Claude Certified Associate Practice Test

Answer: C
Domain 2 includes editing, adapting, and comparing outputs for the intended audience. Engineers need mechanism and terminology; executives need the decision, its risk, and the next action, so the yardsticks differ, which is option C. Option A rewards length, and the executive draft should usually be the shorter one. Option D outsources judgment to the model that wrote the drafts and scores prose, not fitness for a reader. Option B guarantees a wrong result for one audience, because a single rubric cannot value brevity and depth at the same time.

Answer: B
Language models can produce specific-looking details, including subsection numbers, that are wrong; that is the hallucination pattern named in Domain 2. Option B checks both the citation and the summary against the authoritative text before a compliance audience relies on it. Option A treats a tool setting as verification; web search can surface a source, but it does not confirm that this summary reads it correctly. Option C relies on self-reported confidence, which is not an accuracy signal. Option D improves presentation while leaving correctness untested.

Answer: A
Domain 2 separates accuracy from completeness. Nothing in the update is false, so the defect is that information the reader needs for the decision, the budget risk and the deadline, is missing; that is option A. Option C would require fabricated content, and none is present. Option D would require a false statement, and the update is factually accurate. Option B concerns detail level and vocabulary for a particular reader; the update's problem is omitted facts, which would be missing for any audience.

Answer: D
Validation tests provenance, freshness, and disagreement. Option D does all three: each material claim is traced to where it came from, its date is checked, and conflicts are surfaced rather than smoothed over. Option B tests format, not truth; a tidy table can be uniformly wrong. Option C confuses repetition with confirmation; vendor marketing is copied across sites, and two pages restating one press release are still one source. Option A removes the very trail a reviewer needs, so nothing in the table could be checked afterward.

Answer: A
Comparing two outputs meaningfully starts with defined criteria applied to both, which is option A: accuracy against the product documentation, whether every step is covered, and whether it reads clearly. Option B uses length as a proxy for quality, which can favor an article that is shorter because it skipped a step. Option C collects opinions with no shared standard, so agents could be reacting to different things. Option D uses a search-ranking signal that reflects indexing and past traffic, not whether the article's content is correct or complete.

Answer: C
Sorting into one of a fixed set of categories is a categorical, single-correct-answer task, and the grading approach named for exactly that is exact match; that is option C. Option A applies a subjective quality scale to a task that has one right label, not a spectrum of quality. Option B measures word overlap, which is irrelevant to whether the assigned category is the correct one. Option D checks similarity between unrelated tickets over time, which says nothing about whether any single ticket was classified correctly.

Answer: B
Evaluation guidance is explicit that measurable success criteria and a diverse test set, including edge cases, come before a workflow ships, not after; that is option B. Option A checks one case and calls it representative, which a report type never seen in that one sample would fail silently. Option C skips evaluation entirely based on first impression, the failure mode the test-set practice exists to prevent. Option D has the model that produced the summaries also judge them, with no independent cases to catch a systematic blind spot.

Answer: D
The verification effort a draft needs scales with what in it is novel or consequential, not with how the language reads; that is option D, which targets exactly the citations, statutes, and case-specific facts that could be wrong. Option A judges by writing style, which has no relationship to factual correctness. Option B assumes reused language means the new, case-specific content added around it is also safe, which does not follow. Option C checks two sections by position in the document rather than checking the sections that actually contain unverified, case-specific claims.

Answer: C
Materials, dimensions, and price are precise, categorical fields, which is exactly what exact-match checking suits, and sampling first with expansion on failure keeps the check scalable; that is option C. Option A and option B both use a surface trait, photo count or word count, that has no logical connection to whether the stated facts match the spec sheet. Option D has the model that produced the descriptions also judge whether they drifted from the source, which is not an independent check.

Answer: B
The requirement is a single yes/no compliance check, and the grading approach named for exactly that is a binary pass/fail rubric; that is option B. Option A scores a quality spectrum, tone, for a task that only needs a presence check. Option C measures general text overlap with an old template, which would flag emails as different even when the required statement is present and correctly worded. Option D would fail nearly every email, since acknowledgment emails legitimately vary in wording while still containing the required statement.

Answer: B
A specific number attributed to an earnings call needs to be checked against the actual transcript or release before an internal briefing relies on it; that is option B. Option A treats professional-sounding prose and a named source as proof, neither of which confirms the figure is right. Option C changes the attribution to something vaguer without resolving whether the number itself is correct. Option D substitutes the model's self-reported confidence for an independent check, and a stated high confidence is not evidence the figure is accurate.

Answer: B
Evaluation guidance calls for defining measurable success criteria and testing against them before something ships, rather than approving on a read-through; that is option B, applied to both versions before either is used. Option A substitutes one person's impression for a defined check. Option C uses a vague, subjective feeling and a prior year's tone as the bar, neither of which is a measurable criterion. Option D treats internal distribution as low-risk enough to skip evaluation, but a board memo with inaccurate or incomplete figures carries its own real consequences.

Answer: A
Nothing in the summary is false and no required section is missing, so the defect is that the register and vocabulary do not match the intended reader, which is exactly what an audience-fit finding names; that is option A. Option B would require a false statement, and the summary is factually correct. Option C would require a missing required element, and every section is present. Option D would require fabricated content not grounded in any source, and the terminology, while inaccessible, is not invented.

Answer: A
Tone and empathy are subjective quality dimensions, and the evaluation guidance names model-graded rubrics on a Likert scale as the fit for exactly this kind of criterion; that is option A. Option B is for categorical tasks with one correct answer, and a reply's wording legitimately varies while still being empathetic. Option C measures text overlap with a reference document, which says nothing about tone. Option D measures how similar two transcripts are to each other, not whether either one is empathetic or on-brand.

Answer: D
A number that names its source is not the same as a number confirmed against that source; the memo could misread a line item or apply the wrong formula. Option D is the only action that independently reproduces the figure from the underlying financial statement. Option A treats a source citation as proof rather than a claim to be checked. Option B asks the same model to re-verify its own work, which does not catch an error in how it read the P&L in the first place. Option C avoids verifying the exact figure by blurring it, but a board deck built on an unverified number carries the same risk whether it is precise or rounded.

Answer: D
Domain 4 starts with requirements and use-case analysis: map the process before changing it. Option D names what has to be known to decide where Claude assists and where a person stays accountable. Option A designs the output before the process; wording is downstream of decision points and owners. Option B is a configuration choice that presupposes the steps are already defined; picking Manual, Auto, or Skip approval before the map exists is premature. Option C schedules and tools a launch for a process nobody has yet examined.

