CertSafari

    Free Anthropic Claude Certified Associate – Foundations (CCAO-F) - December 7 onwards Sample Questions

    35 free sample questions from our bank of 447+, covering every exam domain, with answers and detailed explanations. Updated November 2026.

    1: Task Definition and Prompting

    1.1: Prompt Construction

    1.Priya, an HR manager, opens a new chat and types: "Write an announcement about our new parental leave policy." Claude returns a polished announcement with leave durations and pay figures that do not match the real policy. What change to the request most directly fixes the problem?

    1. A.Paste the written policy into the chat, name the staff as the readers, and tell Claude to draw its figures from that text alone.
    2. B.Give Claude the role of a senior HR communications writer, name the staff as the readers, and ask it to match the company's usual tone.
    3. C.Ask Claude to re-read its draft, correct the figures it doubts, and repeat that pass until the leave numbers look consistent.
    4. D.Add a line telling Claude to stay accurate and avoid guessing about the details, and leave the one-line request itself as it is.
    Show answer & explanation

    Correct answer: A — Paste the written policy into the chat, name the staff as the readers, and tell Claude to draw its figures from that text alone.

    • A. Correct. Claude has no access to the company's policy in a new chat, so the real figures must be supplied. Pasting the document, naming the readers and limiting Claude to that text removes the gap it was filling with guesses.
    • B. Incorrect. A role and a matched tone improve how the message sounds, but they add no policy facts. Claude still lacks the real durations and pay figures and will keep filling them in with plausible invented ones.
    • C. Incorrect. Re-reading its own draft cannot reveal which figures are wrong, because Claude has never seen the real policy. Repeated passes can make the invented numbers consistent with each other without making them correct.
    • D. Incorrect. A general instruction to be careful does not give Claude the missing policy. With the one-line request unchanged it still has no accurate durations or pay figures to use, so the inaccuracies remain.

    1.1: Prompt Construction

    2.A consultant attaches interview notes from eight clients and asks Claude to list the key themes. The answer includes themes that appear nowhere in the notes. Which prompt change best grounds the response in the source material?

    1. A.Tell Claude to add widely documented industry trends, merge them with the clients' remarks, and order themes by relevance.
    2. B.Tell Claude to paraphrase each theme in its own words, group related remarks into broader categories, and keep it concise.
    3. C.Tell Claude to draw on the attached notes alone, cite a supporting excerpt per theme, and flag topics the notes skip.
    4. D.Tell Claude to reason step by step through the notes, rate its confidence in each theme, and answer in a fixed format.
    Show answer & explanation

    Correct answer: C — Tell Claude to draw on the attached notes alone, cite a supporting excerpt per theme, and flag topics the notes skip.

    • A. Incorrect. Adding industry trends brings in themes that none of the clients mentioned, which is the problem the consultant needs to remove. Merging them with the clients' remarks also hides which themes came from the notes.
    • B. Incorrect. Paraphrasing and grouping remarks into broader categories gives Claude room to introduce abstractions the clients never expressed. Nothing here limits the answer to the notes or lets a reader check a theme against them.
    • C. Correct. Limiting Claude to the attached notes and requiring a supporting excerpt for each theme makes every theme traceable to a client statement. Flagging topics the notes skip gives Claude a way to report a gap instead of filling it.
    • D. Incorrect. Step-by-step reasoning, self-rated confidence and a fixed format do not restrict Claude to the notes. Claude can still list an unsupported theme and rate it confidently, and a confidence score is not evidence from the source.

    1.2: Task Decomposition and Sequencing

    3.You are a marketing manager producing a 15-page onboarding guide for a new software product. You have product notes, a support-ticket export and an approved style guide. One request for the whole guide returned a generic draft that missed several ticket themes. Which approach best applies task decomposition?

    1. A.Resubmit the single request with a senior-technical-writer role, a minimum word count per chapter and a note to cover the support tickets in depth.
    2. B.Have Claude draft an outline from the notes and tickets, revise it yourself, then request one chapter at a time with that outline attached as reference.
    3. C.Add the ticket export to a Project's knowledge base, then request the full guide again and let Claude retrieve the relevant tickets while it drafts.
    4. D.Turn on extended thinking, then resubmit the full guide request and let Claude plan the structure internally before drafting the fifteen pages.
    Show answer & explanation

    Correct answer: B — Have Claude draft an outline from the notes and tickets, revise it yourself, then request one chapter at a time with that outline attached as reference.

