28 free CCAR-P sample questions

1. Q1 · Integration · 19.0%

   A field-service agent has grown to thirty-four tools over three releases. Traces from the last month show eleven of them called fewer than five times each, and the misroutes engineers complain about cluster on a group of tools with similar names. The team has one sprint to spend on the agent and wants to start where the evidence points.

   • AGroup the thirty-four definitions into six named categories in the system prompt so the model narrows its choice first.
   • BPull the call count and the misroute rate for each of the thirty-four, then retire and merge on what those two show.correct
   • CRaise the reasoning effort on this route so the model spends longer deciding which of the thirty-four tools applies.
   • DRewrite every tool description onto one template stating the arguments, the return shape and the system behind it.

   Bloat has two separable costs and they need different remedies. Tools nobody calls cost schema tokens on every request and are candidates for removal; tools that overlap cost accuracy and are candidates for merging. The two measurements tell the team which of the thirty-four is in which category, and a sprint spent without them is a guess.

2. Q2 · Integration · 19.0%

   A platform team is standing up the company's first shared agent. Tool requests have arrived from six departments, and the current draft attaches every internal API the company runs so that the agent can do anything a person could. An architect is asked for a rule the team can apply to decide what ships in the first release.

   • AAttach the read endpoints from all six departments now, and add the write endpoints once the agent has run for a quarter.
   • BAttach whatever each department asks for and rely on the tool descriptions to keep the agent off the ones it should not use.
   • CAttach the smallest set that lets the agent finish the tasks it was commissioned for, and add to it when a task provably fails.correct
   • DAttach one general tool per department that forwards the request onward and lets that department's own service route it.

   Every tool in the set is paid for on every request in schema tokens and in the model's selection accuracy, whether or not it is used. Commissioned tasks are the only evidence available before launch that a capability earns its place, and a failure to complete one is the evidence that justifies adding the next.

3. Q3 · Integration · 19.0%

   A shared engineering assistant carries tools for code review, deployment, incident triage and cost reporting. Incident work misroutes often; review work is reliable. The team has agreed to split the assistant into several and wants the split to reduce misroutes rather than relocate them. Which TWO choices should shape it? Select TWO.

   Select 2 answers

   • ADraw the boundaries so that an incoming request carries enough for a router to tell the four kinds of work apart.correct
   • BGive each of the four resulting assistants the whole tool set, so any of them can finish a request without a handoff.
   • CKeep one assistant and add worked examples drawn from the incident requests that were misrouted over the last quarter.
   • DLeave the tools that only incident work uses out of the review assistant, so that path cannot reach for them at all.correct
   • ESend every request to the incident assistant first, on the grounds that incidents are where a wrong tool costs the most.

   A split only helps if the classification it depends on can be made from the request itself, and if each resulting configuration is genuinely narrower than the original. Without the first the misroute moves to the router; without the second each assistant faces the same choice it faces today.

4. Q4 · Integration · 19.0%

   An assistant reads a partner's data feed. The partner changed the meaning of an amount field from minor units to major units without changing its type, and the assistant's figures were wrong for a fortnight before customers reported it. The team is choosing two standing signals that would have surfaced this earlier. Which TWO? Select TWO.

   Select 2 answers

   • AThe rate at which advisers override the figure the assistant suggests, tracked daily against its own baseline.correct
   • BThe share of results from the partner's feed that fail the schema check the harness applies before using them.
   • CThe number of requests the assistant serves each day, compared against the same day of the previous week.
   • DThe distribution of the values arriving in that field, watched for a shift against its recent history.correct
   • EThe ninety-fifth percentile latency of the partner's feed, on the grounds that such changes accompany a deployment.

   The change was structurally valid, so everything that inspects shape carried on passing. What moved was the values themselves and what people did with the results, so a distribution watch catches it at the boundary and the override rate catches it at the outcome. Together they bracket the pipeline between input and consequence.

5. Q5 · Solution Design & Architecture · 17.0%

   A distributor's operations director asks for an assistant that "handles supplier queries". Walking the process, you find eleven inbound query types; four of them carry most of the volume. Every type ends the same way, with an update in the ERP that only two trained staff can make. The programme has one quarter.

