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

Answer as you read

The CCDV-F paper: every question

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

1. Question 1Domain 2: Applications and IntegrationSingle answer

   Northgate Bank's product owner states two requirements for a mortgage assistant: the first token of an answer must appear within three seconds at the 95th percentile, and every quoted rate must match the bank's rate table, which is updated several times a day as markets move. The team must turn these two statements into a design before any prompt is written. Which design meets both requirements exactly as stated?

   Select the best answer.

   AEmbed the rate table in the system prompt with prompt caching, refresh the prompt each night, and stream the answer from a fast model tier so tokens appear early.

   BQueue mortgage questions to the Message Batches API so every rate can be re-checked before results are released, and show the customer a progress indicator meanwhile.

   CFetch the current rate through a tool call at request time, then stream the answer from a fast model tier so the first tokens reach the customer early.

   DUse the highest-capability model with extended thinking so it reasons carefully about which rate applies to the question, and stream its final answer to the customer.
   Show the answer
2. Question 2Domain 1: Agents and WorkflowsSingle answer

   Dunmore Freight, a logistics firm, processes 9,000 customs declarations a day through four fixed steps: extract fields from the shipper's PDF, classify the goods, validate the classification against the tariff table, and file the declaration. Every step is known in advance, auditors need to see which step produced each value, and finance wants a predictable cost per declaration. An engineer proposes an autonomous agent that chooses its own steps with a tool loop. Which architecture better fits these constraints?

   Select the best answer.

   AA single autonomous agent with all four capabilities exposed as tools, so it can decide the order of operations and recover from unexpected document layouts on its own.

   BA manager agent that spawns a subagent for each declaration and lets every subagent choose its own tools, aggregating the results into one filing report.

   COne large prompt that asks the model to carry out every step within one response, returning the final declaration as structured JSON to cut per-call overhead.

   DA predefined workflow that chains one Claude call per step, with deterministic code between the steps that validates each output before the next step runs.
   Show the answer
3. Question 3Domain 2: Applications and IntegrationSingle answer

   Brightline Software's in-app help assistant returns answers of about 600 tokens. Users stare at an empty panel for roughly nine seconds and then the complete answer appears at once; session recordings show 22 percent of users navigate away before it arrives. The product manager wants the perceived wait reduced this sprint without changing the model tier and without shortening the answers. Which change delivers that?

   Select the best answer.

   AReduce max_tokens to 250 so the answer completes in roughly a third of the time, and add a link that lets the user request the remainder when they need it.

   BRequest the response with streaming enabled and render every text delta in the panel the moment it lands, so the opening words are visible inside the first second.

   CSwitch the help assistant to a quicker model tier, which lowers the time to the complete answer without requiring any change to the code that renders the panel.

   DAdd a cache breakpoint on the system prompt so the model begins generating sooner, which shortens the blank period without touching how the panel renders the answer.
   Show the answer
4. Question 4Domain 2: Applications and IntegrationSelect 2

   Granite Accounting's tax-query assistant is edited by three analysts through an admin console that writes the production prompt directly. Last Tuesday the assistant began omitting the required disclaimer; it took two days to discover that a wording change made on Monday had caused it, and nobody could say who made the change or what the earlier text had been. Which TWO practices address the cause of that outage? (Select two.)

   Select exactly 2. 0/2 selected.

   ARestrict console access to a single senior analyst so that only one person can make prompt changes and the source of any regression is immediately obvious.

   BStore the prompt in version control next to the application code, with every change made through a reviewed pull request that records who changed what and when.

   CPin the model to a specific version so the prompt behaves identically across releases, which prevents wording changes from having effects nobody anticipated.

   DRun the assistant's evaluation suite, including a check that the disclaimer is present, in the deployment pipeline so a prompt change that fails cannot reach production.

   EAsk the analysts to add a dated comment at the top of the prompt describing each edit, so the change history is visible inside the prompt text itself.
   Check answerShow the answer
5. Question 5Domain 1: Agents and WorkflowsSingle answer

   Alder Health, a hospital group, is deploying a discharge-summary agent. Its compliance office requires that the agent's runtime, tool execution and all patient records remain inside the hospital's own cloud account, with logs retained under the hospital's existing controls. The engineering lead wants the lowest operational burden and is weighing an Anthropic-hosted managed agent deployment against running the agent harness on the hospital's own infrastructure and calling the model over the API. Which deployment model satisfies the stated constraint, and for what reason?

   Select the best answer.

   ASelf-hosted: run the agent loop and tool execution inside the hospital's own account and call the model API from there, since the constraint governs where execution and records live.

   BAnthropic-hosted: the managed deployment satisfies the constraint, because model inference runs on Anthropic infrastructure in either design and the only difference is which party operates the harness.

   CAnthropic-hosted: the lead's goal of minimal operational burden is decisive, and the managed platform's own retention controls stand in for the hospital's logging requirements.

   DSelf-hosted: a hosted agent has no way to invoke tools that reach systems inside a customer's network, so the discharge-summary tools would be unreachable from the managed runtime.
   Show the answer
6. Question 6Domain 2: Applications and IntegrationSingle answer

   Dunmore Freight classifies shipment exceptions with a Claude tool call whose input schema declares the cause field as a free string. Over 10,000 classifications the field held 61 distinct spellings of six intended causes, such as Customs hold, customs-hold and Held at customs, and the routing code's exact-match lookup sent 18 percent of exceptions to a manual queue. The six causes are fixed and will not change this year. Which schema change fixes the routing?

   Select the best answer.

   AKeep the cause field as a free string and normalise it in the routing code with lower-casing and fuzzy matching, so all 61 spellings collapse onto the six causes.

   BAdd the six values to the tool description as examples and instruct the model to copy one of them verbatim, keeping the schema type as a free string.

   CSplit the cause field into six boolean fields, one per cause, so the model marks the applicable one and the router checks each flag in turn.

   DDeclare the cause field as an enum of the six permitted values in the input schema, so the model selects one and the router matches exactly.
   Show the answer
7. Question 7Domain 3: Claude CodeSingle answer

   Lantern Media, a publisher, wants Claude Code to review every pull request in its CI pipeline: read the diff, apply the repository's review checklist, and return findings as JSON that the pipeline posts as a comment. No engineer is at a terminal when it runs, and the job must never modify files or push commits. Which way of operating Claude Code fits this job?

   Select the best answer.

   AStart an interactive session on the CI runner with auto-accept enabled so no permission prompt blocks the job, and parse the terminal transcript afterwards to extract the findings.

   BDefine a custom slash command in the repository that a reviewer invokes from their own session, and have the pipeline wait until the resulting comment appears on the pull request.

   CRun Claude Code in headless mode with the prompt passed as an argument and JSON output selected, restricting the allowed tools to read-only operations for the duration of the job.