    • A. A role, a length floor and a reminder about the tickets are still prompt tweaks to one request. The guide is produced in a single pass with no reviewed structure, so missed ticket themes stay likely and nothing is checked along the way.
    • B. Correct. Splitting the work into an outline you review and then one request per chapter gives each step a narrow goal. Ticket coverage is fixed in the approved outline and can be checked chapter by chapter.
    • C. A Project's knowledge base changes where the material is stored, not how the work is divided. The guide is still one large request, and retrieval gives no assurance that each ticket theme is assigned to a chapter or reviewed.
    • D. Extended thinking lets Claude reason before answering, but the plan stays internal and the output is still one fifteen-page draft. There is no outline to approve and no chapter-level checkpoint where a missing theme could be caught.

    1.2: Task Decomposition and Sequencing

    4.A consultant asked Claude for a six-section strategy memo in one chat. Sections 1 and 3 to 6 are approved, but section 2 is too formal for an audience of store managers. What is the most efficient next step?

    1. A.Use the retry button on the memo reply to regenerate it, then check whether the second section comes back friendlier and keep that version of the memo.
    2. B.Edit the original request to add that readers are store managers and the voice should be friendlier, then regenerate the memo and review the new full draft.
    3. C.Ask Claude to soften the voice across the full memo so the sections read evenly, then have the team sign off on the revised version of each part again.
    4. D.Reply in the same chat naming the second section and its store-manager readers, ask for a more casual rewrite of that section, and leave the other five as written.
    Show answer & explanation

    Correct answer: D — Reply in the same chat naming the second section and its store-manager readers, ask for a more casual rewrite of that section, and leave the other five as written.

    • A. Retrying the reply regenerates all six sections, so the five approved ones can change as well. It also gives Claude no information about the store-manager readers or the tone wanted, so a friendlier second section would be luck.
    • B. Regenerating the memo from an edited request produces six new sections, so wording in the five approved ones can shift. The consultant would have to review the full draft again, which is more effort than fixing the single section that is off.
    • C. Softening the voice across the memo rewrites sections that were already approved to fix a problem in only one. That forces a new round of sign-off from the team on text that was settled, which is the opposite of an efficient fix.
    • D. Correct. A reply in the same conversation keeps the memo context. Naming the section and its readers and asking for a more casual rewrite limits the change to that section, so the five approved sections stay intact and need no further review.

    1.3: Prompt Diagnosis and Iteration

    5.A marketing coordinator asks Claude: "Give me the single best tagline for our new reusable water bottle." The prompt already includes the product details, the brand voice and the target customers. Claude returns one safe, generic tagline. Two follow-ups saying "make it more creative" produce near-identical lines. She needs something distinctive for the launch next week. What is the best way to proceed?

    1. A.Request fifteen candidate taglines, each built on a different angle such as humor or durability, then shortlist a few to refine together.
    2. B.Start a new conversation with the same one-tagline request, then compare the result with earlier replies and keep the strongest.
    3. C.Add the mission statement and founder story to the prompt, then resend the same request for one tagline and review the result.
    4. D.Assign Claude the role of an award-winning advertising copywriter, then ask it for one final tagline and review the result.
    Show answer & explanation

    Correct answer: A — Request fifteen candidate taglines, each built on a different angle such as humor or durability, then shortlist a few to refine together.

    • A. Correct. A request for one line pushes Claude toward the safest wording, and rewording it twice changed nothing. A wide set of lines from different angles suits a creative task. The coordinator can then pick promising ones and refine them in the same conversation.
    • B. Incorrect. The request still asks for one line, so a fresh conversation brings the same pull toward safe, generic wording. Rerunning it only gives another draw of that kind of line, so it is not a reliable way to find something bold before next week's launch.
    • C. Incorrect. The prompt already includes the product details, brand voice and target customers, so missing context is not the problem. Extra background such as a mission statement or founder story leaves the one-answer request unchanged, and it will keep producing a safe line.
    • D. Incorrect. A copywriter role is a real prompting technique, but it does not change the shape of the request. Claude is still asked for one final tagline, so it will still pick the safest line instead of offering a range of ideas to choose from.

    1.3: Prompt Diagnosis and Iteration

    6.A department head attaches three agency proposals and types: "Analyze these proposals." Claude returns an accurate, lengthy side-by-side comparison of every section but no recommendation. The department head has a $20,000 budget, must pick one agency by Friday and cares most about delivery speed, but none of this was in the prompt. What is the best diagnosis and fix?

    1. A.Unfocused output: no format was requested, so ask for a one-page summary table comparing price, timeline and scope across the three agencies.
    2. B.Missing evidence: Claude saw only the written proposals, so attach each agency's case studies and client references and ask for an evaluation.
    3. C.Overloaded input: three proposals in one request split the focus, so analyze each in its own chat and compare the three results yourself.
    4. D.Missing criteria: Claude did not know what mattered, so give the $20,000 cap, Friday deadline and speed priority, and ask for one pick.
    Show answer & explanation

    Correct answer: D — Missing criteria: Claude did not know what mattered, so give the $20,000 cap, Friday deadline and speed priority, and ask for one pick.

