat by design. Any fault or injected instruction now reaches every plant system at once.

The objective is to evaluate connection protocols and choose the integration mechanism that fits several clients sharing many tools. That is the case MCP is designed for: one server per system, connected by every agent, with credential handling done once at the server, which resolves the three inconsistent auth patterns and the three-week onboarding. The second change guards against capability bloat, which the platform team named as a constraint: each agent registers only the tools its workflows need. A shared library with every description loaded is the monolithic context strategy and gives every agent every capability. Agent-to-agent borrowing adds a hop and blurs who is authorised. A single plant agent trades duplication for the widest possible attack surface.

How do you prove a change worked, and cut the bill without touching quality?

Evaluation, Testing & Optimization is 16 percent. Item 6 is a causal question with a confound: something else changed the day after the prompt did, and a decision is due by Friday. Only one design isolates the prompt from everything else that week. Item 7 is cost optimisation with quality fixed by a stakeholder who will not accept a sample as evidence. The lever for a high-volume, latency-tolerant job is not the lever for a long stable prefix.

Question 6 · Written for this post · Domain 4 · single response

Brightwell Retail believes a rewritten prompt for its product-recommendation assistant raises checkout conversion. The team switched all traffic to the new prompt on Monday and reports that this week's conversion is 2.1 points higher than last week's. A storewide promotion started on Tuesday. The head of e-commerce wants a decision on keeping the new prompt by Friday and asks how the team should establish whether the prompt caused the lift. What should the architect recommend?

  • A. Run both prompts concurrently on randomly assigned sessions for the rest of the week, with conversion as the pre-declared metric and a sample size set before looking at results.
  • B. Keep the week-over-week comparison but subtract the promotion's estimated effect using last year's promotion uplift, so the residual difference can be attributed to the prompt.
  • C. Score both prompts against the team's offline evaluation set of 500 recommendation cases and keep the new prompt if its judged relevance is higher, since relevance drives conversion.
  • D. Alternate the prompts day by day for the rest of the week and compare the conversion on new-prompt days with the conversion on old-prompt days, adjusting for the promotion.
Reveal the answer and the reason for every option

Answer: A. Randomise concurrently to isolate the cause.

A. Correct. Random assignment at the session level exposes both prompts to the same promotion, the same days and the same shoppers, so the difference in conversion can be attributed to the prompt alone.

B. Subtracting an estimated promotion effect turns one uncertain figure into two. Last year's uplift came from a different catalogue and a different week, and the residual cannot support a Friday decision.

C. An offline relevance score is evidence of quality, and worth keeping as a gate, but it does not measure conversion. Higher judged relevance may or may not translate into checkouts under a live promotion.

D. Alternating by day confounds the prompt with weekday traffic and with the promotion's own daily pattern. A day or two of each is also far too few observations to separate a 2.1-point effect from noise.

A/B testing answers a causal question about a business outcome, and the stem loads it with a confound: a promotion started the day after the switch. Only a randomised concurrent split isolates the prompt from everything else happening in the store that week, and declaring the metric and sample size in advance stops the team reading the result before it is meaningful. Adjusting last week's figure with a borrowed uplift is an estimate stacked on an estimate. The offline set measures judged relevance, which is a fine quality gate but not the conversion the head of e-commerce asked about. Day-by-day alternation leaves the promotion and the weekday pattern tangled with the prompt, since some days carry more promotion traffic than others.

Question 7 · Written for this post · Domain 4 · single response

Carrowmore University summarises about 40,000 lecture transcripts each night for its accessibility service. The job runs on the mid-tier model with a 200-token system prompt, and each transcript is different. Nobody reads the summaries until the following afternoon. The monthly bill has grown past the department's budget, and the accessibility office has said summary quality must not change because students rely on it. Which change most directly reduces the cost?

  • A. Enable prompt caching on every request so that the repeated portion of each call is served from cache and billed at the reduced rate for cached input tokens.
  • B. Submit the transcripts through the Batch API each night, keeping the same model and prompt, once compliance confirms batch processing is covered by the university's agreement.
  • C. Move the job to the fastest model tier, which costs a fraction per token, and confirm afterwards with a sample of students that the summaries still read acceptably to them.
  • D. Raise the job's concurrency so that all 40,000 transcripts finish earlier in the night, reducing the total time the pipeline runs and therefore the compute the department pays for.
Reveal the answer and the reason for every option

Answer: B. Latency-tolerant volume is what batch is for.

A. Caching cuts cost on the repeated part of a request. Here that part is a 200-token prompt in front of a unique transcript of thousands of tokens, so the saving is a rounding error on the bill.

B. Correct. Batch processing lowers the per-token price for work that can wait, which this job can by a full night, and it leaves model, prompt and output unchanged, so the quality constraint holds.