   • ACover all eleven query types as drafted replies for the team to send, and leave every ERP update with the two trained staff as it works today.
   • BCover the four highest-volume types end to end, including the ERP update, and leave the remaining seven queries with the team for now.correct
   • CStart with the two rarest types, where a mistake costs least, and widen coverage once the pattern has proved itself over a full quarter.
   • DClassify all eleven types on arrival and route each into the right team queue, leaving drafting and the ERP update with staff as now.

   Everything in this process funnels through one step that two people can perform, so that step is what determines how much work actually clears. A slice that goes all the way through it converts into throughput; anything that stops short of it just produces more material for the same queue. Taking the four types that carry the volume makes the first quarter's result large enough to read, and the seven remaining types stay exactly as they are rather than being half-changed.

6. Q6 · Solution Design & Architecture · 17.0%

   An underwriting assistant drafts risk summaries. Underwriters edit them before filing, and only the filed version is written to the case system — the draft is overwritten. Six months on, nobody can say whether the drafts have got better, and each proposed prompt change is argued from recollection.

   • AAdd a rating control beside each draft and review the scores by prompt version monthly, so underwriters have somewhere to register dissatisfaction.
   • BScore a nightly sample of drafts with a model against a written rubric, and treat the trend in that score as the quality measure for releases.
   • CSend underwriters a short weekly survey about draft quality, so problems are described while the cases they came from are still fresh and easy to look up.
   • DStore the generated draft alongside the filed version against the same case, so what underwriters changed becomes both the measurement and the examples.correct

   The underwriters are already producing the highest-quality signal available — a corrected version of every draft, made by the people whose judgement the system is imitating — and the design throws it away. Keeping both versions turns ordinary work into a labelled pair per case, which supports a measurement that moves when quality moves and supplies concrete examples of what good looks like. No new task is added to anybody's day.

7. Q7 · Solution Design & Architecture · 17.0%

   A coordinator hands code-migration sub-tasks to worker agents that share its filesystem but not its conversation. Workers keep stopping to ask about decisions the coordinator settled an hour ago, and two workers produced changes in visibly different styles. Which TWO changes to the delegated task would address this? Select TWO.

   Select 2 answers

   • AState in each task the exact files to change and the condition the finished work has to satisfy for the coordinator to accept it.correct
   • BAttach the coordinator's conversation so far to each task, so a worker starts with the whole background the coordinator has built up.
   • CSpell out the conventions the output must follow in the task itself, rather than relying on a worker to infer them from the code it opens.correct
   • DHave workers pause and put a question to the coordinator whenever something is unclear, with the coordinator answering while the run continues.

   A worker begins with the repository and the sentence it was sent, and nothing else — the coordinator's reasoning, its decisions and its sense of house style are all invisible. Anything the work depends on therefore has to travel in the brief: the scope, what finished means, and the conventions the result is judged by. That is what stops one worker asking and another guessing, and it is the discipline that makes delegation cheaper than doing the work inline.

8. Q8 · Solution Design & Architecture · 17.0%

   A team is deciding between a fixed workflow and an agentic design for handling escalated support cases. The route through a case varies: some need one system checked, some need five, and which ones depends on what earlier checks return. Every action the design would take is a read, and a wrong answer is caught by the customer's reply. Select TWO.

   Select 2 answers

   • AThe route cannot be enumerated up front, because which checks matter depends on what previous checks turned up in that particular case.correct
   • BMistakes surface quickly and are cheap to put right, so the design does not have to be correct on the first attempt to be worth running.correct
   • CEscalated cases arrive in enough volume that the engineering effort of building either shape would be repaid within the first year of operation.
   • DAn agentic design can be extended later without rework, which matters because support processes are revised more often than most other workflows.

   Two conditions justify paying for an agent's extra cost, latency and unpredictability: the path genuinely cannot be laid out in advance, and errors are detectable and recoverable so the freedom to explore does not become the freedom to do damage. This case has both — the checks depend on each other's results, and every action is a read whose mistakes come back as a customer reply. Take either condition away and a fixed route becomes the better buy.

9. Q9 · Evaluation, Testing & Optimization · 16.0%

   A Claude-backed triage step flags transactions for manual fraud review. Roughly one in two hundred is genuinely fraudulent. The team reports 99.1 percent accuracy on their evaluation set and the fraud team says the tool is missing cases they care about. Both statements are supported by the data.