    • A. Incorrect. The comparison was already accurate, and a tighter table of price, timeline and scope is still a comparison. Without the decision, budget and priority, Claude has no basis for recommending an agency.
    • B. Incorrect. Case studies and references add more material to compare, but the output was not short of material. What was missing was the department head's decision, budget and priority, which no agency document contains.
    • C. Incorrect. Claude handled all three proposals accurately in one pass, so volume was not the problem. Separate chats would yield three isolated analyses, still without criteria, and leave the final pick to the department head.
    • D. Correct. Claude was asked only to analyze, so it described everything and recommended nothing. Giving it the budget, the deadline and the speed priority, and asking for a single pick, supplies the criteria to weigh the proposals and reach a recommendation.

    2: Output Evaluation and Quality Control

    2.1: Output Evaluation and Fitness for Purpose

    7.A finance analyst has Claude draft a two-paragraph status email for the CFO, who wants the bottom line first and no jargon. Claude's draft opens with three sentences of pipeline methodology and uses terms like "ETL backfill" and "idempotent reruns" before reaching the result. What is the main fitness-for-purpose problem?

    1. A.The draft is too thin on technical support for a CFO, so it needs more methodology detail placed ahead of the conclusion.
    2. B.The draft contains spelling and grammar errors, so a human editor should proofread and correct it before the CFO receives the email.
    3. C.The draft misstates the pipeline results, so the figures should be re-verified against the source data before style edits are made.
    4. D.The draft's order and vocabulary suit a technical reader, not the CFO, so Claude should be asked to reorder and simplify it.
    Show answer & explanation

    Correct answer: D — The draft's order and vocabulary suit a technical reader, not the CFO, so Claude should be asked to reorder and simplify it.

    • A. Incorrect: the CFO asked for the bottom line first, and the draft already puts methodology ahead of it. Adding more technical support would worsen that mismatch, not fix it.
    • B. Incorrect: nothing in the scenario points to spelling or grammar mistakes. The draft's weakness is its structure and vocabulary relative to the CFO, which proofreading would leave untouched.
    • C. Incorrect: no sign is given that the results are wrong. The draft does reach a result, just too late and in terms the CFO will not follow, so re-verifying the data would not address the mismatch.
    • D. Correct: the draft is written for a technical reader, while the CFO needs the result first and plain language. Fitness for purpose includes audience fit, and Claude can revise the draft once told the reader's needs.

    2.3: Review Depth and Escalation

    8.A department head asks Claude to summarize a 40-page supplier contract. The summary says the agreement renews automatically unless cancelled with 30 days' notice. The head must decide this week whether to renew, a commitment of about $200,000. What is the best next step?

    1. A.Ask Claude to re-read the contract and restate the notice terms, then proceed if it gives the same 30-day period again.
    2. B.Ask Claude to cite the clause number for the notice period, then proceed once the citation is specific and consistent.
    3. C.Run the contract through a second AI tool for its own summary, then proceed if both summaries show 30 days' notice.
    4. D.Locate the notice clause in the contract itself, then have the legal team confirm the notice terms before deciding.
    Show answer & explanation

    Correct answer: D — Locate the notice clause in the contract itself, then have the legal team confirm the notice terms before deciding.

    • A. Asking Claude to re-read the contract tends to reproduce the same reading. A repeated 30-day figure is not a check against the document itself, and a $200,000 commitment needs one.
    • B. A clause number that Claude supplies can be wrong or point to different wording. A specific-looking reference does not show the notice period was read correctly, and nobody has compared it with the contract text.
    • C. A second AI summary can repeat the same misreading of the contract wording. Agreement between two generated summaries is not a comparison with the source, so a $200,000 decision would still rest on unverified output.
    • D. Correct. The decision is binding and costs about $200,000, so the notice terms must be checked in the original contract. The legal team, who own contract interpretation, should confirm them before the head decides.

    2.3: Review Depth and Escalation

    9.A support lead wants every Claude-drafted customer reply sent without a person reading it, because a one-week sample showed 95% were fine. The replies include order-status acknowledgments and also replies to complaints that promise refunds or compensation. What is the best approach?

    1. A.Have Claude send replies directly and audit a weekly random sample afterwards, correcting and apologizing for errors found.
    2. B.Have a reviewer approve each reply before it is sent, acknowledgments included, and keep the weekly sample as a quality check.
    3. C.Have Claude rate its confidence in each draft and send high-confidence ones directly, holding the rest for a person to read.
    4. D.Have a person read replies that commit money or answer a complaint before sending, and audit a sample of the routine ones.
    Show answer & explanation

    Correct answer: D — Have a person read replies that commit money or answer a complaint before sending, and audit a sample of the routine ones.