   DInstall Claude Desktop on the CI runner and connect it to the repository, so the review runs with the same permissions and context as a developer would have.
   Show the answer
8. Question 8Domain 1: Agents and WorkflowsSelect 2

   Wexford University's research office runs a grant-compliance agent that, for each of 120 active grants, pulls a spending ledger of roughly 30,000 tokens, checks it against the funder's rules, and writes a finding. Reviewing all grants in one session exhausts the context window by the fortieth grant, and the run takes four hours. The team plans to delegate each grant's check to a subagent. Which TWO benefits can they correctly expect from that change? (Select two.)

   Select exactly 2. 0/2 selected.

   AEach subagent works in its own context window, so a grant's 30,000-token ledger never accumulates in the parent's context; only the finding is returned to it.

   BSubagents share the parent's conversation history automatically, so each one already knows what earlier grant checks concluded without being told by the orchestrator.

   CThe grants are independent of one another, so subagents can run concurrently and the four-hour wall-clock time falls towards the duration of the slowest few checks.

   DA fresh context per subagent makes each finding deterministic across reruns, since none of the earlier grants' details can influence the model's sampling.

   ESubagents inherit the parent's instructions and the funder's rules implicitly, so the orchestration code needs to pass each one only the grant identifier.
   Check answerShow the answer
9. Question 9Domain 2: Applications and IntegrationSingle answer

   Ashcombe & Reade, a law firm, sends its 38,000-token contract-review playbook with every request, about 2,000 requests a day, and input tokens are 91 percent of the bill. The playbook changes once a month; the clause under review differs on every request and lawyers wait for the answer interactively. An associate added cache_control to the request but the cache hit rate stayed near zero. Which arrangement produces cache hits?

   Select the best answer.

   AKeep the request as it is but route it through the Message Batches API, where repeated content across the requests in a batch is cached automatically.

   BPlace the playbook first in the request as a stable prefix with the cache breakpoint at its end, and put the varying clause after that breakpoint.

   CPut the clause first so the model reads the specific question before the playbook, and set the cache breakpoint on the clause block itself.

   DCompress the playbook to a 4,000-token summary, which lowers the input cost enough that cache hits no longer matter to the monthly bill.
   Show the answer
10. Question 10Domain 2: Applications and IntegrationSingle answer

    Open Shelter, a nonprofit, refined a volunteer-intake prompt in claude.ai until answers were well formatted and consistently declined to give legal advice. Moving the same text into the user turn of an API call produced plain unformatted output and, in 6 of 40 test conversations, tentative legal guidance. The model is the same and no sampling settings were changed. What explains the difference?

    Select the best answer.

    AThe API applies no default system prompt and no product behaviours; whatever the consumer interface supplied implicitly must now be written into the application's own system prompt.

    BThe API serves a differently tuned build of the model than the consumer product, so the same prompt cannot be expected to produce equivalent behaviour across the two surfaces.

    CThe API defaults to a higher sampling temperature than the consumer interface, and lowering it will restore both the formatting and the refusal behaviour the team observed.

    DThe consumer interface post-processes model output into formatted text and filters legal content, so the application must add an equivalent filtering layer rather than change its prompt.
    Show the answer
11. Question 11Domain 1: Agents and WorkflowsSingle answer

    Cobalt Insurance wrote its own agent harness on the Messages API with a lookup_policy tool. In traces from 200 claims, the model requests the tool and the harness logs the tool_use block, but it never executes the lookup: it drops the block from the history and asks the model to answer anyway, so customers receive replies quoting coverage figures the tool never returned. Which correction to the loop fixes the behaviour?

    Select the best answer.

    ASet tool_choice so the lookup_policy tool is required on every request, which stops the model answering from its own knowledge before the tool has been run at least once.

    BIncrease the max_tokens limit by a wide margin so the model has room to request the tool and write the final answer inside a single response, removing the need for a second call.

    CLoad the full policy table into the system prompt so the model no longer depends on a tool call completing before it drafts the figures in its answer.

    DWhen a response ends with a tool_use block, run the tool, return a tool_result carrying the same id in the next user turn, and call the model again until it stops normally.
    Show the answer
12. Question 12Domain 2: Applications and IntegrationSelect 2

    Hollis Estates, a property listings company, must regenerate 250,000 listing descriptions before a site relaunch in five days and has chosen the Message Batches API for the cost saving. The engineer designing the pipeline has only used the synchronous Messages API before and wants to know what changes in application behaviour the batch design must accommodate. Which TWO statements accurately describe what the pipeline must handle? (Select two.)

    Select exactly 2. 0/2 selected.

    ARequests within one batch share a conversation, so the pipeline can send the style guide once and each later description request in the batch will inherit it.

    BBatch results are delivered as a stream of deltas per request, so the pipeline must reassemble partial text exactly as a streaming client would.

    CResults for a batch can complete in any order, so each request needs a custom_id and the pipeline must match results back to listings by that id.

    DBatch pricing applies only once a batch exceeds a minimum number of requests, so the pipeline must accumulate listings until that threshold is reached before it submits.

    EA batch may take up to 24 hours to finish, so the pipeline must poll or check batch status and tolerate that delay instead of blocking on each request.
    Check answerShow the answer
13. Question 13Domain 2: Applications and IntegrationSingle answer

    Piper Eats, a food-delivery company, produces a one-line summary of each of up to 300 reviews on a restaurant page. The service loops over the reviews with a synchronous client, one call after another, and the page's p95 build time is 45 seconds. Each summary is independent of the others, and the account's rate limit is well above the current request rate. What is the most direct way to cut the build time?

    Select the best answer.

    AConcatenate all 300 reviews into a single request and ask for one combined summary, so the page needs a single round trip regardless of how many reviews it shows.

    BSwitch each call to the fastest model tier, which shortens every individual summary and therefore shortens the sequential loop as a whole in proportion.

    CIssue the calls concurrently with the async client, bounded by a semaphore sized under the rate limit, and gather the results as they complete.

    DSpawn an unbounded thread per review so all 300 requests leave at the same moment, and add a retry loop to absorb whatever rate-limit responses come back.
    Show the answer
14. Question 14Domain 4: Eval, Testing, and DebuggingSingle answer

    Corriedale Telecom's integration retries every non-200 response from the Claude API immediately, up to ten times. During a traffic spike the logs showed 429 responses climbing with each retry, several 400 responses retried ten times with an identical malformed request, and one 401 retried ten times after a key rotation. Which retry policy matches each error type to a suitable recovery?

    Select the best answer.

    ARetry 429, 529 and other 5xx responses with exponential backoff and jitter, honouring any retry-after header; do not retry 400 or 401, and surface them for a fix.

    BRetry every non-200 response with a fixed 30-second delay between attempts, which spaces the requests out enough for the service to recover from any error type.

    CStop retrying entirely and fail each request back to the caller on the first non-200 response, since automated retries were what amplified the incident in the first place.