C. Switching tiers is a quality change, and a student sample after the fact is not the eval the accessibility office would accept as evidence that summaries are unchanged. The constraint rules this out as a first move.

D. Cost is billed per token processed, not per hour the pipeline runs. Finishing earlier changes the schedule and may raise rate-limit pressure, while the token count and therefore the bill stay the same.

Cost optimisation starts with naming the binding constraint. Quality is fixed by the accessibility office, so any lever that changes the model or the prompt is off the table until an eval says otherwise. The workload is high-volume and nobody needs a result for hours, which is exactly the shape that batch processing is priced for: the same model, the same prompt, the same output, at a lower rate, with compliance coverage confirmed first. Caching pays when a long, stable prefix dominates each request, and a 200-token prompt in front of a unique transcript is the opposite case. A cheaper tier is a quality change made without an eval, and concurrency changes when the tokens are billed, not how many there are.

Which controls stop an injected instruction from moving money?

Governance, Safety & Risk Management is 14 percent. The select-two below reconstructs an incident with two named causes: an instruction arrived inside a customer message, and an irreversible tool executed before any check ran. Each key closes one cause. The distractors are the controls teams reach for first, and each is downstream of the harm or a probability rather than a control anyone can audit.

Question 8 · Written for this post · Domain 5 · select two

Eastvale Transit Authority runs a Claude agent for fare-card customers that can issue refunds and update account details. Last month a customer's message contained embedded instructions, and the agent issued a refund to a card number named in that text rather than to the authenticated account. An output filter is in place, but the refund had already executed before the reply was screened. The authority processes around 9,000 refunds a month and cannot add staff. Which two controls should the architect add? (Select two.)

  • A. A deterministic pre-execution check on the refund tool that verifies the destination card belongs to the authenticated account, denying by default and logging every denial.
  • B. A stricter output filter that scans each reply for refund amounts and card numbers and blocks the message when the values do not match the account on file.
  • C. Input screening that flags instruction-like content in a customer message and withholds tool access for that turn, so the agent can reply in text but cannot act on it.
  • D. A move to the most capable model tier, since a stronger model is more likely to recognise and ignore instructions embedded in a customer's message.
  • E. A weekly report of all refunds by destination card, reviewed by the finance team so that any refund to a card outside its account is reversed.
Reveal the answer and the reason for every option

Answer: A and C. Authorise before the call, screen the input.

A. Correct. The refund tool is where the irreversible action happens, so a deterministic ownership check before execution, denying by default, is the control that would have stopped this refund regardless of what the model believed.

B. A better output filter still runs after the refund has executed. It could stop the customer seeing a confirmation, but the money has already moved, which is the harm the incident describes.

C. Correct. The injected instructions arrived in the customer's message, and the agent can also update account details. Screening input and withholding tool access on a flagged turn stops such content driving any tool call, refund or otherwise, without adding staff.

D. A more capable model may follow fewer injected instructions, but it is a probability, not a guarantee, and it leaves the tool free to refund any card the model names. Capability is not an authorisation control.

E. A weekly report finds wrong refunds after they clear and depends on reversing them, which at 9,000 refunds a month is a recovery process, not a control that prevents the next one.

The incident has two named causes and each key closes one. The refund reached the wrong card because nothing checked authorisation before the tool executed; an output filter only ever sees the reply, and by then the money has moved. Tool-call authorisation must be deterministic, decided before the call, and fail closed, so the destination-card check on the tool itself is the load-bearing control. Injected instructions entered through the customer message, so input screening that withholds tool access on a flagged turn stops them reaching the refund or account tools at all, and it needs no extra staff. A stricter output filter is still downstream of the side effect. Model capability reduces susceptibility but is not a control anyone can audit. A weekly report is detective, and reversing refunds after the fact is recovery, not a safety design.

What do you leave behind at handoff, and how do you keep a whole team inside the rules?

Stakeholder Communication & Lifecycle Management is 14 percent and Developer Productivity & Operational Enablement is 7, the smallest domain, and they share a theme: enabling the people who inherit or operate the system. Item 9 asks which handoff artefact to prioritise when a team cannot say why a design choice was made and is about to change it. Item 10 asks how a team's Claude Code rules should be enforced once individual configuration has drifted. In both, a document people may or may not read loses to a mechanism that holds on its own.

Question 9 · Written for this post · Domain 6 · single response

St. Aldric Health's architect rolls off a Claude-based prior-authorisation assistant in two days. The internal team inherits a repository that includes the evaluation harness and its latest scores, a diagram and a dashboard. In the handover meeting they ask why the pipeline calls a legacy pricing service through a tool rather than reading the pricing table already in the retrieval index, and nobody in the room can answer. The team plans several cost changes over the next quarter. Which artefact should the architect prioritise producing before leaving?