    • A. A weekly audit only finds mistakes after the replies have gone out. A wrongly promised refund or compensation has already reached the customer by then, and the 95% average says nothing about how often the high-stakes complaint replies fail.
    • B. Pre-send approval protects the refund and complaint replies, but applying it to routine order-status acknowledgments spends reviewer time where an error costs little. Review effort should scale with the consequence of a mistake, which a blanket rule ignores.
    • C. A model's own confidence is not tied to the cost of an error. A fluent, confident reply that promises an unauthorized refund would be sent unread, while a harmless acknowledgment might be held. Routing has to follow what the message commits the company to.
    • D. Correct. Review is tiered by consequence: a person reads the replies that commit money or respond to complaints before they go out, where an error is costly. Sampling the routine acknowledgments keeps oversight proportionate and still catches quality drift.

    2.4: Output Adaptation and Format Selection

    10.A team lead asks Claude to rewrite a project update for a Slack channel where colleagues read on phones between meetings. Claude produced a formal, multi-section report with headings. Which revision suits the channel?

    1. A.Keep the report as is but add a bolded summary line and a linked contents list at the top, for jumping to a specific section.
    2. B.Save the report as a PDF and post it in the channel with a one-line teaser, leaving the full detail in the attached file.
    3. C.Rewrite it as a brief post with no section titles, opening with the status and the one action needed, then three short bullets.
    4. D.Keep the report's sections but cut each to one short sentence and bold the key figure in each, leaving the structure intact.
    Show answer & explanation

    Correct answer: C — Rewrite it as a brief post with no section titles, opening with the status and the one action needed, then three short bullets.

    • A. Incorrect. A summary line and contents list leave the formal multi-section report in place and add navigation on top of it. Phone readers between meetings still scroll a long document in a channel meant for quick messages.
    • B. Incorrect. The update itself is not rewritten for the channel. Readers on phones must open an attachment to learn anything beyond the teaser, which costs more effort than a message they can read in the feed.
    • C. Correct. A brief post that opens with the status and the one required action, followed by three short bullets and no section titles, matches the casual, scannable style of a chat channel read on a phone.
    • D. Incorrect. Trimming every section to one sentence shortens the text but keeps the formal, sectioned report format. The key action stays spread across several labelled parts instead of leading the message.

    2.1: Output Evaluation and Fitness for Purpose

    11.An accounts payable clerk has Claude extract vendor, invoice number and total from 80 scanned invoices into a table that will drive this week's payment run. The clerk compares five randomly chosen rows with the original invoices and all five match. What is the soundest next step?

    1. A.Spot-check ten more random rows against the originals, then release the payment run if those also match the invoices.
    2. B.Have Claude re-extract the invoices in a second pass, then release the payment run if both tables agree on each row.
    3. C.Check the ten largest totals against the originals, then release the payment run with the smaller invoices left as extracted.
    4. D.Compare each row with its invoice, or reconcile the grand total with the batch total, then release the payment run.
    Show answer & explanation

    Correct answer: D — Compare each row with its invoice, or reconcile the grand total with the batch total, then release the payment run.

    • A. Incorrect. Fifteen checked rows still leave 65 unverified, and each of those rows triggers a payment. One misread vendor, invoice number or total would go out as a wrong payment, so a larger sample does not give the coverage this run needs.
    • B. Incorrect. A second extraction by Claude is not independent of the first, because the same model can misread the same smudged or oddly laid-out invoice in the same way. Two tables agreeing proves consistency, not that they match the invoices.
    • C. Incorrect. Vendor and invoice number feed the payment as well as the total, so a wrong vendor or a duplicated invoice number on a small invoice still sends money to the wrong place. Leaving the smaller rows unchecked leaves that exposure open.
    • D. Correct. Each row drives a payment, so the table needs checking beyond a sample: either a row-by-row comparison with the invoices, or a reconciliation of the table's grand total with the batch total, before the run is released.

    2.2: Hallucination Detection and Verification

    12.A procurement analyst uploads six evaluation documents about a software tool to Claude: three favorable and three critical. The analyst asks for a balanced summary for a steering committee. The summary lists four benefits and one minor drawback, and the critical documents' main concern, limits on exporting data, is missing. What should the analyst do before the summary is shared?

    1. A.Check the summary against the six source documents, add the missing data-portability concern with a citation, and review how much weight each side gets
    2. B.Ask Claude to score how evenly its summary covers the favorable and critical documents, and share the summary once it rates itself 8 out of 10 or higher
    3. C.Share the summary as written with a note telling the committee that it leans toward the favorable documents, and attach the six originals for reference
    4. D.Regenerate the summary twice with an instruction to weight favorable and critical documents equally, and share the draft that lists the most drawbacks
    Show answer & explanation

    Correct answer: A — Check the summary against the six source documents, add the missing data-portability concern with a citation, and review how much weight each side gets

    • A. Correct. The skew and the missing export concern are only visible by comparing the output with the six documents. Restoring the concern with a citation and reviewing the weighting fixes the summary before the committee sees it.
    • B. Incorrect. A score from the model that wrote the skewed summary is not an independent check. It cannot show which content from the critical documents was left out, so the export concern could stay missing while the score is high.
    • C. Incorrect. The summary is already known to omit the main concern of the critical documents. A note and attached originals pass the checking to the committee and leave a one-sided summary in circulation.
    • D. Incorrect. The number of drawbacks listed does not measure faithfulness to the six documents. A regenerated draft can still omit the export concern or add criticism the documents do not support, and neither is caught without comparing against the sources.