    DRetry 400 and 401 with backoff, since they indicate the service is under load, and fail 429 immediately because the rate limit will not clear within the lifetime of a request.
    Show the answer
15. Question 15Domain 1: Agents and WorkflowsSingle answer

    Barleycorn Grocers, a retailer, runs a Claude Code agent that edits its pricing-rules file. The CLAUDE.md instructs the agent to run the pricing linter after every edit to that file, yet in 11 of the last 60 sessions the linter never ran and a malformed rule reached staging. The team wants the linter to run every time the file changes, regardless of what the model decides to do. Which mechanism gives that guarantee?

    Select the best answer.

    ARelocate the instruction to the first line of CLAUDE.md, mark it as mandatory in capital letters, and duplicate it in the system prompt to give it extra weight with the model.

    BRegister a PostToolUse hook that runs the linter whenever an edit tool touches the pricing-rules file, since hooks execute as code outside the model's decision.

    CExpose the linter as a custom tool with a description stating it must be called after every edit, so the model has a clear and unambiguous signal to invoke it.

    DAdd a verification subagent that reviews each session transcript at the end and runs the linter if it finds that the main agent skipped the step.
    Show the answer
16. Question 16Domain 2: Applications and IntegrationSingle answer

    The Harbour City transit authority, a public agency, receives 60,000 complaints a month, each triaged by a clerk into one of 14 queues. A Claude-based triage system passed its offline evaluation at 93 percent agreement with clerks. The director wants the clerks reassigned next month. Misrouted safety complaints carry legal exposure, and there is no budget for a permanent parallel team. Which implementation approach best manages the transition?

    Select the best answer.

    ACut over fully on the planned date, since the offline evaluation already met the bar, and rely on production dashboards to catch any misrouting quickly.

    BKeep the system in shadow mode indefinitely and use its output only as a suggestion the clerks may accept, so that no complaint is ever routed by the model alone.

    CExtend the offline evaluation to a larger historical sample until agreement exceeds 99 percent, and defer any production deployment until that threshold has been reached.

    DRun the system in shadow mode against clerk decisions for a fixed period, promote it queue by queue where agreement holds, and keep a rollback path for each queue.
    Show the answer
17. Question 17Domain 2: Applications and IntegrationSelect 2

    Alder Health's clinical-notes assistant keeps one long conversation open per clinician for the whole shift and attaches a running shift memory of earlier encounters to each request; a clinician moves between 20 or more patients. In a review of 500 generated notes, 7 referenced a medication belonging to the previous patient. Retrieval returns the correct record each time, so the error arises from what else is in the context. Which TWO changes remove the cross-patient contamination? (Select two.)

    Select exactly 2. 0/2 selected.

    AAdd an instruction at the start of each encounter telling the model to forget everything about the previous patient before it reads the new record.

    BOpen a fresh conversation for every patient encounter, so no earlier patient's record or generated note remains in the context when the next note is written.

    CAdopt whichever model tier offers the largest context window, so the earlier patients' details are less likely to crowd out the current record.

    DEnable prompt caching on the conversation so the earlier turns are served from cache and do not influence generation for the current patient.

    EScope each request to the current patient's record and keep patient details out of the shift memory that persists across encounters.
    Check answerShow the answer
18. Question 18Domain 1: Agents and WorkflowsSingle answer

    Ashcombe & Reade built a due-diligence agent that reviews acquisition contracts against a six-category checklist. On a 400-contract data room the single agent's answers were sound for the first 30 contracts, then it began skipping categories and citing clauses from earlier contracts; token usage per contract had grown fivefold by the sixtieth. The partners need a report per contract and one consolidated risk summary. Which change to the architecture addresses the degradation?

    Select the best answer.

    AKeep one agent but enable compaction so the conversation is summarised whenever it approaches the context limit, preserving the checklist, the open risks and the most recent findings.

    BUpgrade the agent to the top capability tier so that it can hold more contracts in working memory before the quality of its reviews begins to decline.

    CIntroduce a supervisor that dispatches each contract to a fresh subagent with the checklist and only that contract, then merges the returned structured findings into the consolidated risk summary.

    DWrite each completed review to a memory file and instruct the agent to consult the file before every new contract so it stops confusing one with another.
    Show the answer
19. Question 19Domain 2: Applications and IntegrationSingle answer

    Northgate Bank's procurement policy requires all model inference to run through the cloud provider where the bank already holds contracts, identity controls and audit logging, so the team must call Claude through that provider's platform rather than through Anthropic's API directly. The prototype was built against the first-party SDK. What should the team expect when they move the integration?

    Select the best answer.

    AThe same Claude models are reached through the provider's endpoint, but client configuration and authentication move to the provider's identity system, so code and credentials change while prompts carry over.

    BThe provider hosts distinct Claude variants that were trained specifically for that platform, so the prompts and the evaluation suite must be rebuilt from scratch before any results can be compared.

    CThe move is transparent: the first-party SDK detects the provider from the API key and switches endpoints, so no code changes and no authentication changes are required at all.

    DThe provider exposes Claude only through batch processing, so the bank's interactive use cases must be redesigned around asynchronous jobs before the migration can begin.
    Show the answer
20. Question 20Domain 2: Applications and IntegrationSelect 2

    Emberforge Games, a studio, is using Claude Code to migrate 900 source files from callback style to async/await. The first pass changed 200 files in one session and produced a single 14,000-line commit; the test suite then reported 31 failures, and finding which change caused each one took a week. The lead must plan the remaining 700 files. Which TWO practices should the plan adopt? (Select two.)

    Select exactly 2. 0/2 selected.

    AWork in small batches of related files, run the test suite after each batch, and commit only when the batch is green.

    BRun the remaining 700 files in one session at the highest model tier, so the model keeps every file in context and applies the pattern consistently.

    CHave the model regenerate the affected tests alongside each migrated file so the suite always matches the new code and stops producing failures.

    DHave a person review each batch's diff before it merges, applying the same code-review standard the studio uses for changes written by engineers.

    EDisable the failing tests until the migration is complete, then re-enable and fix them in one final pass so the migration is not interrupted.
    Check answerShow the answer
21. Question 21Domain 2: Applications and IntegrationSingle answer

    Cobalt Insurance's claims-intake service is configured with an alias that points at whichever release of its tier is newest. After an overnight release, the JSON the service parses acquired an explanatory sentence before the object and 40 percent of intakes failed parsing until an engineer patched the parser at 6 a.m. The insurer's change-control policy requires every production behaviour change to be tested first. Which configuration practice prevents a repeat?

    Select the best answer.

    AKeep the alias so the service always benefits from model improvements as they ship, and add a parser fallback that strips any leading prose before the JSON object begins.

    BPin the service to a specific dated model version in configuration and adopt new releases only after the evaluation suite has passed against them in staging.