Answer: D
Domain 4 places a control where an error would do harm. Misassigned owners create missed work, so option D inserts a focused check at the distribution step while keeping the drafting gain. Option A accepts a known error rate on the field that matters most. Option B removes the useful content instead of correcting it; a summary without owners does not drive action. Option C asks for a promise the model cannot keep; an instruction changes phrasing, not the model's ability to know who agreed to what in the room.

Answer: A
Domain 4 measures workflow value end to end. A faster draft that needs more fixing may cost more overall, so option A looks at total cycle time, classifies the errors, and fixes the step that produces them before more teams inherit it. Option B scales a net loss because one local metric improved. Option C moves the cost off the books; correction time is part of the same workflow. Option D stops learning at the moment the data becomes useful, so the weak step would never be identified.

Answer: C
Domain 4 asks associates to communicate Claude's value and its limits. Option C names the task it speeds up, the benefit, and the controls that stay with people, which a stakeholder can hold the team to. Option B overstates: Research citations make claims easier to check; they do not make them true. Option D inverts the dependency; a Project applies a brief, it does not discover one, and vague requirements produce vague synthesis. Option A promises replacement of the reviewers who catch errors, which misrepresents both the tool and responsible use.

Answer: D
Domain 4 asks where to place a control that must never be skipped. A hook runs a shell command before or after a defined action, so the logging step executes outside the model's discretion every time; that is option D. Option A is an instruction, and an instruction is a request the model can still miss under load, which is exactly the failure the lead wants to close. Option B adds reasoning depth, not a guaranteed action; thinking harder about whether to log is not the same as logging. Option C keeps the default auto trigger, which lets Claude decide per ticket whether the tool is worth calling, the opposite of a non-negotiable step.

Answer: B
Domain 4 covers matching a tool's triggering behavior to how often it should actually fire. The default tool_choice is auto, letting Claude decide per turn whether a described tool's capability matches the request, and that boundary is promptable with light system-prompt language, which is option B. Option A forces a ticket on every message, including ones the bot already resolves, creating tickets nobody needs. Option D adds a manual trigger customers do not know to use, so unresolved issues can pass through the conversation with no ticket at all. Option C removes the tool and reintroduces the manual step the workflow was built to automate, pushing the work back onto agents reading full transcripts.

Answer: C
Domain 4 covers integrating Claude into workflows that must run on a fixed cadence without a person present to trigger them. A routine is a scheduled, recurring Claude Code run in the cloud that keeps executing even when the user's machine and the analyst are both offline, which is option C. Option A depends on the analyst's memory and availability, exactly the failure vacation weeks expose. Option B asks Claude to remember and self-initiate a task, but the model does not wake itself on a calendar; something still has to trigger it. Option D is a working session a person opens manually, so it inherits the same single point of failure as option A.

Answer: A
Domain 4 covers recognizing a workflow shaped for a specific cost lever. A stable long system prompt reused across many requests with only the final turn changing is exactly prompt caching's best use case: a cache hit reads the repeated prefix at a fraction of the base input price instead of reprocessing it every time, which is option A. Option D removes the manual the classification depends on, trading a cost fix for a correctness loss. Option B trims the part of the prompt that already changes every request and leaves the large stable block untouched, missing the actual savings. Option C changes models without the evaluation step that would show whether the cheaper model still meets the classification bar, risking accuracy for an unconfirmed savings.

Answer: D
Domain 4 covers designing a workflow's cost and latency profile, not only its capability. Tool use adds token overhead from the tools array itself, the tool_use and tool_result blocks, and a small hidden system prompt for tool-enabled requests, a real cost and latency lever at high volume, so option D scopes the tool set to what the workflow needs. Option B is factually wrong; tool definitions are not free, they consume tokens on every request whether or not they get called. Option A maximizes exposed capability while ignoring that every additional tool raises the per-request cost across thousands of daily calls. Option C ties the decision to staffing rather than to the request-level overhead that actually drives cost and latency at this volume.

Answer: B
Financial claim-adjustment recommendations are a high-risk use case under Anthropic's Usage Policy, which requires two safeguards before such content or a decision is used: a qualified professional reviews it before it is finalized, and the affected person is told AI helped produce it, at the start of the interaction. Option B supplies both. Option D reviews only a monthly sample after decisions have already reached customers, review after the fact rather than before finalization. Option A conditions disclosure on the reviewer agreeing with the number, but the policy requires disclosure regardless of whether the recommendation turns out to be right. Option C makes the calculation visible in an artifact, a presentation choice that satisfies neither the review requirement nor the disclosure requirement on its own.

Answer: D
Domain 4 places a control where an error would cause harm, not evenly across every step. Payout approval is a finance decision, one of the Usage Policy's High-Risk Use Cases requiring a qualified professional to review it before it is finalized, so option D puts the required gate there and lets the low-risk draft move faster. Option C spends review effort on the draft explanation at the same rate as the payout, slowing the safe step without reducing the real risk. Option A puts the gate on the wrong step; a wording issue in a draft is cheap to fix, while an unreviewed incorrect payout has already left the business. Option B removes the review after a clean streak, but a clean history does not bound the size of the next error.

Answer: A
Domain 4 covers connecting Claude Code to the external systems a workflow actually depends on. The Model Context Protocol lets Claude Code connect to servers such as a ticketing system, so it can read and write Jira and a deployment board directly as part of the pipeline, which is option A. Option B keeps the manual copying the engineer wants to remove, just moved into the chat window. Option C has Claude infer status instead of reading the source system, producing a guess where a live lookup exists. Option D uses a hook for the wrong purpose; a hook enforces a fixed script action like reformatting text, it does not give Claude read or write access to an external ticketing or tracking system.

Answer: B
The Exam Guide's Workflow Integration and Solution Design domain covers mapping a full process before automating it: inputs, decision points, risks, owners, handoffs, and success measures. Having already listed inputs, decision points, and risks, the map is incomplete without naming who owns each step, who the output hands off to, and what measure will show the pilot worked, which is option B. Option A designs prompt wording, a downstream implementation detail that depends on the map, not a mapping step itself. Option D is a communications task unrelated to defining the process. Option C looks for a person to blame rather than an owner accountable for a step going forward, a different question from the one the mapping step asks.