   • AReport how many fraudulent transactions are caught and how many clean ones are flagged, as two figures.correct
   • BRebalance the evaluation set so fraudulent and clean transactions appear in roughly equal numbers in it.
   • CRaise the threshold at which a transaction is flagged, so the accuracy figure reflects the review capacity.
   • DWeight the accuracy calculation so a missed fraud counts for more than a wrongly flagged clean transaction.

   At a base rate of one in two hundred, a step that flags nothing at all scores 99.5 percent, so the headline figure is dominated by the majority class and says almost nothing about the behaviour the fraud team is judging. The two error types have different costs and different owners, so they need reporting separately: catch rate for the fraud team, false-flag rate for whoever staffs the review queue.

10. Q10 · Evaluation, Testing & Optimization · 16.0%

    A multilingual support assistant is scored on a set drawn proportionally from production traffic, which is 78 percent English. A prompt revision raises the overall score by three points and ships. Complaints from the German and Japanese markets rise sharply over the following fortnight while the headline score stays up.

    • ASet a minimum sample size per language and treat any language below it as out of scope for the gate.
    • BScore the revision on a second set drawn from the fortnight's complaints, which represents the affected users.
    • CRaise the proportion of non-English cases in the set until each language carries roughly the same weight.
    • DScore and report each language separately, so a gain concentrated in one cannot conceal a loss in another.correct

    A single number over a proportional sample lets a large improvement in the majority segment pay for a regression in the small ones, and nothing in the reporting shows it happening. Per-segment scores make the trade visible before release, and the gate can then require that no segment falls even where the aggregate rises.

11. Q11 · Evaluation, Testing & Optimization · 16.0%

    An architect is choosing which measurements may block a release of a Claude-backed service and which are reported for information. The candidates are answer quality on a scored set, the rate of unsupported claims, mean tokens per request, and the share of sessions ending in an escalation. Which TWO are suitable as blocking gates? Select TWO.

    Select 2 answers

    • AAnswer quality on the scored set, which is measured before release on a fixed set of cases.correct
    • BThe rate of unsupported claims, which is measured on the same set with the sources available.correct
    • CMean tokens per request, which is computed from usage figures on the release candidate's traffic.
    • DThe share of sessions ending in escalation, which is computed from production sessions each week.
    • ETime to first token at the ninety-fifth percentile, which is observed once real traffic is flowing.

    A blocking gate has to produce a verdict before the release goes out and has to move for reasons the release controls. Quality and groundedness on a fixed scored set satisfy both: they are available pre-release and a change in them is attributable to the candidate. Escalation share and observed tail latency are only available afterwards, and token count is a cost signal rather than a statement about whether the release is fit to ship.

12. Q12 · Evaluation, Testing & Optimization · 16.0%

    A team is designing how a model-graded scorer will be used to gate releases, and wants its verdicts to be defensible when a blocked release is challenged. Which TWO practices most directly support that? Select TWO.

    Select 2 answers

    • AMeasuring how often the scorer agrees with human reviewers on a held-out sample.correct
    • BStoring the scorer's per-case verdicts and reasons so a blocked release can be examined.correct
    • CRunning the scorer at the highest reasoning depth so its verdicts are as considered as possible.
    • DUsing the same model for scoring as for generation so both share an understanding of the task.
    • EReporting only the aggregate score so individual case disagreements do not derail the review.

    Defensibility rests on two things: evidence that the instrument tracks human judgement, and a record specific enough to argue with. Agreement measured against reviewers on a held-out sample supplies the first, and per-case verdicts with reasons supply the second, so a challenge becomes a discussion about particular cases rather than about whether the scorer can be trusted at all.

13. Q13 · Governance, Safety & Risk Management · 14.0%

    A bank runs three Claude-backed surfaces — an internal adviser assistant, a customer chat, and a batch letter generator — each owned by a different team. Each grew its own output-screening instructions. A wording rule compliance added last month reached only two of them, and nobody noticed for six weeks. You are asked to restructure this.

    • AGive the compliance team direct edit rights on all three system prompts, so that a new rule reaches every surface on the day it is agreed rather than weeks later.
    • BMove screening behind one service the three surfaces call on the generated text, with the rule set versioned centrally and each surface pinned to a released rule version.correct
    • CPublish the compliance wording as a shared prompt fragment that all three teams import into their own system prompts at build time, so the wording stays identical everywhere.
    • DAdd a nightly job that diffs each surface's system prompt against the current compliance rule set and raises a ticket against whichever team has drifted behind.