    CKeep the alias but lower the temperature to zero, which makes the output format stable enough that a release change cannot alter the structure of responses.

    DPin the prompt text in version control while leaving the model alias in place, since the prompt, not the model, determines the format of the output.
    Show the answer
22. Question 22Domain 1: Agents and WorkflowsSingle answer

    Brightline Software, a SaaS company with two backend engineers, wants an internal agent that can read its repository, edit files, run the test suite, and ask for approval before any shell command, with a pilot due in three weeks. One engineer prefers writing the agent loop directly on the Messages API for full control; the other proposes the Claude Agent SDK. Given the team size and the deadline, which approach is the better starting point, and why?

    Select the best answer.

    AA custom loop on the Messages API, because the SDK's built-in tools cannot be restricted, so an approval step before shell commands is only achievable in hand-written code.

    BA custom loop on the Messages API, because the SDK is limited to the interactive Claude Code terminal and cannot be embedded in an internal service that the team controls.

    CThe Claude Agent SDK, because it supplies the loop, file and shell tools, permission handling and context management that a custom build would have to write and maintain.

    DThe Claude Agent SDK, because it removes the need for a system prompt and permission configuration, letting the model infer the repository's rules from the code it reads.
    Show the answer
23. Question 23Domain 2: Applications and IntegrationSingle answer

    Riverbend Energy, a utility, generates outage notices with Claude and posts each one to its customer-messaging REST endpoint. The HTTP client times out after 10 seconds and retries; the messaging service usually completed the first request anyway, so 3 percent of customers received duplicate notices last month. The endpoint accepts a client-supplied key on each request and uses it to detect repeats. Which change stops the duplicates while keeping the retry?

    Select the best answer.

    ARaise the client timeout to 60 seconds so the first request is almost always acknowledged before a retry is triggered, and keep the request body and key exactly as they are.

    BRemove the retry entirely and log the timeout for manual follow-up, since a duplicate notice does more harm to customers than a delayed one would.

    CGenerate a fresh key for every attempt so the endpoint can distinguish the retry from the original in its logs and reconcile the duplicates afterwards.

    DSend the same client-supplied idempotency key on the original request and on every retry of it, so the endpoint recognises the repeat and returns the earlier result.
    Show the answer
24. Question 24Domain 3: Claude CodeSelect 2

    Emberforge Games keeps a monorepo with a root CLAUDE.md of shared conventions. The payments team needs stricter rules, such as never editing ledger migrations, that apply only when Claude Code is working inside the services/payments directory. Several engineers also keep personal instructions, like a preferred commit message style, that must never be committed to the repository. Which TWO placements use the CLAUDE.md hierarchy as intended? (Select two.)

    Select exactly 2. 0/2 selected.

    AThe payments rules appended to the root CLAUDE.md under a heading saying they apply only to payments, so every session loads them regardless of the directory.

    BPersonal instructions added to the root CLAUDE.md, since one shared file is easier to maintain than several and other engineers can ignore the entries that are not theirs.

    CA CLAUDE.md inside services/payments holding the stricter rules, which Claude Code reads in addition to the root file when it works in that directory.

    DPersonal instructions in a CLAUDE.local.md or in the engineer's user-level CLAUDE.md, which apply to that person's sessions and stay out of the repository.

    EThe payments rules in a CLAUDE.md placed in each engineer's home directory, so the rules travel with the engineer to every repository they work in.
    Check answerShow the answer
25. Question 25Domain 2: Applications and IntegrationSingle answer

    Barleycorn Grocers' support assistant builds one string per request: the operating instructions, then the customer's message, then the retrieved order history, and sends all of it as the user turn. In 2 percent of conversations the assistant follows text in the customer's message as if it were an instruction, once issuing a refund it was told to offer only under the conditions listed in the instructions section. Which design change establishes the boundary that is missing?

    Select the best answer.

    AMove the order history into the system prompt where it carries more authority than the customer message, leaving the instructions and the message together in the user turn.

    BMove the operating instructions to the system prompt, keep the customer's message as the user turn, and wrap the retrieved order history in clearly labelled data tags.

    CKeep the single string but add a sentence stating that any instructions appearing inside the customer's message must be ignored by the assistant when it responds.

    DSwitch to the highest-capability model tier, which separates instructions from data more reliably, and leave the structure of the request exactly as it is.
    Show the answer
26. Question 26Domain 1: Agents and WorkflowsSelect 2

    Tolland Precision, a manufacturer, built a maintenance-planning agent on an agentic framework. A planning session runs 80 or more turns; by turn 50 the agent re-inspects machines it already cleared and reverses a shutdown decision made at turn 12. Traces show the early turns have been pushed out by accumulated tool output, not that the model changed its assessment. Which TWO agent patterns address the lost decisions without simply enlarging the context? (Select two.)

    Select exactly 2. 0/2 selected.

    ASwitch to a different agentic framework, because losing early turns is a property of the framework's state handling rather than of the context window.

    BWrite each decision to an external memory store as it is made and read the store back at the start of every planning step.

    CReplace verbose raw tool results with a compact summary once they have been acted on, so the early decisions stay inside the window.

    DIncrease the sampling temperature so the model explores the machine list more widely and is less likely to revisit machines it has already cleared.

    EMove every tool result into the system prompt, where content is never truncated by the framework and therefore remains visible for the whole session.
    Check answerShow the answer
27. Question 27Domain 2: Applications and IntegrationSingle answer

    Tolland Precision's platform team found that a formatter hook runs in Claude Code for three engineers and not for the other nine; the hook is defined in one engineer's user-level settings file and two others copied it. The team wants the hook applied to everyone working in the repository, without overriding personal preferences such as each engineer's editor. Where should the hook definition live?

    Select the best answer.

    AIn the shared project settings file, .claude/settings.json, committed to version control so every checkout applies it, with personal preferences left in each engineer's own local settings.

    BIn the repository's CLAUDE.md, as an instruction that the formatter must run after each edit, since CLAUDE.md is committed and read at the start of every engineer's session.

    CIn each engineer's .claude/settings.local.json, distributed by an onboarding script so all twelve receive the same hook while the file itself stays out of version control.

    DIn the user-level settings of every engineer, added by asking the nine remaining engineers to copy the definition from the one whose hook already works correctly.
    Show the answer
28. Question 28Domain 2: Applications and IntegrationSelect 2

    Wexford University wants a study assistant for 30,000 students. Stakeholders listed: a warm, encouraging tone; a peak of 400 concurrent conversations during exam weeks; answers that cite the assigned course reading; conversation logs that must stay in the university's region and be deleted after 90 days; and a 300-word cap on replies. The architect is separating infrastructure requirements from functional ones before design begins. Which TWO items on the list are infrastructure requirements? (Select two.)

    Select exactly 2. 0/2 selected.