Answer: C
Domain 4 places a control at the step where an error would cause harm, sized to the actual risk. The purchase order is submitted through a client tool the team's own application executes, and that call is exactly where a review gate can be inserted; a mis-sized order above the threshold has already caused a warehousing problem, so option C puts the required review there while letting routine small orders move faster. Option B reviews the low-stakes, high-volume orders and leaves the costly ones unchecked, inverting the risk. Option A reviews everything at the same rate, spending scarce reviewer time on orders that rarely need a change. Option D treats a successful system receipt as proof the quantity and price were correct, but a confirmation only means the order arrived, not that its size was right.

Answer: C
Domain 4 covers producing output another system can consume reliably, not just output a person finds readable. Structured Outputs constrains Claude's response to conform to a given JSON schema, the modern, reliable way to guarantee the exact field structure an import job needs, making option C the fit. Option B relies on prefilling a partial response to bias output toward JSON, a workaround no longer supported starting with Claude 4.6-generation models, which return an error on such requests. Option D asks nicely for consistency in free text, which still leaves every field name and type to chance. Option A produces something readable by a person but reintroduces a manual copy step into a system that is supposed to import automatically.

Answer: A
Domain 6 requires following organizational AI policies and escalating uncertainty. Option A stops the non-compliant use and routes the question to whoever can approve or amend the policy. Option D substitutes personal judgment about usefulness for the governance process. Option B treats the model as the policy authority; Claude can summarize a policy, but it cannot grant an exception. Option C moves company work to a consumer account where organizational retention controls, admin oversight, and commercial terms no longer apply, which increases the risk the policy exists to manage.

Answer: A
Domain 6 applies data-sensitivity and privacy controls to the data itself. Option A removes or masks the regulated identifiers before they leave the controlled system, which is what the policy requires. Option B cites a true training exclusion, but the policy restricts sharing the data at all, and no-training is not the same as permitted disclosure. Option C misuses incognito mode; it keeps a chat out of history and memory, it does not make an upload of regulated data compliant. Option D obscures labels while the names and numbers remain in the file.

Answer: D
Employment decisions are a high-risk use under Anthropic's Usage Policy: a qualified person must review the output before it is acted on, and affected people must be told AI was involved. Historical hiring data can encode past bias, so option D examines data, criteria, and legal exposure and keeps a human in the loop. Option C treats biased outcomes as ground truth. Option A hides the method, which defeats the disclosure and review the policy requires. Option B makes the model the sole decision-maker, exactly what the human-in-the-loop requirement prohibits.

Answer: D
Healthcare content is a high-risk use case under Anthropic's Usage Policy, which requires a qualified professional to review such material before it is disseminated. Option D grounds the draft in approved sources and adds that clinical review. Option A treats fluent wording and the presence of citations as accuracy; neither is. Option C shifts verification to the public after the risk has already been created. Option B removes the trail a reviewer would use to check claims, making the review harder rather than the handout safer.

Answer: C
Anthropic's Usage Policy requires that consumer-facing chatbots disclose that the user is interacting with AI, and for high-risk categories that disclosure must come at the start of the interaction rather than after the fact; a general-purpose customer-facing widget falls under this same disclosure standard, so option C is correct. Option A confuses data retention with disclosure — whether chats are logged has no bearing on whether visitors are told they are talking to AI. Option B treats a policy page nobody reads mid-conversation as equivalent to telling the visitor at the point of interaction, which the policy does not accept. Option D discloses only on request, leaving every visitor who does not ask believing they may be speaking with a person, which is exactly the outcome the disclosure requirement exists to prevent.

Answer: C
The Universal Usage Standards prohibit harassment, and a set of message variants built specifically to be sent repeatedly at one individual until they comply with a demand is a harassment campaign regardless of the topic, which makes option C correct. Option B focuses on the tone of each individual message, but a coordinated, repeated campaign aimed at one person is harassment by its pattern and intent, not only by whether any single message is rude. Option A narrows harassment to protected-characteristic-based conduct, but the prohibition covers targeted harassment generally, not only bias-motivated harassment. Option D treats using the sender's real account as a mitigating factor, but attribution of the sender does not change that the content is built for a repeated, targeted campaign against the reviewer.

Answer: D
The Universal Usage Standards' fraud provisions specifically cover plagiarizing or submitting AI-assisted work as one's own without required disclosure or permission, which is exactly what happens when a largely AI-drafted article is presented to a client as entirely original human work — option D. Option A treats the writer's review of the final text as sufficient cover, but review does not create the disclosure the policy requires when a client expects original human authorship. Option C narrows the fraud category to counterfeiting and phishing only, but the same provision explicitly extends to undisclosed AI-assisted submissions. Option B focuses on factual accuracy, which is a separate concern from the misrepresentation of authorship the policy addresses here — an accurate article can still be submitted fraudulently as to who wrote it.

Answer: C
Housing is one of the Usage Policy's High-Risk Use Cases: a qualified person must review the AI-generated recommendation before it is finalized, and the applicant must be told AI was involved, at the point the decision affects them, which is option C. Option A sends the AI output straight to the applicant with no review step at all. Option D reviews denials but lets approvals — which still reflect an AI-generated recommendation — go out with no review whatsoever, and even the denial review happens after a week's delay rather than before that specific denial is sent. Option B discloses AI's role in advance paperwork signed before the background check runs, not connected to the actual decision the applicant receives, which does not satisfy disclosure tied to that decision.

Answer: C
Organizational AI-policy compliance means identified gaps get closed and tracked, not merely observed, so requiring the three non-compliant widgets to add the disclosure and confirming the fix is the correct response — option C. Option D treats the disclosure requirement as solely Anthropic's to enforce, but the company deploying the chatbot is the one obligated to disclose AI to its own customers, and internal compliance exists to catch exactly this kind of gap. Option A defers any correction for a full quarter and relies on self-certification instead of verifying the fix, leaving customers undisclosed to in the meantime. Option B is disproportionate: it disables compliant features along with the non-compliant ones, which fixes nothing about the three widgets that actually lack the label and disrupts unrelated, policy-compliant work.

Answer: D
Legal applications are a High-Risk Use Case under Anthropic's Usage Policy, which requires both a qualified professional's review before the content is finalized or disseminated, and disclosure to the people affected that AI helped produce it, at the outset rather than later — option D. Option A lets the model's own citation check stand in for the required review; Claude confirming its own citations is not a qualified professional's review, it is more model output. Option C substitutes a paralegal's proofread of tone for the attorney review the policy specifically requires for legal content. Option B makes disclosure conditional on the recipient asking, but the policy requires proactive disclosure to affected people, not disclosure on demand.