  • A. A decision log that records each significant design choice, the alternatives rejected, the assumptions it depends on, and who owns revisiting it when those assumptions change.
  • B. A fully annotated architecture diagram showing every component, data flow and integration, with the current version of each dependency and the environment in which it runs.
  • C. A recorded walkthrough of the production system with the architect narrating how each component behaves under load and where the team should look first when something fails.
  • D. A simplification of the pipeline before handover that reads prices from the retrieval index instead of the legacy service, so the team inherits one fewer integration to maintain.
Reveal the answer and the reason for every option

Answer: A. A handoff must record why, not only what.

A. Correct. The team's question is about intent. A decision log with the rejected alternatives and the assumptions behind each choice is what lets them judge whether a planned cost change undoes a deliberate safeguard.

B. The team already has a diagram and can see the tool call in it. Annotating versions and flows explains what is connected, not why the pricing service was chosen over the index.

C. A recorded walkthrough transfers operational knowledge, which matters, but two days of narration will not capture the reasoning behind each design choice in a form the team can consult when a change is proposed.

D. Making that change in the last two days does the very thing the missing reasoning should prevent: prices change on their own timeline, so a retrieved copy goes stale, and the safeguard is removed with no eval run and nobody left to explain it.

Handoff documentation is judged by one test: whether an architect who missed the design sessions could make a safe change using the documents alone. The question the team asked is a why question, and the reason the pricing service is a live tool call rather than a retrieved table is that prices change on their own timeline and a retrieved copy would go stale. Without that reasoning recorded, the next cost change may swap the call for the cheaper retrieval and reintroduce the stale-fact defect. A diagram shows what exists, not why; a walkthrough shows behaviour; simplifying the pipeline on the way out makes the risky change itself, unrecorded. The decision log is the artefact that answers the question actually asked and protects the planned changes.

Question 10 · Written for this post · Domain 7 · single response

Quillstone, a SaaS company, rolled Claude Code out to 60 engineers, each of whom configured it individually. Last week one engineer's session ran a migration command against the shared staging database and dropped a table. A survey found eleven different permission setups across the team, and engineering leadership wants the AI-assisted refactoring speed the team has gained to continue. Which configuration approach should the architect recommend?

  • A. Publish an internal page listing the recommended permission settings and hooks, and ask every engineer to apply them to their personal configuration by the end of the sprint, with a reminder in standup.
  • B. Restrict the whole team to plan mode, in which Claude Code reads and proposes changes but never edits files or runs commands, so no session can reach the staging database at all.
  • C. Commit a project settings file to each repository with the team's allow and deny rules and hooks, and place the destructive database commands in a managed policy that personal settings cannot override.
  • D. Send every engineer an identical personal settings file to replace their own, and re-send it whenever the team's rules change so that all sixty configurations stay aligned with the standard.
Reveal the answer and the reason for every option

Answer: C. Enforce team rules in the configuration hierarchy.

A. A recommended-settings page is advice. Sixty engineers applying it by hand produced eleven setups the first time, and nothing stops the next new joiner from skipping it or a colleague from loosening a rule to get unblocked.

B. Plan mode prevents the drop-table command by preventing every edit and command, which removes the AI-assisted refactoring speed leadership asked to keep. It trades the whole benefit for one safeguard a deny rule provides more narrowly.

C. Correct. Project settings travel with the repository, so every engineer gets the same rules and hooks automatically, and a managed policy makes the destructive-command denial hold above any personal configuration.

D. Emailing a personal settings file aligns configurations for a day. Engineers edit personal settings, new joiners are missed, and the re-send process is exactly the manual step that let eleven variants appear.

Configuring Claude tools for a team means enforcing the shared rules through the tool's own configuration hierarchy rather than through documents people may or may not follow. A project-level settings file committed to each repository gives every engineer the same allow and deny rules and hooks the moment they open the repo, and a managed policy for the destructive commands makes the denial hold regardless of what an individual's personal settings say. That closes the staging incident without slowing the refactoring work leadership wants to keep. An internal page and an emailed file are guidance, not enforcement, and drift back into eleven setups. Plan mode for everyone stops the destructive command by stopping every command, which gives up the productivity the constraint said to protect.

Where did these CCAR-P practice questions catch you? A miss map by domain

Score the ten, then count misses by domain rather than in total. Ten CCAR-P questions cannot produce a percentage that means anything against a scaled cut of 720, and the score report gives percent-correct by domain, so that is the unit to practise reading. Each row names the items here, the objective to reread if you missed one, and the free mock's untimed filter for that domain.