    AHandling 400 concurrent conversations at peak, which sets the API throughput tier, the concurrency limits and the queueing that the deployment must be sized for.

    BThe warm, encouraging tone, which determines the system prompt wording and the examples used to shape how the assistant phrases its replies.

    CCiting the assigned course reading in every answer, which determines the retrieval design and the format the model must follow when it responds.

    DThe 300-word cap on replies, which determines the max_tokens setting and the output constraint written into the prompt for every response.

    EKeeping logs in the university's region and deleting them after 90 days, which determines where the system is hosted and how its storage is configured.
    Check answerShow the answer
29. Question 29Domain 5: Model Selection and OptimizationSingle answer

    Harrowgate & Voss, a law firm, uses the Claude API to produce clause summaries that go into a court filing bundle. The compliance partner requires that any summary reproduced during an audit be byte-identical to the one filed. The engineering team set temperature to 0 and reports that in a 500-run test, 11 runs still differed by a few words. The partner asks how the requirement can be met without changing the model. What is the most accurate response?

    Select the best answer.

    ASet temperature to 0 and top_p to 0 at the same time so the sampler is left with exactly one candidate token at every step, which the team can document for the auditor as a fully reproducible configuration.

    BSend the same request three times and keep the majority text, since agreement across independent runs demonstrates that the model has converged on the deterministic answer for that clause.

    CStore the filed summary itself as the record of what was produced, alongside the pinned model version and prompt hash, because low temperature narrows sampling but does not promise identical output.

    DTell the model in its system prompt to answer deterministically and to reuse its earlier phrasing whenever it meets a clause it has summarised before.
    Show the answer
30. Question 30Domain 5: Model Selection and OptimizationSingle answer

    Coastline Telecom runs a customer-support assistant on Claude with a median time-to-first-token of 600 milliseconds and a target of under one second. Support leads report that the assistant handles routine plan questions well but reasons poorly on the roughly 4 percent of turns that involve reconciling a disputed bill across several months of statements. A developer proposes enabling extended thinking with a large budget on every request. Given the latency target, what is the most defensible change?

    Select the best answer.

    AEnable extended thinking on every request with a generous budget, because a single consistent reasoning depth across all turns is easier to test and monitor, and the median latency should still land near the target.

    BMove the whole assistant to the largest model tier instead of enabling thinking, since a more capable model reasons better on every turn without needing a separate thinking budget.

    CLeave the configuration as it is and add a system prompt instruction asking the model to think step by step before answering any question that concerns a billing dispute.

    DRoute the bill-reconciliation turns to a call with extended thinking enabled and an appropriate budget, and keep routine turns on the current configuration so their latency is unchanged.
    Show the answer
31. Question 31Domain 5: Model Selection and OptimizationSingle answer

    Fable & Thread, an online retailer, uses Claude to tag incoming product listings with one of 60 categories. Overall accuracy on a labelled set is 97 percent, but the model confuses the categories 'loungewear' and 'sleepwear' in 40 percent of listings that belong to either. The team has already rewritten the instructions twice to define both categories precisely and the confusion rate did not move. The prompt is cached, so added tokens cost little per call. What change is most likely to fix the confusion?

    Select the best answer.

    AAdd three to five labelled examples to the prompt that sit on the boundary between the two categories, showing which features decide each case.

    BRaise the temperature so the model explores more of the category space before committing to a label, which should let it weigh the distinguishing features of each listing more fully.

    CMove the two category definitions from the system prompt into the user message so that they appear immediately next to the listing text the model is classifying.

    DRemove the two confusable categories from the taxonomy and route any listing that mentions either one into a single combined holding category for a merchandising analyst to split by hand.
    Show the answer
32. Question 32Domain 5: Model Selection and OptimizationSelect 2

    Kestrel Freight, a logistics firm, generates delivery-exception notes through the Claude API from 200 depot terminals. During the 6 a.m. dispatch peak, 18 percent of requests return HTTP 429, and the team's hand-written wrapper retries each failure immediately up to ten times. Since that wrapper shipped, the 429 rate at peak has risen to 31 percent. Depots need each note within a few seconds. Which two changes to the client should the team make? (Select two.)

    Select exactly 2. 0/2 selected.

    ARaise the per-request timeout so that the client waits through the rate-limited window instead of receiving the 429 response at all.

    BRetry with exponential backoff and jitter, honour any retry-after header the API returns, and cap the number of attempts per request.

    COpen a fresh HTTP client for each retry so that a throttled connection is never reused and the retried request can be accepted by a different edge.

    DMove the depot terminals to the Message Batches API, since batch submissions avoid per-minute request limits and each note is returned within the 24-hour processing window.

    EPut a bounded concurrency limit or queue in front of the peak so that the depots' combined request rate stays under the account's quota instead of bursting past it.
    Check answerShow the answer
33. Question 33Domain 5: Model Selection and OptimizationSelect 2

    Solventis, a SaaS company, is adding a drafting assistant to its browser application. Product wants tokens to appear on screen as Claude generates them, and a security review has flagged that the prototype calls the Messages API directly from the browser with the company's API key in the JavaScript bundle. Responses average 900 output tokens and take about 12 seconds to complete. Which two changes does the design need? (Select two.)

    Select exactly 2. 0/2 selected.

    AMove the Claude call to a backend service that holds the API key, and have that service consume the Messages API streaming response as it arrives.

    BKeep the browser calling the API directly but obfuscate the key inside the bundle and rotate it weekly, since streaming requires a direct client connection.

    CRelay tokens from the backend to the browser over a websocket or server-sent events connection so the user sees text as it is generated.

    DWait for the full response on the backend and then send it to the browser in one payload, because streaming through an intermediate service is not supported.

    EOpen a websocket from the browser straight to the Messages API endpoint so tokens flow without a backend hop, passing the API key as the first message.
    Check answerShow the answer
34. Question 34Domain 5: Model Selection and OptimizationSingle answer

    Pinecrest Council, a local government agency, generates planning-committee report drafts with Claude using extended thinking and long outputs. Requests are sent through the council's API gateway, which closes any connection that carries no bytes for 60 seconds. About one draft in five fails with a connection reset after roughly a minute, and the officer who requested it is waiting at their screen. The team must keep using the gateway. What change is most likely to resolve the resets?

    Select the best answer.

    ASubmit each draft through the Message Batches API so that no connection stays open during generation, and poll for the result when it is ready.

    BRequest the draft with streaming enabled so that events flow continuously over the connection and the gateway never sees a 60-second silence.

    CRaise the client-side timeout in the SDK configuration to twenty minutes so that the client keeps waiting past the point at which the gateway would otherwise close the connection.