Answer: D
The Universal Usage Standards prohibit weapons development, explicitly including biological, chemical, radiological, and nuclear weapons, and this prohibition is not lifted by a stated downstream purpose like a training deck, so the correct response is to decline, option D. Option B accepts the stated framing at face value and treats intent as a safe harbor, but the policy prohibits producing this category of content regardless of the requester's claimed purpose. Option A assumes trimming out quantities makes step-by-step synthesis guidance for a mass-casualty agent acceptable, but the prohibited content is the synthesis pathway itself, not only the precise measurements. Option C adds a disclaimer on top of the same prohibited instructions, and a disclaimer does not change that the underlying content is prohibited weapons-development material.

Answer: B
The Universal Usage Standards prohibit election interference, specifically naming deceptive political content and large-scale automated contact with voters or officials, and this request does both at once — impersonating election officials and mass-messaging voters about their polling location — so the correct answer is B. Option A treats accurate polling information as a cure, but the prohibition covers the deceptive impersonation and mass automated contact themselves, not just factual errors in the content. Option C tries to exempt the request because it targets voters rather than officials, but the policy's election-interference prohibition covers deceptive contact with voters directly, including this kind of impersonation. Option D substitutes a sampled human review for compliance, but a policy violation embedded in the message's premise — impersonating officials at scale — is not fixed by spot-checking a subset before sending.

Answer: B
Academic testing is listed among the Usage Policy's High-Risk Use Cases, so a qualified professional — here, an educator — must review the AI-generated score and placement before it is finalized, and students must be told AI is being used, at the start of the process rather than buried elsewhere, which is option B. Option A only reviews after a parent appeals, which is review after finalization, not before. Option C skips review for a subset of students based on their own score, which is not a review at all for that group. Option D places the disclosure in a handbook distributed months before the exam is even scored, not tied to the actual use of AI on that student's exam, which does not satisfy an upfront, connected disclosure.

Answer: A
The Universal Usage Standards prohibit cyberattacks, including denial-of-service tooling and unauthorized attempts to disrupt a system; a request to knock a specific competitor's site offline is exactly that, so the correct response is to decline, option A. Option D relabels the request as a stress test, but a legitimate stress test targets systems the requester owns or has authorization to test, not a competitor's live site during a launch, which is what makes this an attack. Option B tries to split the request into an allowed half and a withheld half, but supplying the traffic-generation logic still builds the core of the prohibited tool. Option C limits the prohibition to critical-infrastructure targets only, but the Universal Usage Standards' cyberattack prohibition covers unauthorized attacks and disruption generally, not solely critical infrastructure.

Answer: D
Domain 1 asks candidates to adapt the prompting strategy to the task type. Brainstorming calls for divergent output: several options, each tied to the need it serves, with the choice deferred to the human, which is option D. Option B converges immediately and hides the alternatives the associate wanted to compare. Option C is drafting work for a decision that has not been made, so the memo would rest on a theme nobody chose. Option A strips the reasoning that makes options comparable; bare scores say nothing about why a theme fits the audience.

Answer: D
Task decomposition is a Domain 1 objective: a request that asks for theme extraction, cross-role comparison, and an executive brief in one turn gives Claude no checkpoint, so parts get dropped. Option D sets a deliverable per stage that the associate reviews before the next stage starts. Option C repeats the same overloaded request and changes nothing about why it fails. Option A discards the evidence the analysis exists to examine, so any themes would be invented. Option B removes the source material rather than the overload; a brief written without the transcripts cannot be checked against them.

Answer: B
Iterating a prompt means giving diagnostic feedback: name what is missing and the requirement it must meet, then ask for the revision. Option B does that and hands Claude the policy language to satisfy. Option D substitutes self-assessment for review; a model's confidence rating is not evidence that the two points were covered. Option A throws away the working structure and the feedback, and a fresh context has no reason to produce the omitted points. Option C changes register, not substance; the two mandatory points stay absent in a friendlier draft.

Answer: B
Anthropic's prompting guidance opens with 'Be clear and direct' and 'Add context to improve performance': a generic draft is the symptom of a prompt that never said who the email is for, what it must achieve, or what the deliverable looks like. Option B supplies exactly those missing inputs. Option A adds emphasis, not information, and the guidance for current models is to dial back capitalized MUST/CRITICAL language. Option C asks for reasoning depth and volume, neither of which fills a context gap. Option D removes the little direction the prompt had, so the draft drifts further from the campaign.

Answer: B
Examples should be diverse enough to cover the range of real cases without teaching an unintended pattern; a set that only shows angry callers teaches Claude that every reply opens with an apology, so option B adds calmer, neutral examples to correct the imbalance. Option D shrinks the set but keeps only the angry-caller pattern, so the same unintended generalization persists. Option C keeps the skewed examples and adds a prohibition that fights what the examples are actually demonstrating. Option A removes the examples, which are one of the most reliable tone-steering tools, instead of fixing what they demonstrate.

Answer: C
Anthropic's long-context guidance places long documents and inputs near the top of the prompt, above the query and instructions, because queries placed at the end can measurably improve response quality on multi-document and long-document tasks; that is option C. Option D repeats the question rather than repositioning the transcript, so the document still follows the query. Option B adds a summary but keeps the transcript after the question, the ordering the guidance advises against. Option A shortens the question without moving the transcript, so the structural problem the paralegal is seeing remains.

Answer: B
Examples are one of the most reliable ways to steer format, tone, and structure; Anthropic's guidance recommends three to five relevant, diverse examples wrapped in example tags, which is option B. Option A describes the target in the abstract, and abstract description is a weaker steering signal than demonstrated examples. Option D states one adjective, professional, and leaves structure undefined, so postings will still vary by recruiter. Option C asks Claude to imagine a standard without ever showing one, which reintroduces the same inconsistency the agency is trying to remove.

Answer: C
Asking Claude to check its own answer against stated criteria before finishing reliably catches errors, especially on coding and math tasks, so option C directs that verification at the exact edge case the patch is still failing. Option A discards the specific failure case for a fresh approach, with no guarantee it addresses that case either. Option D asks for brevity, which does not confirm whether the edge case now passes. Option B produces an explanation for a different audience, not a check of whether the patch actually handles the reported case.

Answer: B
Since prefilling the last turn is no longer supported on current-generation models, the documented replacement for skipping a preamble is a direct system-prompt instruction, such as asking Claude to respond directly without preamble, which is option B. Option D is another form of the unsupported prefill approach, just phrased as an unfinished sentence, and would fail the same way. Option A asks Claude to generate the unwanted preamble and then remove it itself, an extra step that does not stop the preamble from being produced. Option C changes reasoning depth and overall length, not the presence of an opening pleasantry.