    3: Capability, Model, and Context Selection

    3.1: Capability Selection

    13.A team lead asks which capability decides whether Claude can follow a reusable, packaged set of instructions and files for a specialized task, such as producing a branded slide deck, when a request matches it. What is it?

    1. A.Artifacts
    2. B.Skills
    3. C.Connectors
    4. D.Memory
    Show answer & explanation

    Correct answer: B — Skills

    • A. Incorrect: artifacts are the rendered outputs shown beside the chat, not packaged procedures.
    • B. Correct: Skills bundle instructions and resources that Claude can load when a task matches, which makes specialized, repeatable outputs consistent.
    • C. Incorrect: connectors give Claude access to external tools and data sources; they do not define how a task should be done.
    • D. Incorrect: memory stores user preferences and context between chats rather than packaging task procedures and files.

    3.1: Capability Selection

    14.A product manager asks for a customer-interview synthesis. The manager's stem facts: interview transcripts live in a shared Google Drive folder, the output must be a polished one-page document the manager can keep editing, and the manager wants figures such as theme counts calculated accurately. Choose two capabilities that together best fit. Choose two.(Select 2)

    1. A.Code execution, to run a script over the transcripts and tally how often each theme comes up
    2. B.Memory, to carry the main points of the transcripts over from the manager's earlier conversations
    3. C.Google Drive connector, to open the interview files in the place where the team keeps them
    4. D.Research mode, to pull public customer-interview benchmarks and compare themes with the transcripts
    5. E.Slide-deck Skill, to package the synthesis into the company's standard presentation template
    Show answer & explanation

    Correct answers: A, C — Code execution, to run a script over the transcripts and tally how often each theme comes up; Google Drive connector, to open the interview files in the place where the team keeps them

    • A. Correct: running a script over the transcripts produces exact theme tallies, which meets the manager's need for accurately calculated figures. Estimating counts by reading would not give that accuracy.
    • B. Incorrect: memory keeps preferences and context from past conversations, and nothing indicates the transcripts were ever discussed with Claude. It is not a reliable store for the full text of source documents that sit in a shared folder.
    • C. Correct: the Google Drive connector lets Claude open the transcripts where they already live, so nobody has to download or paste them first. Without a way to read the files, no synthesis can start.
    • D. Incorrect: research mode gathers public material, and the manager asked for a synthesis of their own interviews, not a comparison with outside benchmarks. It adds no way to read the private Drive transcripts or to calculate exact counts.
    • E. Incorrect: a slide-deck Skill produces a presentation, but the deliverable is a one-page document the manager can keep editing. It also neither reads the transcripts nor calculates theme counts.

    3.2: Model and Effort Selection

    15.A user believes that choosing Claude Opus and turning on extended thinking guarantees a correct answer. What does extended thinking actually do?

    1. A.It lets Claude reason before it replies, which can improve results on hard problems, takes more time and usage, and does not ensure accuracy.
    2. B.It lets Claude search the web before it replies, which improves how current its facts are, takes more time and usage, and still needs checking.
    3. C.It lets Claude hand a hard question to a larger model, which can improve results on it, takes more time and usage, and does not ensure accuracy.
    4. D.It enlarges the context window Claude works with, which helps on long conversations, takes more time and usage, and still needs checking.
    Show answer & explanation

    Correct answer: A — It lets Claude reason before it replies, which can improve results on hard problems, takes more time and usage, and does not ensure accuracy.

    • A. Correct: extended thinking gives Claude a reasoning phase before its reply. That can help on hard problems, but it costs extra time and usage and does not make the answer reliably right, so the output still has to be checked.
    • B. Web search is a separate capability that has to be enabled on its own. Extended thinking changes how Claude reasons before answering and does not fetch live sources, so it does nothing for how current the facts are.
    • C. Extended thinking runs inside the model the user already picked. It does not pass a question to a larger model, so choosing Opus and enabling it still yields that one model, reasoning for longer.
    • D. The context window is a fixed property of the model and conversation. Extended thinking does not enlarge it. It gives Claude a reasoning phase before it replies, so long conversations gain nothing from it.