    DReduce max_tokens on each request so the draft finishes within a minute, and ask officers to request the remainder of the report in a second call.
    Show the answer
35. Question 35Domain 5: Model Selection and OptimizationSingle answer

    Brightwater Mutual, an insurer, triages first-notice-of-loss messages into one of eight claim queues. The classifier runs on the highest-capability Claude tier, costs about 3,100 dollars a day, and averages 2.4 seconds per message against a 1-second service target. On a 5,000-message labelled evaluation set, the fastest tier scores 97.8 percent and the highest tier scores 98.6 percent; the mid tier has not been run. Misrouted claims are corrected by a human within an hour at low cost. Which model choice is most defensible?

    Select the best answer.

    AStay on the highest-capability tier because the 0.8-point accuracy difference is measurable, and a regulated claims process should always run on the most capable model the insurer can afford.

    BMove to the mid-tier model as a compromise that splits the accuracy difference between the measured tiers, roughly halves latency and cost, and avoids committing to either extreme for now.

    CStay on the highest tier but enable extended thinking so that the remaining 1.4 percent of misroutes are reasoned through more carefully before a queue is assigned.

    DMove to the fastest tier, keep the evaluation set as a release gate, and monitor misroute corrections, because it meets the latency target at a fraction of the cost with a loss the process already absorbs.
    Show the answer
36. Question 36Domain 5: Model Selection and OptimizationSingle answer

    Stonebridge Accounting, an accounting firm, extracts line items from scanned invoices with Claude and parses the result into its ledger. After a new model version was released, the team switched production to it the same afternoon because the release notes described higher accuracy. Within a day, 6 percent of invoices failed to parse: the new version began returning currency amounts with symbols where the old one returned plain numbers, and the parser rejects them. What should the team have done before switching, and why?

    Select the best answer.

    ARead the release notes and migration guidance more closely and delay the switch until Anthropic documents the output format changes that affect invoice-style extraction, since a vendor's migration notes are the authoritative record of behavioural differences.

    BSwitch models but raise the parser's tolerance to accept any string in the amount field, because a newer model's output should be trusted over a legacy parsing rule written for the old version.

    CRun the existing evaluation set against the new version in a non-production lane first, because behaviour can shift between releases in ways an accuracy headline does not describe.

    DKeep production on the old version indefinitely, because migrating between model releases introduces a category of risk that an accounting workflow handling client invoices cannot justify taking on.
    Show the answer
37. Question 37Domain 5: Model Selection and OptimizationSingle answer

    Redwood State University runs a course assistant that sends the same 38,000-token course handbook with every student question, followed by the question itself. The team enabled prompt caching by adding a cache_control breakpoint, expecting input cost to fall by most of the handbook's share. After a week, usage logs show cache_read_input_tokens at zero on every request and cache_creation_input_tokens on every request, including bursts of questions sent a few seconds apart. Each request currently places the student's question first, then the handbook, then the breakpoint. What is the most likely cause and fix?

    Select the best answer.

    ACaching requires the model to see identical requests in full, so the fix is to normalise every student question into a fixed template before it is sent, making each request match the last.

    BThe variable question sits before the handbook, so the cached prefix never matches; place the handbook first with the breakpoint after it and put the question afterwards.

    CThe handbook is over the maximum cacheable length, so the fix is to split it into two halves and mark each half with its own cache_control breakpoint.

    DCache entries expire after a short period of inactivity, so the fix is to schedule a background keep-alive request every minute that resends the handbook and keeps the entry resident between questions.
    Show the answer
38. Question 38Domain 6: Prompt and Context EngineeringSingle answer

    Vantage Robotics, a manufacturer, runs a Claude agent that diagnoses failed firmware builds. A typical session makes 30 to 50 tool calls, and several of them return build logs of 20,000 tokens or more. Engineers observe that after about the twentieth call the agent starts repeating steps it already completed and forgets constraints stated at the start of the session, even though the model's context window has not yet been exhausted. The diagnosis must continue in the same session because later steps depend on earlier findings. What is the most defensible change?

    Select the best answer.

    AAdopt a model with a bigger context window so that all of the build logs fit and no earlier turn has to be dropped.

    BStart a fresh session after every ten tool calls, carrying over only the original task description and the agent's most recent message, so each stretch of work begins with a nearly empty context window.

    CTruncate or summarise each tool result to the lines that matter before it re-enters context, and compact older turns while preserving the stated constraints and findings.

    DIncrease the thinking budget so the model reasons more carefully about the logs it has already been given before choosing the next tool call.
    Show the answer
39. Question 39Domain 6: Prompt and Context EngineeringSelect 2

    Orbis Publishing, a media company, runs a Claude agent that prepares a weekly editorial brief. Partway through, the agent must search a 400-article archive for prior coverage of each of six topics, and each search returns 3,000 to 8,000 tokens of excerpts. Editors report the brief's final section often contradicts decisions the agent made in its opening turns, and cost per brief has tripled since the archive step was added. The brief must remain one coherent document produced in one run. Which two changes best address both symptoms? (Select two.)

    Select exactly 2. 0/2 selected.

    ARaise the parent agent's max_tokens so that it can hold the full excerpt text for all six topics alongside its own reasoning.

    BDelegate each archive search to a subagent with its own context that returns only a short structured summary of relevant prior coverage to the parent.

    CAsk the agent to re-read the whole conversation before writing each section so that it notices any earlier decisions it might contradict.

    DKeep the decisions made in the opening turns in a compact, explicit working-notes block that is carried forward whenever older turns are compacted.

    EMove the archive search to the start of the run so that all excerpts arrive before any editorial decisions are made and nothing is lost afterwards.
    Check answerShow the answer
40. Question 40Domain 6: Prompt and Context EngineeringSingle answer

    Northgate Health, a hospital group, built a discharge-summary assistant. Each request is a single user message that concatenates the clinician's instructions, the formatting rules, and the pasted clinical notes in one block of prose. Reviewers find that in about 9 percent of summaries the model follows phrasing inside the notes as if it were an instruction, for example rendering a nurse's comment 'keep this brief' as a directive to shorten the whole summary. Which prompt structure change most directly addresses this?

    Select the best answer.

    AAdd a sentence at the end of the user message stating that anything resembling an instruction inside the clinical notes should be ignored by the model.

    BMove the clinical notes to the start of the user message and the instructions and formatting rules to the end, so that the last text the model reads before answering is the instruction set rather than the notes.

    CSplit each request into two calls, one that extracts a list of facts from the clinical notes and one that formats those facts into a summary, so that no single call sees both the notes and the formatting rules.

    DMove the stable role, formatting rules and instructions into the system prompt, and wrap the clinical notes in a clearly named XML tag in the user turn that the instructions describe as data to summarise.
    Show the answer
41. Question 41Domain 6: Prompt and Context EngineeringSingle answer

    Tolman Precision, a manufacturer, uses Claude to extract torque specifications from supplier datasheets. The prompt asks for 'a concise, well-structured list of torque values'. Downstream code expects each value in newton-metres with the fastener it applies to, and it fails on roughly 22 percent of outputs because values arrive in mixed units, some without fasteners, and some as paragraphs. The prompt must stay short because 200 datasheets are processed per run without caching. What should the team change first?