Answer: A
The first draft went wrong because the prompt never named which section mattered, so a scope-adding revision, naming the finished-goods forecast specifically, is the direct fix, which is option A. Option D changes the model but keeps the same underspecified request, so the new model has no more information about which section to summarize than the last one did. Option C expands the request to the whole report instead of narrowing it to the needed section, reintroducing the ambiguity in a different form. Option B repeats the identical vague request and hopes for a different, uncontrolled outcome rather than fixing what caused the wrong section to be picked.

Answer: D
For long-document tasks, asking Claude to first quote the relevant passages, for example into a dedicated quotes section, before doing the analysis grounds the response in the source text and reduces drift; that is option D. Option A reverses the order, answering from memory first and checking only when uncertainty is already suspected, which is exactly the pattern the quoting technique prevents. Option C shortens the answer without adding any grounding step, so a short wrong answer is just as likely as a long one. Option B summarizes the whole ordinance broadly, which does not connect a specific detailed question to the exact clause that answers it.

Answer: B
Anthropic's current guidance is that heavy capitalized emphasis, such as MUST and CRITICAL, can cause behaviors to overtrigger on current models, and normal phrasing that states the actual condition works better; that is option B. Option A adds more of the emphasis that is already causing the overtriggering, which would make the problem worse, not better. Option D removes the escalation criteria entirely, so Claude has no basis for flagging real delays either. Option C asks for an internal state change Claude cannot perform through a prompt instruction; the fix is the wording of the criteria, not a request to disregard training.

Answer: A
Anthropic's prompting guidance recommends giving Claude a role in the system prompt, even a single sentence, to focus tone and behavior for a specific use case; that is option A. Different requesters produce different drafts because nothing anchors the voice, and a role prompt fixes that anchor regardless of who asks. Option D adds reasoning depth, which affects thoroughness, not voice consistency. Option C raises effort, a cost and latency lever, not a tone control. Option B pushes the inconsistency back onto the team instead of fixing the prompt, and still leaves no shared voice for Claude's own drafts.

Answer: A
Anthropic's guidance is to tell Claude what to do rather than what to avoid; stating the desired positive form, flowing prose paragraphs, is a clearer format instruction than a list of prohibitions, which is option A. Option B lists three things not to do without saying what the output should look like instead, which is the weaker framing the guidance warns against. Option D changes reasoning depth, not formatting, and list structure can persist regardless of the thinking setting. Option C accepts the wrong format first and asks for a manual conversion afterward, adding a step instead of preventing the mismatch.

Answer: B
Domain 5 covers managing connectors and knowledge sources: access, authority, and retrieval quality all need checking before a workflow depends on them; that is option B. Option A confuses reachability with approval; inheriting the owner's permissions means the connector can see drafts and superseded copies too. Option C widens access and removes the approval step at the moment the folder structure is known to be ambiguous. Option D uses the private-project setting to hide provenance from the people who must review outputs, the opposite of good governance.

Answer: C
Configuration quality depends on maintained knowledge. Project knowledge is read by every chat in the project, so the fix is at the source: option C. Option D instructs Claude to treat superseded policy as current, which turns a stale answer into a deliberate one. Option A keeps both versions and hopes retrieval prefers the newer; RAG ranks passages by relevance to the question, not by which document is current, so either version can be returned. Option B clears memory, which is not where the handbook lives; the outdated file would still be retrieved on the next question.

Answer: D
Domain 5 asks associates to inform, maintain, and update configurations. Instructions, knowledge, and permissions drift out of date as policies, sources, users, and goals change, so option D combines a calendar with event triggers. Option C waits for harm before looking. Option B confuses persistence with correctness; saved instructions stay exactly as written while the business moves on. Option A ties review to vendor events only; a new model or interface is one trigger, but a changed policy or source document is the more common reason a project starts giving wrong answers.

Answer: B
Effective system-level instructions, a Domain 5 objective, set durable process rules: which definitions govern, how to handle missing data, how to separate fact from opinion, and what structure to follow; that is option B. Option A optimizes for impression and lets the metric set and layout drift, so weeks cannot be compared. Option C forbids the clarification that missing or ambiguous data requires. Option D instructs Claude to fabricate figures and hide that they are estimates, which is the hallucination risk an instruction should prevent.

Answer: B
On paid plans, Projects use Retrieval-Augmented Generation that can expand effective knowledge capacity roughly tenfold once knowledge approaches the base context limit, chunking and retrieving by relevance rather than always reading every file in full; so the real question is whether retrieval quality holds up on real questions, which is option B. Option C assumes full-document reading is the only retrieval mode, which is exactly what RAG chunking replaces once the base limit is approached. Option D jumps to deleting sources before checking whether retrieval already handles the volume. Option A is a non sequitur; nothing here changes which plan the account is on, and the RAG expansion is itself a paid-plan behavior.

Answer: B
Team/Enterprise Projects are shared at an individual permission level, so the owner can change one person's access without waiting on anything else; a personnel change is exactly the kind of trigger that should prompt an out-of-cycle update, which is option B. Option D treats the calendar as the only trigger, leaving edit access open on a live Project for six weeks after the person who held it is gone. Option A is true only in the narrow sense that the account itself is gone; any lingering session, integration, or credential tied to that access is a real exposure the fixed schedule does not address for weeks. Option C hands responsibility for revoking access to the person who is leaving instead of the owner who controls the Project's sharing settings.

Answer: B
Auto memory is built as Claude works and saves learnings across that developer's sessions, but it is not the same as the shared, versioned CLAUDE.md file every teammate's session reads at the start; carrying a discovery to the whole team still means writing it into CLAUDE.md, which is option B. Option D describes an automatic sync from memory to CLAUDE.md that does not exist; the file is a maintained document, not a generated log. Option A invents a Skills behavior that has no basis in how memory or Skills work. Option C is simply false about what auto memory can hold; nothing restricts it to pricing or context-limit facts.

Answer: B
A Project's knowledge base is the set of documents every chat in that Project can reference, so a policy change that bans a data category is a configuration change that has to reach documents already sitting in that knowledge base, not just new ones: option B. Option A treats the policy as forward-only, leaving already-banned data live and retrievable in Projects the policy is meant to cover. Option D asks Claude to ignore data that remains uploaded and retrievable, which does not remove the exposure the policy addresses. Option C changes a file's name while the banned identifiers stay in its content, which a document upload exposes regardless of what the file is called.