    A rule that lives inside three separately-owned prompts is three implementations that drift. Making screening a called service turns the rule set into one deployable artefact with a version each surface pins to, so a compliance change is a release with a known blast radius, and coverage is a property of the call graph rather than of whether three teams remembered. It also gives one place to log what was screened and why.

14. Q14 · Governance, Safety & Risk Management · 14.0%

    An operations agent can restart services, rotate credentials, and delete storage buckets through tools. Its system prompt says bucket deletion requires an operator to confirm in chat first. In a rehearsal, an operator typed a paragraph describing cleanup goals and the agent deleted a bucket without asking. Where should the requirement live?

    • AIn a post-run reconciliation job that compares deletions against the approval log every hour and restores anything deleted without a matching operator confirmation recorded.
    • BIn a stronger system prompt that restates the confirmation requirement immediately before the tool list and repeats it in the delete tool's own description field.
    • CIn the tool layer, which refuses the delete call until it receives an approval token issued by a separate operator action rather than by anything the model produced.correct
    • DIn a second model call that reviews the proposed delete against the transcript and decides whether the operator's message amounted to an explicit confirmation.

    Anything the model can decide, the model can be talked out of — including whether a confirmation happened. Moving the requirement to the tool layer makes approval a fact the harness holds rather than a conclusion the model draws: the call cannot execute without a token minted by an operator action outside the conversation. That is the difference between steering and enforcement, and irreversible actions need enforcement.

15. Q15 · Stakeholder Communication & Lifecycle Management · 14.0%

    A operations director opens a kick-off meeting by saying the company needs an AI assistant for its support team, and that the board has approved budget for the quarter. She has brought a slide listing features she has seen elsewhere. You have an hour with her and two team leaders.

    • AEstablish what the support team does today, where the time actually goes, and which of those points a wrong answer would be tolerable at — before discussing features.correct
    • BWalk through the feature list and establish which items are feasible on the current data, so the quarter's scope is agreed before the team commits to any of it.
    • CAgree the success measure for the quarter, so the work has a target and the director can report progress to the board against something concrete and defensible.
    • DMap the systems the support team uses and what each holds, since the assistant's usefulness will be bounded by which of them it can reach reliably.

    A feature list is a stakeholder's hypothesis about a solution, and building against it means inheriting an analysis nobody has checked. Understanding the current work, where effort concentrates, and where error is survivable tells you which features would actually matter — and it is the only part of this that is hard to recover later. Everything else can be revisited once the problem is understood.

16. Q16 · Stakeholder Communication & Lifecycle Management · 14.0%

    A team is scoping an assistant that will summarise engineering incident reports for a monthly reliability review. The sponsor wants it to produce good summaries. Two engineers on the call describe good differently — one means comprehensive, the other means short enough that people read it.

    • AHave the sponsor arbitrate, since the disagreement is about what the review needs and the sponsor owns the review and its audience.
    • BBuild both variants and let the review group choose after a month, so the decision is made on real summaries rather than in the abstract.
    • CCollect a set of past incidents with summaries the review group considers good, and use them to define the target concretely enough to evaluate against.correct
    • DAdopt the shorter definition, since a summary that is not read has no value regardless of how comprehensive it is or how accurate its content.

    Quality words like good and comprehensive are placeholders for a judgement nobody has written down, and arguing about them in the abstract rarely converges. Worked examples the audience endorses turn the disagreement into something inspectable — you can see whether the two engineers actually differ or are describing the same output — and they double as the evaluation set the build will need.

17. Q17 · Governance, Safety & Risk Management · 14.0%

    An engineering platform team is about to let an agent apply infrastructure changes overnight without a human present. They have capacity for two changes before the first unattended shift and want the two that most reduce the consequence of a bad run rather than its likelihood. Select TWO.

    Select 2 answers

    • AScope the credentials the agent runs under to the specific resource types and environment the task needs, rather than the operator role it currently borrows.correct
    • BRaise the model's effort setting for this workload so the agent reasons more carefully about each change before it decides to apply anything at all.
    • CRequire the agent to produce a written plan and post it to the team channel before applying, so the record exists for review the following morning.
    • DPut a hard ceiling on the run so it terminates rather than continuing to act once it passes an agreed number of applied changes or spend threshold.correct
    • EAdd examples of previous safe changes to the system prompt so the agent has concrete precedent for the shape of change the team considers routine.