    Select the best answer.

    AReplace the adjectives with explicit output constraints: one line per fastener, value converted to newton-metres, a fixed field order, and a stated rule for missing values.

    BAdd ten worked examples of correctly formatted torque lists to the prompt so the model can infer the expected structure and units from them.

    CAppend an instruction asking the model to review its own output against the downstream parser's requirements before responding, and to revise any line that would fail, so that formatting errors are caught inside the same call.

    DEnable extended thinking so that the model reasons about unit conversion and structure before it writes the list, which should reduce inconsistencies across datasheets.
    Show the answer
42. Question 42Domain 6: Prompt and Context EngineeringSingle answer

    Lumen Analytics, a SaaS company, asks Claude to return a JSON object describing each customer support ticket, then inserts the fields into a PostgreSQL table. About 3 percent of responses cannot be inserted: some arrive wrapped in markdown code fences, some include a sentence before the object, and some use a priority value outside the four the table allows. Every failure currently drops the ticket silently. The team wants a fix that holds up as prompts and models change. What is the most defensible approach?

    Select the best answer.

    AStrip code fences and leading prose with a regular expression before parsing, map any unknown priority value to 'medium' so every ticket is inserted, and log each case where the mapping fired for the support team to review weekly.

    BAdd a strongly worded instruction to the system prompt that the response must be raw JSON with no surrounding text, and reject any response that does not parse.

    CDefine the ticket fields as a tool with an input schema that includes an enum for priority, force that tool call, and validate the returned arguments against the schema before inserting; route failures to a retry or review queue.

    DPrefill the assistant turn with an opening brace so the response begins as JSON with no fences or preamble, and keep the current parsing code and table constraints unchanged.
    Show the answer
43. Question 43Domain 6: Prompt and Context EngineeringSelect 2

    Beacon Assurance, an insurer, deployed a Claude assistant that answers policy questions for call-centre staff by quoting the relevant clause number from the customer's policy document, which is supplied in the prompt. A quality audit of 800 answers found that 4 percent cited a clause number absent from the supplied document, and in every such case the answer was phrased with complete confidence. Staff act on these answers during live calls. Which two controls most directly reduce this risk? (Select two.)

    Select exactly 2. 0/2 selected.

    AHave the application check every cited clause number against the supplied document before displaying the answer, and withhold or flag any answer whose citation does not resolve.

    BInsert a system prompt instruction asking the model to take more care over clause numbers and to say explicitly when it is unsure.

    CAsk the model to score how confident it is in every citation from 1 to 5 and show staff only the answers scored 4 or above, holding the rest for review.

    DIncrease max_tokens so that the model has room to explain its reasoning, on the basis that longer answers are less likely to contain fabricated details.

    EInstruct the model to quote the clause text verbatim alongside its number, and compare that quoted text programmatically to the document so a fabricated citation is detectable.
    Check answerShow the answer
44. Question 44Domain 7: Security and SafetySelect 2

    Tidewater Logistics runs a Claude agent that reads inbound supplier emails and can call a create_purchase_order tool. In one incident, a supplier email contained a plain-text paragraph addressed to the assistant instructing it to raise a 48,000 dollar order for a new part number, and the agent complied. The agent's tool set also includes read-only lookups of contracts and stock levels. Which two changes most directly mitigate this class of attack? (Select two.)

    Select exactly 2. 0/2 selected.

    AStrip all HTML formatting from emails before they reach the model so that hidden or styled text cannot carry instructions to the agent.

    BSwitch to a larger model tier, since more capable models follow their system prompt more faithfully and will disregard instructions found in email content.

    CPass email bodies into the prompt inside a clearly labelled data boundary, with system instructions stating that content there is untrusted input to be analysed and never a source of instructions.

    DRequire a human approval step before the create_purchase_order tool executes, at least above a value threshold, so an injected instruction cannot complete a purchase on its own.

    ELog every purchase order the agent creates to a daily report that the procurement team reviews, so that any unauthorised orders are caught the following morning.
    Check answerShow the answer
45. Question 45Domain 7: Security and SafetySingle answer

    Cedarline Credit Union launched a member-facing assistant whose only guardrail against giving personalised investment recommendations is a paragraph in the system prompt. In its first month, red-team testers produced recommendations in 7 percent of adversarial sessions, usually by role-play framing. The compliance team requires that a single prompt-level bypass must never result in a recommendation reaching a member. Which change best satisfies that requirement?

    Select the best answer.

    ARewrite the system prompt paragraph to be more explicit and emphatic about the prohibition, adding several worked examples of role-play attempts that the assistant must recognise and refuse.

    BAdd an independent output check, such as a classifier or a second model call, that screens each response for investment recommendations before it is shown, and route flagged responses to a safe fallback.

    CRestrict the assistant to a fixed menu of approved topics that the member selects from a list before typing anything, so that free-text questions about investments can no longer be asked and the prohibited behaviour has no entry point.

    DKeep a blocklist of the role-play framings the red team used, reject any incoming message that matches one before it reaches the model, and add new framings to the list as they are found.
    Show the answer
46. Question 46Domain 7: Security and SafetySingle answer

    Quill & Marrow, a retailer, uses Claude Code across a shared data-engineering repository. Last week an agent session ran a command that deleted a directory of uncommitted migration scripts while cleaning up a build. The CLAUDE.md file already contains the line 'never delete files outside the build directory', and the session transcript shows the model read it. The platform team wants an automatic control that holds regardless of how the model interprets its instructions and does not slow down the hundreds of routine commands a session runs. What should they implement?

    Select the best answer.

    AMove the deletion rule to the top of CLAUDE.md, restate it in stronger language, and add a reminder to every custom slash command that touches the file system.

    BSet the repository's settings.json so that every Bash command requires an engineer's approval before it runs, so a human sees each command and rejects any deletion that reaches outside the build directory.

    CA PostToolUse hook that records every Bash command to an audit log so the team can identify which session caused a deletion and restore the scripts from backup.

    DA PreToolUse hook that inspects Bash tool calls before they run and blocks any deletion command whose target path falls outside the build directory, returning a denial the model sees.
    Show the answer
47. Question 47Domain 7: Security and SafetySingle answer

    Halcyon Airways discovers that a production Claude API key was committed to a public repository three days ago inside a configuration file. Usage dashboards show a spike in requests from unfamiliar IP ranges over the past 24 hours. The key is used by two services and a nightly batch job, all deployed from the same environment. What should the team do first?

    Select the best answer.

    ARevoke the exposed key in the Claude Console immediately, issue replacement keys scoped per service, and update the deployments, accepting a brief interruption while the services roll over.

    BRewrite the repository history to remove the file containing the key, then force-push so that the key is no longer visible to anyone browsing the repository.