Answer: C
Enterprise plans provide admin controls, usage analytics, and observability, including an Analytics API and OpenTelemetry, over Cowork sessions, which is the built-in way to get org-wide visibility: option C. Option A substitutes manual, unreliable self-reporting for the tooling built for this purpose and does not scale across hundreds of employees. Option B removes connector access altogether, which stops the workflows the officer is trying to observe rather than monitoring them. Option D assumes no configuration is available or needed, when Enterprise specifically adds the oversight tooling this scenario calls for.

Answer: A
A Project's knowledge base holds documents that every chat in the Project can reference, and paid-plan Projects retrieve relevant passages by relevance rather than requiring the whole document to be re-supplied on every question, which is the direct fix here: option A. Option D keeps the same paste-the-whole-manual pattern and only asks for a faster summary, not a different retrieval mechanism. Option C trades one problem for another, asking the technician to know which of 900 pages holds the answer before asking the question. Option B changes which model answers, but Haiku 4.5's smaller 200K-token window makes repeatedly pasting a 900-page manual harder, not easier, and does not address why the manual is being re-pasted at all.

Answer: C
Team and Enterprise Projects can be shared at specific permission levels, 'Can view' (read-only) or 'Can edit' (full modification), so a client who should only read outputs and sources gets 'Can view': option C. Option D grants edit access and substitutes an informal request for an access control, leaving the instructions and knowledge base open to accidental change. Option A shares org-wide at edit level, which is broader than the client needs and exposes the Project to every employee too. Option B sidesteps the Project's sharing settings entirely and hands over static transcripts, which will not stay current as the Project's knowledge and instructions are updated.

Answer: D
CLAUDE.md is a markdown file at the project root that Claude Code reads at the start of every session, and it is meant for exactly this: architecture decisions and preferred libraries the whole team should follow, which is option D. Option C leaves the fix to auto memory, which is built during a session as Claude works and is not guaranteed to capture and generalize a specific past decision the way a written standard does. Option A only reaches whichever chat happens to raise the topic again and leaves every other session, and every other engineer, without the standard. Option B depends on a new hire finding and correctly interpreting an old chat thread instead of Claude Code itself being configured to know the rule.

Answer: B
Claude Cowork connects to workplace tools such as Google Drive, Gmail, and DocuSign so it can act on real documents and data; the configuration task is to enable the specific systems the workflow touches, which for new-hire paperwork is Drive and DocuSign: option B. Option A enables unrelated connectors the task does not use, widening what Cowork can reach for no benefit to this workflow. Option D substitutes a generic connector for the actual system in use, which does not connect to the onboarding files sitting in Google Drive. Option C avoids connectors altogether and pastes content by hand, giving up the direct document access Cowork's connectors are built to provide.

Answer: A
A Project's custom instructions are read in every chat within that Project, so putting the checklist there gives every analyst's chat the same standing rules without repeating it. Option B depends on one analyst manually relaying the checklist and drifts as soon as someone forgets or edits it. Option C stores the checklist in the lead's personal memory, which is tied to that individual and does not configure the shared Project or reach the other analysts. Option D relies on an outside email and each analyst's own notes, neither of which Claude reads inside the Project, so the checklist would need to be retyped into every chat.

Answer: C
Projects hold two durable things: project instructions and a project knowledge base that every chat in the project reads, which is exactly what recurring research with a stable brief needs; hence option C. Option A leans on memory, which saves topics Claude notices rather than a curated document set, and each project has its own memory space anyway. Option B rebuilds the reference set from the open web on every run, so the sources drift and are never maintained. Option D publishes confidential documents into a shareable artifact, a data-handling failure, not a workspace.

Answer: A
Domain 3 aligns model selection with cost, speed, and quality. Anthropic's model guidance recommends starting efficiency-first for high-volume, straightforward tasks: begin with Claude Haiku 4.5, test on a sample, and upgrade only if a capability gap appears, which is option A. Option C pays the highest per-token price and the slowest latency in the lineup for drafts that need no deep reasoning. Option D chooses by tier label rather than by measured fit. Option B picks a plausible model but skips the evaluation that proves it meets the threshold.

Answer: D
Domain 2 includes selecting output formats: artifacts, inline text, or structured data. Claude creates an artifact for content that is substantial, self-contained, and likely to be edited or reused; a quick inline answer suits a one-time reply. So the deciding factor is option D. Option B is appearance, not workflow. Option C forces one format onto tasks with different review and reuse needs. Option A is the wrong goal: a stakeholder deliverable should be reviewed regardless of format, and the format choice does not change that duty.

Answer: B
Domain 3 covers context limits and memory: knowing when to restart, summarize, or persist. Stale assumptions in a long thread keep steering answers, so option B carries forward only what is still true and states the task fresh. Option D mistakes capacity for relevance; a larger context window still contains the outdated turns and gives them weight. Option A adds more history, including more stale material, to the problem. Option C removes the requirements that define the task while leaving the outdated assumptions in place, which inverts the fix.

Answer: B
Claude Code is the agentic coding tool that reads a codebase, edits files, runs commands, and integrates with dev tools, available across the terminal, IDE extensions, a desktop app, and the web, which is exactly this bug-fix workflow: option B. Option A's Project knowledge base lets chats reference uploaded files, but it does not edit the repository or run the test suite; the developer would still do that work by hand. Option C describes Cowork accurately as a working session in a user's files and connected apps, but Anthropic distinguishes Cowork's general office work from Claude Code's developer-focused agentic coding inside a codebase, which is what running a test suite requires. Option D's artifacts support in-place editing and version history for content in the chat window, but they do not execute commands or run a test suite against a repository.

Answer: D
Tool use lets the API call functions the application defines: the backend sends a tools array, Claude responds with a tool_use block naming the function and arguments, the backend executes it, and returns a tool_result, all without a person opening a Claude app; that is option D. Option A's Project knowledge is read within chat conversations by people using the app, not invoked programmatically from a separate backend's automatic reply loop. Option B's Cowork is a working session inside Anthropic's own desktop, web, or mobile app acting on a user's files and connected apps, not a headless component embedded in another system's request-response cycle. Option C's Routines run recurring Claude Code jobs such as a scheduled PR review inside a codebase; they are not built for a per-ticket backend integration returning replies to a support system.