    Both chosen controls bound what a bad run can reach: narrow credentials limit which resources are touchable at all, and a hard run ceiling limits how much damage accumulates before something stops. Neither depends on the agent behaving well, which is the property you need when nobody is watching. The others make a good run more likely without changing what a bad one costs.

18. Q18 · Governance, Safety & Risk Management · 14.0%

    A team is designing the review step for an assistant that drafts responses to regulatory correspondence. Two changes are proposed and both fit this quarter. The team wants the pair that makes a reviewer's disagreement with the draft likely to be expressed rather than absorbed. Select TWO.

    Select 2 answers

    • AMake rejection as cheap as acceptance, so returning a draft does not require the reviewer to write the replacement themselves before they can move on.correct
    • BRecord the reviewer's edits and rejections as structured signal feeding the next iteration, so disagreement changes the system rather than only the document.correct
    • CShow the reviewer the model's confidence in each section, so attention is drawn to the parts the assistant itself considers least certain before they read.
    • DAdd a second reviewer for every draft, so a reviewer inclined to accept is paired with a colleague who may take a different view of the same wording.
    • ESet a minimum review time per draft, so a reviewer cannot clear an item faster than it takes to read the correspondence it responds to.

    Disagreement is absorbed when it is expensive to express and pointless to record. Making rejection cheap removes the incentive to accept a mediocre draft rather than take on the rewrite, and capturing edits as structured signal means an objection improves the system instead of dying in one document. Together they change both the cost and the consequence of dissent, which is what the team asked for.

19. Q19 · Stakeholder Communication & Lifecycle Management · 14.0%

    A steering group will decide next week whether to fund an assistant for a legal team's contract review. The architect has one week of discovery. They want the two findings that would most change the decision, rather than the two that are easiest to produce. Select TWO.

    Select 2 answers

    • AWhether examples of correctly reviewed contracts exist in enough volume and consistency to define the target and measure against it.correct
    • BWhat happens when a review is wrong — who is affected, how it is caught today, and how far the consequence can be reversed once it has happened.correct
    • CHow long a lawyer currently spends on a contract review, so the saving per review can be estimated and multiplied across annual volume.
    • DWhich contract management systems the team uses, and whether their interfaces would allow an assistant to read and write review records.
    • EHow the legal team feels about the proposal, since adoption will depend on their willingness to use whatever the programme delivers.

    Both findings can stop the project. Without a body of consistent worked examples there is no way to define correct or to evaluate, which makes the build unmeasurable; without knowing what a wrong review costs and whether it can be undone, nobody can say what level of autonomy is acceptable or how much review is needed. Either answer can make the proposal unfundable, which is what a decision-changing finding means.

20. Q20 · Stakeholder Communication & Lifecycle Management · 14.0%

    A supplier is drafting the service schedule for an assistant it will operate for a client. The architect is asked which two commitments the supplier can hold across a three-year term without depending on things outside its control. Select TWO.

    Select 2 answers

    • AA maximum time to acknowledge and begin work on an incident, measured from the point the client raises it through the agreed channel.correct
    • BA notice period and a defined process before any material change to the assistant's behaviour, including a change of underlying model.correct
    • CA minimum measured accuracy against the client's live traffic, reported monthly and backed by a service credit when the figure is not met.
    • DA maximum response time for the assistant, held constant across the term regardless of how the client's query volume or complexity develops.
    • EA commitment that the assistant's answers will not require correction by the client's staff more than a stated number of times per month.

    Both commitments are about the supplier's own conduct — how fast it responds to an incident, and how it manages change — so they hold regardless of what the client's traffic does or how the model behaves. The rejected options all depend on things the supplier does not control: the mix of questions asked, how volume grows, and how the client's staff judge an answer.

21. Q21 · Claude Models, Prompting & Context Engineering · 13.0%

    A payments platform runs eleven Claude-backed services, every one of them pinned to the same model inside a shared client library. A regional capacity event degrades all eleven inside the same minute, and restoring service takes forty minutes of hand-editing. The architecture review asks what to change first.