    3.3: Context and Memory Management

    16.After moving from retail to a logistics company, Dana notices Claude still frames advice around her former employer's store-operations tools. She wants this to stop in all future chats without losing her other saved preferences. What should she do?

    1. A.Open the memory settings, choose Reset memory, then let Claude relearn her current role and preferences from new chats.
    2. B.Open Settings, go to Your Claude, choose Manage next to memory, then edit or delete the entry about her retail job.
    3. C.Open a new chat, tell Claude to disregard her retail background, then rely on that message in later conversations.
    4. D.Open future chats in incognito mode, then let Claude answer without drawing on the retail entries in her memory.
    Show answer & explanation

    Correct answer: B — Open Settings, go to Your Claude, choose Manage next to memory, then edit or delete the entry about her retail job.

    • A. Incorrect: resetting memory erases the whole memory, including the preferences she wants to keep. Relearning them from new chats means she loses them in the meantime, and nothing guarantees they come back.
    • B. Correct: the memory settings list what Claude has stored and let her edit or delete individual topics. Fixing the stale job entry stops the old framing in later chats while her other saved preferences stay in place.
    • C. Incorrect: an instruction given in a chat shapes only that conversation. The stored memory still contains the retail entry, so later chats can keep framing advice around her old employer.
    • D. Incorrect: incognito chats do not draw on memory at all, so her other saved preferences go unused as well. The retail entry also stays stored, and she would have to use incognito in every future chat instead of fixing it.

    3.3: Context and Memory Management

    17.Kofi pasted a client's contract terms into a normal chat. Later, while reviewing memory, he sees a topic that names that client and its pricing. He wants it removed and wants future client work kept out of memory. What should he do?

    1. A.Pause memory in settings, which erases entries that mention clients and keeps new client details from being saved.
    2. B.Delete the entry in memory settings, then use incognito chats for future work with confidential client details.
    3. C.Ask Claude in an incognito chat to forget the client, which removes the stored entry from the account's memory.
    4. D.Move the client work into a Project, since Projects delete any earlier memory entries tied to the same client name.
    Show answer & explanation

    Correct answer: B — Delete the entry in memory settings, then use incognito chats for future work with confidential client details.

    • A. Wrong: pausing memory stops Claude from using and adding memories but keeps what is already saved, so the client entry would remain stored.
    • B. Correct: deleting the entry in the memory list removes it at the source, and incognito chats are not saved to memory or history, so future confidential work leaves no new entry.
    • C. Wrong: an incognito chat does not read or change stored memory, so a forget request made there cannot remove the existing entry.
    • D. Wrong: a Project does not delete earlier account memory entries; it only gives its chats their own context, so the existing client entry would still be there.

    4: Workflow Design and Integration

    4.1: End-to-End Deliverable Workflows

    18.An analyst used Claude to model a three-year cost projection for a proposal that will be sent to an external client. The code ran without errors and the summary looks reasonable. Who should confirm the projection's assumptions and figures before it is sent?

    1. A.The client's own finance team, who review the projection once the proposal arrives and return corrections in a revised version.
    2. B.Claude in a new session, prompted to audit its own projection and flag the assumptions it finds unrealistic before sending.
    3. C.The analyst or a finance colleague, who traces the inputs and tests the assumptions against source data before it goes out.
    4. D.The engineering team, who review the code for logic errors and confirm the calculations run correctly before the proposal goes out.
    Show answer & explanation

    Correct answer: C — The analyst or a finance colleague, who traces the inputs and tests the assumptions against source data before it goes out.

    • A. This is incorrect because the figures have to be confirmed before the proposal is sent, not after. Letting the client's finance team find the errors puts the firm's credibility at risk and means the client may already have acted on wrong numbers.
    • B. This is incorrect because a new session still uses the same reasoning and has no access to the firm's real inputs, so it can repeat the original mistakes. An audit by the model can support the review, but it cannot replace an accountable person confirming figures an outside client will rely on.
    • C. This is correct because the analyst or a finance colleague can compare the inputs and assumptions with real source data and answer for the deliverable. Figures going to an external client need that human check before they leave the team, however clean the run and the summary look.
    • D. This is incorrect because reviewing the code only shows the calculations are implemented correctly. The code already ran without errors, and that review says nothing about whether the inputs and assumptions are realistic, which is what a finance-qualified person has to confirm.

    4.4: Capability Communication

    19.An analyst proposes that Claude read a 60-page competitor annual report and pull the revenue, margin and headcount figures for a brief going to the sales team. She says Claude 'reads every page perfectly'. What should the capability note say?