DomainItems hereYour missesObjective to rereadUntimed drill
D1 Solution Design & Architecture (17%)1 and 2__ of 2O06: scope aligned to a business value pillar; O04: multi-agent orchestration designDomain 1 untimed
D2 Claude Models, Prompting & Context Engineering (13%)3__ of 1O10: context window and token optimisationDomain 2 untimed
D3 Integration (19%)4 and 5__ of 2O13: authentication and authorisation gaps; O18: choosing the integration mechanismDomain 3 untimed
D4 Evaluation, Testing & Optimization (16%)6 and 7__ of 2O22: A/B testing a change; O24: token, latency and cost trade-offsDomain 4 untimed
D5 Governance, Safety & Risk Management (14%)8__ of 1O26: guardrails and safety controlsDomain 5 untimed
D6 Stakeholder Communication & Lifecycle Management (14%)9__ of 1O34: documentation and implementation guidanceDomain 6 untimed
D7 Developer Productivity & Operational Enablement (7%)10__ of 1O36: configuring Claude tools for a teamDomain 7 untimed
Count misses by domain: none means move on, one means reread the objective, two means run the domain filter untimedScore the ten itemsthen count misses by domain0 missesMove onThe full mock rechecks it1 missReread the objectiveThen redo the item cold2 missesRun the domain filterUntimed, in the free mock
A Cred Farmer reading of a ten-item score, not an official rule.

Domains with one item here cap at one miss, so treat a miss there as a reread, not a verdict. Two misses in the same domain is the pattern worth acting on: run that domain's filter untimed, then sit the full timed form. The two-week CCAR-P study plan schedules the rereads, the CCAR-P cheat sheet compresses each domain to its decision rules, and why exam dumps are worth so little explains what a memorised answer letter costs.

Ten items are a sample, not a score

Two misses in the same domain is a pattern; one miss anywhere is noise until the full 63-item form confirms it. Nothing on this page is a scaled score, a pass verdict or a readiness figure, and it does not predict an official result.

Question bank dated . Checked against the CCAR-P Exam Guide v1.0 (July 2026) on .

Cred Farmer is not affiliated with Anthropic; the CCAR-P page records what the official sources say and when they were last read.

Key takeaways

  • Read the constraint before the options. Every item here turns on a stated fact: a committee decision, a fixed quality bar, a credential's scope.
  • Count misses by domain, not in total. Ten items give no percentage worth quoting against a scaled cut of 720.
  • The bigger knob is usually the trap. A larger window, a stronger model or a stricter filter each treats a symptom the scenario has already explained.
  • Deterministic decisions stay out of the model. Authorisation, pre-execution checks and team-wide rules hold only when a system enforces them.
  • Practice evidence is not a pass probability. Only measured answers on aligned timed forms move a Cred Farmer readiness score, and it does not predict an official result.

Sit the whole 63-item form under the clock

The free CCAR-P mock is the full 63 items in 120 minutes, no account, with a reason for every option and a raw count by domain beside the official weight. Signed in, five blueprint-aligned timed forms add 315 more items; only measured answers on those forms move a Cred Farmer readiness score, which does not predict an official result.

Start the free CCAR-P mock

Already scored the mock? Sign in for the five timed forms.

Questions

Frequently asked

The follow-up questions people search next.

Are these real CCAR-P exam questions?

No. All ten were written fresh for this post from the CCAR-P Exam Guide v1.0 and its published objectives; none comes from a live exam form or from the gated Cred Farmer bank. The guide's own three sample items are illustrative too, and it says they are not drawn from the live item bank.

How many questions are on the CCAR-P exam?

The CCAR-P has 63 items in 120 minutes, multiple-choice and multiple-response, and each item states how many responses to select. The result is a pass or fail with a scaled score from 100 to 1,000; the cut is 720. The fee is $175 and the credential is valid for 12 months.

Is there a free CCAR-P mock exam?

Yes. The free Cred Farmer mock at /practice/ccar-p runs the full 63-item form under a 120-minute clock with no account, then shows a raw count by domain and a reason for every option. It never reports a scaled score or a pass verdict, because no practice percentage converts to the 720 cut.

How should I score myself on these ten CCAR-P questions?

Count misses by domain, not in total, and treat a select-two item as a miss unless both picks match the key. One miss in a domain means reread the objective in the miss map and redo the item cold a day later; two means run that domain's untimed filter in the free mock first.

What score do I need on practice questions to pass the CCAR-P?

No practice percentage converts to the exam's scaled cut of 720, and Cred Farmer publishes none. Use practice to find the domains that need work. Signed in, only measured answers on blueprint-aligned timed forms move a readiness score, which is practice evidence; it does not predict an official result.

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.

We use cookies and privacy-friendly analytics to understand usage and improve Cred Farmer. Essential features work either way. See our Cookie Policy.