    • AGive each service tier a second model that carries a slice of real traffic daily, so the path is proven.correct
    • BRaise the organisation's rate limits on the pinned model well above peak so capacity events stop biting.
    • CQueue rejected requests during a degradation and replay them once the pinned model returns to normal.
    • DSplit the eleven services across separate API keys so each one's capacity is accounted for on its own.

    Eleven services sharing one model share one failure. The architectural defect is the correlated dependency, and the reason recovery took forty minutes is that the alternative had never run. A second model that carries a small continuous share of production traffic is the only kind of fallback that is known to work when it is needed, because its prompts, its output shape and its cost are already measured.

22. Q22 · Claude Models, Prompting & Context Engineering · 13.0%

    A document-processing service handles a mix where roughly one request in twenty is genuinely hard. A proposal adds a small classifier call that reads each request and picks between a cheap and an expensive model. Modelling shows the classifier itself adds about 300ms and misroutes around eight percent of items. What should the architect weigh?

    • AWhether the two models return the same output shape so the downstream parser needs no changes.
    • BWhether the classifier call can share a cache prefix with the request it is about to classify.
    • CWhether the classifier can be trained often enough to keep its misroute rate below five percent.
    • DWhether misrouted hard items cost more than the routing saves, given how few of them there are.correct

    A router pays a fixed tax on every request to save money on some of them. With one hard item in twenty, the savings pool is the ninety-five percent that get the cheap model, while eight percent misrouting sends a meaningful share of the hard items to a model that will handle them badly. The comparison the design turns on is the expected cost of a wrong route against the expected saving, and if that is unfavourable the cheaper answer is a static assignment.

23. Q23 · Claude Models, Prompting & Context Engineering · 13.0%

    A team is hardening a Claude-backed workflow that issues payments, sends customer email and reads an internal wiki. They are deciding which properties to enforce outside the model rather than request in the prompt. Which TWO belong outside it? Select TWO.

    Select 2 answers

    • AThat a payment above an agreed threshold is not issued without a recorded human approval.correct
    • BThat every outbound customer email carries the current regulatory footer word for word.correct
    • CThat the assistant's tone in customer email stays formal and avoids contractions throughout.
    • DThat wiki passages are summarised rather than quoted at length when the answer is long.
    • EThat the assistant asks a clarifying question when the customer's request is ambiguous.

    The test is whether a violation is tolerable. A payment above threshold and a missing statutory footer are both defects with legal consequence, and neither can be left to a system that produces its output probabilistically, so both are enforced by the harness and by code that appends the footer after generation. Tone, summarisation style and when to ask a question are judgements where an occasional miss is a quality issue, which is what prompting is good at.

24. Q24 · Claude Models, Prompting & Context Engineering · 13.0%

    An architect is deciding what belongs in a skill rather than in an agent's system prompt. The agent handles general engineering questions all day, and about twice a week it is asked to produce a formal architecture decision record, which has a strict template and a long house style guide. Which TWO belong in the skill? Select TWO.

    Select 2 answers

    • AThe decision-record template, with its required sections and the order they appear in.correct
    • BThe house style guide covering how options and consequences should be written up.correct
    • CThe instruction to answer in British English, which applies to everything the agent writes.
    • DThe list of repositories the agent may read, which governs every task it is given.
    • EThe rule that the agent asks before writing to any file, which applies on every turn.

    The test is how often the material is needed. A template and a style guide for a document produced twice a week are exactly what on-demand loading exists for: their descriptions sit in context so the agent knows they exist, and their bulk is read only on the days a record is written. Language, permitted repositories and a write-confirmation rule apply on every turn, so they belong where they are always in force.

25. Q25 · Developer Productivity & Operational Enablement · 7.0%

    Six weeks after Claude Code was rolled out to 240 engineers, the CTO asks whether it is working. The platform team has one figure: 240 licences assigned. Telemetry export to the company's observability stack was configured at rollout and nobody has looked at it since. The CTO wants something at Thursday's leadership meeting.

    • AReport `claude_code.session.count` and `claude_code.active_time.total` by team, since an assigned licence nobody has opened and one used every day are the same number today.correct
    • BReport `claude_code.lines_of_code.count`, since it measures how much of the codebase the tool is now producing and rises directly with the amount of work being delivered.
    • CReport the change in pull requests merged per engineer against the six weeks before rollout, since that is the outcome the investment was justified on originally.
    • DSurvey the 240 engineers on whether they find it useful and report the distribution, since perceived value captures benefits that exported metrics cannot express well.