    1. A.Claude summarizes prose well but is a poor fit for pulling figures out of a long report, so the analyst keys in the numbers herself and uses it for wording.
    2. B.Claude reads an uploaded report at full fidelity, so the analyst copies the figures into the brief as extracted and reviews only the wording before it goes out.
    3. C.Claude is dependable on the opening pages of a long report and weaker later, so the analyst re-checks figures from later pages and uses early ones as returned.
    4. D.Claude finds and summarizes the main points well but can misread exact figures, so the analyst checks each number in the brief against the report page.
    Show answer & explanation

    Correct answer: D — Claude finds and summarizes the main points well but can misread exact figures, so the analyst checks each number in the brief against the report page.

    • A. Incorrect. Claude can locate and extract revenue, margin and headcount figures from an uploaded report. The sound approach is to use that ability and verify the numbers, not to retype them by hand.
    • B. Incorrect. Claude can misread or transpose figures in long documents, so extracted revenue, margin and headcount numbers need checking before a sales brief repeats them. Reviewing only the wording leaves that risk.
    • C. Incorrect. Misreads are not confined to later pages, so figures from the opening pages need checking against the source as much as the rest. Treating early figures as returned leaves errors in a brief sales will repeat.
    • D. Correct. It credits Claude's strength at locating and summarizing content and still requires checking the exact figures against the report, since misreads are possible and sales will quote the numbers.

    4.3: Repeatable Workflow Design

    20.A marketing team applies the same 6-step brand-review procedure, including a fixed checklist and a scoring template, to documents in many different chats and Projects. Members currently paste the procedure into each conversation, and some skip steps. Which approach lets Claude apply the procedure whenever it is relevant, without pasting it each time?

    1. A.Add the six steps and scoring template to a shared Project's instructions, so Claude applies them to each document reviewed in that Project
    2. B.Package the six steps and scoring template as a Skill, so Claude loads and applies them when a document review comes up
    3. C.Add the six steps and scoring template to personal preferences, so Claude applies them to each document discussed in the user's chats
    4. D.Attach the checklist and scoring template as a file to each chat, so Claude applies the steps to each document reviewed there
    Show answer & explanation

    Correct answer: B — Package the six steps and scoring template as a Skill, so Claude loads and applies them when a document review comes up

    • A. Incorrect. Project instructions only apply to chats inside that one Project. Reviews run in other Projects or in standalone chats would get no procedure, so the team would still paste it there.
    • B. Correct. A Skill bundles instructions and resources such as a template, and Claude loads it when the task matches the Skill's description. The same procedure is then available across chats and Projects without pasting.
    • C. Incorrect. Preferences are set per person, so each member must enter them and some will not. They are also meant for general style and background, and they sit in unrelated conversations instead of being applied when a brand review comes up.
    • D. Incorrect. Attaching the file to each chat is the same manual per-conversation step as pasting, and it is the step people already skip. Nothing makes the procedure available in a chat where nobody attached it.

    4.3: Repeatable Workflow Design

    21.A communications manager's Claude workflow drafts the weekly customer newsletter, which is published to 20,000 subscribers. The workflow has run smoothly for two months and the drafts are usually excellent. A colleague proposes removing the manager's read-through to save time, since the Project's instructions and template already ensure a consistent format. What is the best response?

    1. A.Drop the read-through and rely on the Project's template and instructions to keep the facts, figures and claims in each issue accurate before it is sent
    2. B.Drop the read-through for now and track subscriber complaints for a few weeks, restoring the review only if reports of errors in sent issues come in
    3. C.Keep the read-through before sending, with the manager checking dates, figures and claims against source material and approving each issue for release
    4. D.Swap the read-through for a second Claude chat that fact-checks each draft, and send the newsletter once that chat reports no factual problems
    Show answer & explanation

    Correct answer: C — Keep the read-through before sending, with the manager checking dates, figures and claims against source material and approving each issue for release

    • A. Incorrect. A template and standing instructions fix structure and tone, not content. A draft can follow the format perfectly and still contain a wrong date, figure or claim, so these do not make the facts accurate.
    • B. Incorrect. Complaints only arrive after an error has already been published to 20,000 subscribers, and a sent newsletter cannot be recalled. Waiting for reports makes the audience the reviewer.
    • C. Correct. Consistent output shows the format is reliable, not that the facts are. Because the newsletter goes out externally to a large audience, an accountable person should check the content against sources before release.
    • D. Incorrect. A second Claude chat has no access to the real facts behind the newsletter and can repeat the same mistakes. It leaves no accountable person checking the content before it is published to subscribers.

    4.2: Delegation and Workflow Redesign

    22.A manager must tell a long-serving employee that their role is being eliminated. The manager is short on time and suggests using Claude for this step of the restructuring workflow. HR is applying Delegation criteria to the step. Which assignment is most appropriate?