    CSet a spend limit on the workspace so that the unfamiliar traffic cannot run up an unbounded bill, then investigate the source and pattern of the requests over several days before making any changes to running services.

    DAdd the configuration file to .gitignore and rotate the key at the next scheduled maintenance window so that the two services and the batch job are not disrupted mid-week.
    Show the answer
48. Question 48Domain 8: Tools and MCPsSingle answer

    Greywell Energy, a utility, runs a Claude agent that answers field-technician questions using a get_meter_reading tool backed by an internal service. When that service times out, the agent's harness raises an exception, the agent loop terminates, and the technician sees a generic failure. Timeouts affect about 2 percent of calls and usually succeed on a retry a few seconds later. The team wants the agent to recover on its own where possible and, when a reading truly cannot be obtained, to tell the technician so plainly rather than failing. How should the harness handle a tool failure?

    Select the best answer.

    ACatch the exception in the harness and return an empty string as the tool result so that the conversation continues without interruption and the model answers from what it already knows.

    BCatch the exception in the harness, retry the meter service a few times with a short delay, and if every attempt fails re-raise it so the loop terminates as it does today, keeping tool failures out of the model's view.

    CHand the failure to the model as an is_error tool_result block with a short description of what went wrong, so the model can retry, try another approach, or explain the problem to the technician.

    DLet the exception terminate the loop as it does today, but improve the generic failure message shown to the technician so that they know to ask their question again.
    Show the answer
49. Question 49Domain 8: Tools and MCPsSingle answer

    Summit Ridge Clinics, a hospital group, exposes a lookup_appointment tool to a scheduling assistant. The tool's input schema has a single parameter named 'query' typed as a free string, and its description reads 'Look up an appointment'. The backing service requires a patient reference number in the form SRC followed by eight digits. In 31 percent of calls the model passes a patient's name or a date instead, and the service returns nothing. What is the most defensible fix?

    Select the best answer.

    AAdd a paragraph to the system prompt listing the reference-number format and instructing the model to always look up the number before calling the tool.

    BHave the backing service accept names and dates as well as reference numbers, resolving them on the server through a fuzzy match, so that whatever the model passes returns an appointment record.

    CReplace the tool with a broader search_patient_records tool that takes any string and returns all matching records, leaving the model to pick the right appointment.

    DRename the parameter to patient_reference, describe the exact format and give an example, constrain it with a pattern in the schema, and state in the tool description what the tool returns and when to use it.
    Show the answer
50. Question 50Domain 8: Tools and MCPsSelect 2

    Oakhaven College built a registrar assistant with 42 tools, several of which overlap: three different tools return course timetables with slightly different filters, and four return student records at different levels of detail. Tool descriptions are one short phrase each. Evaluation shows the assistant calls a suitable tool on the first attempt 68 percent of the time and often calls two overlapping tools in sequence for one question, doubling latency. Which two changes would most improve tool selection? (Select two.)

    Select exactly 2. 0/2 selected.

    AConsolidate the overlapping tools into one tool per capability with parameters for the filters and detail level, so the model chooses among fewer, clearly distinct options.

    BRewrite every tool description so it names the return value, the situations that call for this tool, and the cases where a named alternative should be chosen instead.

    CAdd a system-prompt rule that limits the assistant to one tool call per question and, when several tools seem to fit, tells it to prefer whichever is listed first.

    DMove to the highest-capability model tier before changing any tool definitions, since a stronger model should be able to disambiguate the existing 42 tools.

    EPrepend the full JSON schema of all 42 tools to the system prompt so that the model has a second view of the tool set when it selects.
    Check answerShow the answer
51. Question 51Domain 8: Tools and MCPsSingle answer

    Larkspur Pharma wants 80 researchers using Claude Desktop to query a regulated document repository through an MCP server. Security insists on three things: each query is authenticated through the company identity provider, access can be revoked centrally, and the repository credentials never leave the data centre. A developer has a working prototype that runs the server as a local process on their laptop over stdio. Which deployment approach fits the requirements?

    Select the best answer.

    ADistribute the stdio server to each researcher's laptop with the repository credentials in a local configuration file managed by the desktop fleet tooling, since stdio is the simplest and most widely supported MCP transport for desktop clients.

    BDeploy the server centrally as a remote MCP server over a streamable HTTP transport with authentication against the identity provider, so that credentials stay on the server side and access is managed in one place.

    CKeep the stdio server on each laptop but have it read the repository credentials from an environment variable that IT sets during device provisioning.

    DExpose the repository's own REST API to Claude Desktop directly and skip MCP, because a protocol layer adds latency without changing where authentication happens.
    Show the answer
52. Question 52Domain 8: Tools and MCPsSingle answer

    Meridian Rail, a public transit agency, wants Claude Code to produce incident post-mortems in the agency's fixed format: a nine-section template, a severity rubric, a script that pulls timeline entries from an exported CSV, and a checklist that must be completed before the document is filed. All the inputs are already on the engineer's machine, and no external system needs to be called. Which approach best fits this need?

    Select the best answer.

    AAuthor a Skill that packages the template, rubric, checklist and the CSV script with instructions on when to apply them, so Claude Code loads the procedure whenever a post-mortem is requested.

    BStand up an MCP server offering a generate_post_mortem tool that wraps the template, rubric and CSV script, deploy it on shared infrastructure, and connect Claude Code to it so the whole procedure runs as a single tool call.

    CAdd the full template, rubric and checklist to the repository's CLAUDE.md so that every session in the repository carries the post-mortem procedure in context.

    DDefine a custom tool in an agent built on the Claude Agent SDK that returns the template and rubric text when called, and instruct the agent to use it for post-mortems.
    Show the answer
53. Question 53Domain 8: Tools and MCPsSelect 2

    Ardent Ledger Bank's analysts want an assistant that, given an uploaded CSV of aggregated transaction figures, computes summary statistics and produces a chart, and separately fetches the latest published base rate from the central bank's public website. The CSVs contain no customer data and the security team has approved sending them to the API. The team has two weeks and no platform engineers available to run infrastructure. Which two approaches fit these constraints? (Select two.)

    Select exactly 2. 0/2 selected.

    ABuild an MCP server that wraps a Python interpreter on the bank's own servers so the analysis runs on infrastructure the bank controls, and deploy it within the two weeks.

    BUse the built-in code execution tool for the CSV analysis, since it runs the code in Anthropic's managed sandbox and needs no infrastructure from the bank.

    CAsk the model to compute the statistics directly from the CSV text placed in the prompt and to state the current base rate from its training data.

    DUse the built-in web fetch or web search tool to retrieve the published base rate, since the source is a public web page and no proprietary system is involved.

    EWrite a custom tool backed by a scheduled job that scrapes the central bank page nightly and stores the rate in a database the assistant queries.
    Check answerShow the answer