    Nothing can be said about impact until you know the tool is being opened, and the licence count cannot distinguish an engineer using it daily from one who logged in once. Session and active-time figures broken down by team also show where adoption is thin, which is the actionable part — an enablement problem in two teams looks nothing like a product that nobody wants.

26. Q26 · Developer Productivity & Operational Enablement · 7.0%

    A platform team wants Claude to review every pull request across twelve repositories and leave comments. An engineer has started writing a CI job that shells out to the CLI, parses the output and posts comments through the GitHub API. He estimates a fortnight and has three days done.

    • AHave him continue but standardise the job as a reusable workflow the twelve repositories call, so the parsing and posting logic exists once rather than twelve times over.
    • BReplace the CI job with a `PostToolUse` hook in each repository, so review happens as changes are made locally rather than after a pull request has been opened.
    • CInstall the maintained GitHub integration with `/install-github-app` and configure the supplied workflow, keeping the bespoke job only if a requirement turns out not to be covered.correct
    • DKeep the CI job and drop the comment posting, publishing the review as a build artefact that reviewers open from the checks tab when they want a second opinion.

    A maintained integration exists for exactly this, and the fortnight is being spent rebuilding output parsing, comment posting and authentication that someone else already maintains. The architect's job here is to stop the team acquiring a bespoke component with an owner, a failure mode and an upgrade burden, and to reserve custom work for the requirement the supported path genuinely does not meet.

27. Q27 · Developer Productivity & Operational Enablement · 7.0%

    You are enabling Claude Code across an organisation of four hundred engineers next quarter. The budget covers licences and two people for a term. Leadership expects the two to run training. You are asked which two commitments to make on the enablement side, given what usually determines whether a rollout like this holds. Select TWO.

    Select 2 answers

    • AA supported default configuration that a team gets without doing anything, so the first day's experience does not depend on each team assembling one for itself.correct
    • BA named route for engineers to raise problems and get an answer, so early friction is resolved rather than becoming the reason a team quietly stops using the tool.correct
    • CA training curriculum delivered to every team in the first month, so all four hundred engineers start from the same understanding of what the tool does and does not do.
    • DA certification requirement before an engineer is given access, so nobody uses the tool on production code before demonstrating they understand its behaviour.
    • EA monthly newsletter of techniques and worked examples, so practice spreads across teams rather than being rediscovered independently by each of them.

    Rollouts of this kind fail at the two points where an engineer meets friction alone: a first day spent configuring rather than working, and an unresolved problem in week two. A default configuration removes the first, and a real support route removes the second — both are the kind of infrastructure two people can sustain for four hundred engineers, unlike anything delivered person by person.

28. Q28 · Developer Productivity & Operational Enablement · 7.0%

    Nine months into an organisation-wide rollout, you are asked which signals would tell you it is going wrong before the annual review does. Usage is high and rising, spend is within budget, and no team has asked to stop. You want the two signals that would surface a real problem while there is still time to act. Select TWO.

    Select 2 answers

    • AA rising share of changes that are reverted or substantially amended shortly after merge, since that is quality being pushed past the review step rather than removed.correct
    • BA widening gap between the teams with the highest and lowest usage, since divergence indicates the enablement effort is reaching some teams and failing others.correct
    • CA falling number of questions in the support channel, since a quiet channel usually means engineers have stopped asking rather than stopped having problems.
    • DA rising average session length, since longer sessions indicate engineers are attempting larger pieces of work than the tool handles reliably in one pass.
    • EA falling proportion of suggested changes that engineers accept, since declining acceptance is the earliest indication that output quality has started to slip.

    Both signals move before the annual review notices anything and both point at something actionable. Post-merge rework says defects are getting past review, which is the failure that high usage figures actively conceal; a widening usage gap says enablement is failing specific teams while the aggregate looks healthy. The measures that look reassuring here — usage, spend, absence of complaints — are exactly the ones that cannot fall until it is too late.

Not affiliated with or endorsed by Anthropic. Domain names and weightings are taken from the published exam guide; always check the official guide before booking.