Answer: C
On current models including Sonnet 5, thinking is adaptive: the model decides whether and how much to think based on an effort setting and the query's complexity, with no fixed budget for the engineer to manage; that is option C. Option A's manual budget_tokens cap is the older, deprecated mechanism, not the current one on Sonnet 5. Option B's prefilled assistant response is no longer supported starting with Claude 4.6-generation models and returns an error rather than starting a reasoning block. Option D's manual chain-of-thought prompting is described as a fallback for models or configurations where thinking is off, not the most reliable technique on a model whose adaptive thinking is already available.

Answer: B
The Batch API prices at about half the base per-token rate specifically for workloads that can tolerate asynchronous processing, and an overnight batch with no one waiting on an individual reply is that case exactly: option B. Option A pays the standard real-time rate on the most expensive model in the lineup for a task with no latency requirement, the precise situation batch pricing exists to discount. Option C's caching discount applies to a repeated shared prefix read from cache; ten thousand distinct, independent line items have no long common prefix to cache, and its discount margin is not interchangeable with the Batch API's. Option D's usage tier governs requests-per-minute, input-tokens-per-minute, and output-tokens-per-minute limits, not the per-token price, so raising it would not lower the bill.

Answer: B
Skills are reusable, packaged workflows, such as a shared /deploy-staging command, that let a team define a process once and reuse it consistently instead of re-describing it each time: option B. Option A's hook enforces an automatic, non-negotiable step around every commit regardless of intent, which is a different problem from giving engineers one shared, invocable deploy command they choose to run. Option C keeps the process as prose in a file read at session start, still asking each engineer to phrase a request in a way that reaches it, exactly the inconsistency the team wants to eliminate. Option D's subagents are for isolated, independent parallel workstreams, not for packaging one shared, repeatedly invoked deployment procedure.

Answer: C
Hooks run shell commands before or after Claude Code actions, such as auto-formatting after every edit or running lint before a commit, and are the way to enforce an automatic, non-negotiable step distinct from an instruction the model could skip: option C. Option A's CLAUDE.md sets standards and context Claude reads at session start, but it remains an instruction the model could still skip under a competing priority, not a guarantee. Option B's auto memory saves learnings across sessions as Claude works; it is not a mechanism for compelling a specific command to run before a specific action. Option D's subagents split independent work into parallel workstreams; they do not gate the very commit already in progress on a required check.

Answer: A
Prompt caching is built for exactly this pattern: a stable long system prompt, tool set, or reference document reused across many requests with only the final turn changing, cached at a breakpoint and read back at roughly a tenth of the base input price; that is option A. Option B's Batch API discount is about tolerating asynchronous processing, not about the prompt repeating; it does not specifically target the reprocessing cost of an unchanged prefix, and this workflow returns replies per request rather than overnight. Option C changes model size, but every model still reprocesses the full repeated prefix at full price without a cache breakpoint, so a smaller model does not remove the repeated-processing cost. Option D is not model-specific: any of the current models supports caching, and the 1-hour TTL is simply the higher-priced alternative to the 5-minute default, not a Fable 5.1 exclusive.

Answer: C
Anthropic's own model guidance says to start with Claude Opus 5 for most workloads and move to Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when evals on Opus 5 at higher effort still fall short, which is exactly this team's evidence. Option A moves to Sonnet 5, the model built for everyday speed and cost efficiency, a step down in reasoning depth from the model that already fell short. Option B moves to Haiku 4.5, the fastest and cheapest model with a smaller 200K context window and no adaptive-effort control, even less suited to a long, interdependent refactor. Option D removes thinking rather than adding capability; Opus 5 only allows disabling thinking at effort high or lower, and cutting the deliberation the eval showed was needed does not recover missed steps.

Answer: D
Fable 5.1, Opus 5, and Sonnet 5 all share a 1M-token context window, so the extraction fits any of them, and the task is a straightforward pass rather than demanding reasoning, so Sonnet 5's lower price is the fitting choice: option D. Option B fails outright: Haiku 4.5's context window is 200K tokens, well short of 700,000, so it cannot hold the archive at all no matter how cheap or fast it is. Option C treats Opus 5's status as a good general starting point as a requirement for large documents, but Sonnet 5's window already covers the archive at a lower per-token price. Option A applies the guidance for demanding, long-horizon agentic work to a single-pass extraction that needs neither Fable 5.1's reasoning depth nor its $10/$50 per-MTok price.

Answer: C
Anthropic describes Claude Cowork as 'Claude Code for the rest of your work,' built to work directly in a user's files, folders, and connected apps such as Google Drive, Gmail, and DocuSign rather than a chat back-and-forth, which matches this multi-app task: option C. Option A names a real connection method, MCP, but Claude Code is positioned as a developer-focused agentic coding tool for a codebase, not the office-file working session this manager needs. Option B loses the direct file-and-app action Cowork is built for, and pasting dozens of contracts one at a time does not scale to reorganizing a folder or routing signatures. Option D's Project sharing permissions govern who can view or edit a project's knowledge; the described capability to act on files and route them through DocuSign as a working session belongs to Cowork, not a Project's connector.

Answer: A
Diagnosis starts by making the failure observable and changing one thing at a time; that is option A. A report that worked and now does not usually reflects a changed input, source document, or instruction, and the comparison finds which. Option B discards the evidence, and a regenerated prompt may reintroduce the same fault. Option C changes three variables at once, so even a fix teaches nothing about the cause. Option D fixes on one suspect first; a model update is one possibility among several, and the cheaper checks come earlier.

Answer: C
The observed fault is structural completeness, so the targeted fix names the sections and verifies them: option C. Anthropic's prompting guidance is to control the format explicitly and state what a complete output contains. Option B changes the sources, which affect content accuracy, not whether a section appears. Option D adds reasoning depth and creative license to a problem of omission; more thinking does not add a section nobody asked for. Option A removes the only structural cue and makes omissions more likely, not fewer.

Answer: D
A durable improvement lives in the maintained configuration, not in one conversation: option D puts it in the shared project instructions or template, records why, and checks it on representative inputs before everyone relies on it. Option A strands the fix in personal history; memory is per user and per project, so colleagues never receive it. Option B removes the yardstick that showed the fix worked, so future regressions go unnoticed. Option C generalizes from one success and skips the testing that distinguishes a fix from a lucky run.

Answer: B
Optimization removes repeated low-value work while keeping quality controls. Reformatting every week is a stable, describable step, so the fix is to put the template in the instructions and check conformance: option B. Option C trades a quality control for time and does not touch the actual bottleneck. Option D increases the volume reviewers must reshape. Option A spreads the manual work across more people instead of eliminating it, raising cost for the same output.

Answer: D
The workload is a textbook prompt-caching case: a stable, repeated prefix, the system prompt plus reference workbook, with only the final turn changing, and caching exists to cut cost and latency for exactly this repeat pattern by reading the unchanged prefix from cache rather than reprocessing it, which is option D. Option A swaps models without addressing why cost and latency scale with report volume; a cheaper model still reprocesses the same 11,000 tokens on every call. Option B removes thinking, which is unrelated to a prompt dominated by a repeated reference document rather than reasoning depth. Option C shrinks the output ceiling, but the growing cost here comes from repeatedly billing the same large input, not from output length.

Answer: A
Prompt caching requires an exact match of all content up to and including the breakpoint; any change earlier in the prompt, including one new bullet inserted above the cached portion, invalidates the cache from that point forward, so every request goes back to paying a cache-write price until the new prefix has been cached again, which is option A. Option B misstates the TTL: caching does not disable itself project-wide on any edit, it simply requires the specific prefix to match on each call. Option C has the length rule backwards; adding text makes a prompt longer, not shorter, and the workload was already well above the roughly 1,024 to 4,096 token minimum before the edit. Option D invents an ITPM cause, but cached input tokens are what generally do not count against ITPM in the first place, and nothing in the scenario describes a 429 rate-limit response.

Answer: B
The presence of a retry-after header is what separates an ordinary rate limit from a spend-cap block: the API's continuously-replenishing token-bucket RPM/ITPM/OTPM limits return 429 rate_limit_error with retry-after stating exactly how long to wait, so option B follows that header and retries with backoff. Option A misdiagnoses the response; a monthly spend-cap 429 carries no retry-after header at all, which is the opposite of what is present here. Option C changes max_tokens, but OTPM is measured on actual output tokens generated, not the requested ceiling, so trimming it does nothing to clear a request-rate limit. Option D assumes an Evaluation-tier problem without evidence; a tier change is a separate process and does not explain a header that already states exactly when to retry.

Answer: C
Thinking improves performance on exactly this kind of task, complex multistep math and analysis where quality depends on intermediate work that would otherwise be skipped or compressed, and adaptive thinking scales with the effort setting, so raising effort for a genuinely multistep reconciliation is the targeted fix, option C. Option D adds format examples but does not address inconsistent intermediate arithmetic, which is a reasoning-depth problem, not a formatting one. Option A shortens the output budget, which pushes toward more compression and more skipped steps, the opposite of what an unreliable multistep calculation needs. Option B chases latency at the moment accuracy is the open problem, and switching models does not itself fix an effort setting that is too low for the task.

Answer: C
Claude Haiku 4.5's thinking is extended and manual rather than adaptive, and the effort parameter is not supported on it, which is exactly why the request was rejected; the model's reasoning depth is instead controlled by a thinking budget configuration, so option C names the actual mechanism. Option B overstates the limitation; Haiku 4.5 does support thinking, just not through the effort parameter. Option D invents a header requirement that has nothing to do with why effort fails specifically on this model. Option A invents a batch-only restriction on effort that does not exist; it is the model, not the request type, that determines whether effort is accepted.

Answer: A
A default ephemeral cache has a 5-minute TTL, so a request cadence of roughly 20 minutes between calls will find the cache already expired almost every time and pay the higher cache-write price instead of the discounted cache-read price; that mismatch between request cadence and TTL is the actual cause, option A. Option B misunderstands what a breakpoint does; adding more breakpoints marks more places to cache, it does not extend how long any one cached segment survives. Option C assumes a billing error, but a 5-minute cache write costing roughly 1.25 times the base input price while a read costs roughly 10 percent of it is the documented, correct pricing relationship. Option D confuses max_tokens, an output-length setting, with cache lifetime, which the two have no connection to.

Answer: C
A 429 with error_code enforced_spend_limit_reached and no retry-after header is the API's distinct signal for an exhausted monthly spend cap, not an ordinary rate limit; unlike a token-bucket RPM/ITPM/OTPM limit, retries do not clear on their own until the cap resets at the start of the next month or the organization requests a higher limit, which is option C. Option A applies backoff logic meant for ordinary rate limits, which recover within minutes because the token bucket replenishes; a spend cap does not replenish that way. Option B misreads the cause as a token-volume problem; splitting requests changes nothing about a spend cap, which counts total spend rather than the size of any one call. Option D invents a tier-demotion mechanism the error does not describe; nothing about enforced_spend_limit_reached involves usage-tier status or a 24-hour reversal.

Answer: C
OTPM is measured on actual output tokens generated, not on the max_tokens ceiling a request asks for, so a high max_tokens setting does not by itself cost more or trigger this limit; the team should check what is really driving output volume, such as verbose responses or a large thinking budget, before cutting a setting that was not the cause, which is option C. Option A assumes a proportional relationship between the requested ceiling and actual usage that the rate-limit mechanics do not describe. Option D invents a rule that ITPM and OTPM are always throttled together; they are separate, independently measured limits. Option B invents an Evaluation-tier-only restriction on OTPM that does not exist; OTPM applies across tiers, and only the limit's numeric value differs by tier.

Answer: B
The default tool_choice auto leaves the triggering decision to the model, and that boundary is promptable: light system-prompt language can nudge Claude toward or away from calling a tool, so an edit aimed at reducing over-eager calls can suppress genuine ones too, exactly what changed right before the regression here, which points to option B. Option A rewrites the tool definition itself, but nothing in the scenario implicates the tool's name, description, or schema, only a system-prompt edit made around the same time as the regression. Option C sets tool_choice to the value that was already the default before and after the edit, so it does not touch the system-prompt language that actually changed. Option D adds formatting examples, but the reported problem is a missing tool call, not disorganized output, so more formatting examples address the wrong defect.

Answer: C
Cached input tokens generally do not count against ITPM on most current models, only uncached input tokens and cache-creation tokens do, so caching the stable repository summary lets the workflow sustain much higher effective throughput against the same ITPM number, which is option C. Option A still sends the same total token volume across more calls, and ITPM counts total tokens per minute regardless of how many requests carry them. Option B describes the Batch API backwards; batch requests are processed asynchronously at a 50 percent discount, not synchronously, so switching to it does not by itself resolve a synchronous-request ITPM ceiling. Option D targets OTPM, the output-token limit, but the problem described is an input-tokens-per-minute limit driven by a large repeated context, not by output volume.