    1. A.Human-retained: Claude drafts talking points and likely questions beforehand, and the manager holds the conversation in person.
    2. B.Human-reviewed: Claude drafts the notification email, and the manager reviews it and sends it to the employee this week.
    3. C.AI-executed: Claude sends the notification through an email connector, and the manager offers a follow-up meeting the next day.
    4. D.Human-reviewed: Claude holds the notification conversation in a chat window, and the manager reviews the transcript that evening.
    Show answer & explanation

    Correct answer: A — Human-retained: Claude drafts talking points and likely questions beforehand, and the manager holds the conversation in person.

    • A. Correct. Telling a long-serving employee their role is eliminated is high-stakes and cannot be undone, and it needs empathy and personal accountability. The step stays human-retained, and Claude only prepares material beforehand.
    • B. Incorrect. Reviewing the draft improves the wording, but the employee still learns of the elimination from a written message rather than from a person. A human check on the text does not supply the presence and accountability this news needs.
    • C. Incorrect. This is AI-executed delivery of irreversible, life-changing news with no human review before it is sent. A follow-up meeting the next day comes after the employee has already read it and cannot undo how they received it.
    • D. Incorrect. Claude would speak with the employee about their dismissal in real time, and the manager only sees the transcript that evening. Oversight comes after the harm is done, so this is not meaningful human review for a sensitive conversation.

    4.1: End-to-End Deliverable Workflows

    23.An analyst uses Claude's research feature to compare how three regulators treat data retention. The comparison goes into a board paper, and the analyst will check the findings before submitting. Which two instructions in the research request make the findings easier to verify? Choose two.(Select 2)

    1. A.Label each statement as either stated directly by a cited source or as Claude's own inference drawn from several sources.
    2. B.Write in a confident tone without hedging, so the comparison reads consistently across all three regulators.
    3. C.Mark which claims come from the regulators' own pages and which come from news or commentary about those regulators.
    4. D.Cite only pages published in the last month, so the comparison is guaranteed to reflect the most current rules.
    5. E.Leave out any claim supported by a single source, even when that source is the regulator's own published rule.
    Show answer & explanation

    Correct answers: A, C — Label each statement as either stated directly by a cited source or as Claude's own inference drawn from several sources.; Mark which claims come from the regulators' own pages and which come from news or commentary about those regulators.

    • A. Correct. Separating what a source says from what Claude inferred tells the analyst which statements to check against a page and which to reason about for themselves.
    • B. Wrong. A confident tone with no hedging hides where evidence is thin, so it makes verification harder and could present uncertain statements as settled.
    • C. Correct. Knowing which claims rest on the regulator's own pages and which on commentary shows the analyst where to start checking and which findings carry the strongest evidence.
    • D. Wrong. A one-month publication window would exclude regulators' rules that were published earlier and remain in force, so it would remove the most relevant sources instead of improving checks.
    • E. Wrong. A regulator's own rule is a strong primary source on its own. Dropping single-source claims would delete valid findings and does not help the analyst check what remains.

    4.4: Capability Communication

    24.A communications team uploads its brand style guide and three approved press releases to a Claude Project, then asks Claude to draft a new product announcement. The head of comms will tell the executive team what to expect from the first drafts. Which statement is most realistic?

    1. A.Drafts will match the brand voice closely because of the uploaded examples, so executives can expect to publish them after a brief proofread for typos.
    2. B.Drafts will ignore the uploaded style guide until the team retrains Claude on it, so executives should expect generic text that staff must rewrite completely.
    3. C.Drafts will start closer to the house style thanks to the guide and examples, so executives should expect staff to check facts and claims before release.
    4. D.Drafts will follow the guide for the first few announcements and then drift, so executives should expect to re-upload the guide before every new draft.
    Show answer & explanation

    Correct answer: C — Drafts will start closer to the house style thanks to the guide and examples, so executives should expect staff to check facts and claims before release.

    • A. Incorrect: examples and a style guide improve voice, but they do not make claims or product details accurate, and a typo-only proofread would miss wrong facts.
    • B. Incorrect: Claude can use uploaded style material in a Project when drafting without any retraining, so expecting fully generic text understates it.
    • C. Correct: supplying the guide and examples moves drafts closer to the house style, while a person still verifies facts and claims before anything is published.
    • D. Incorrect: nothing in the setup makes the guide stop applying after a few drafts, and re-uploading it before every draft is not needed. Claiming this invents a limitation.

    4.2: Delegation and Workflow Redesign

    25.An HR team's leave workflow has Claude approving or denying each request and emailing the employee at once. Decisions depend on team coverage that only the manager knows, and a denial can make an employee miss a booked event, which an appeal cannot undo. Which redesign best reduces over-delegation while keeping Claude's time savings?

    1. A.Claude decides each request using the leave policy and a shared team calendar, and managers get a weekly summary of the decisions sent.
    2. B.Claude decides each request using the leave policy, and a second Claude session rechecks the denials against the policy before emails go out.
    3. C.Claude decides each request using the leave policy, and the email says a denial can be appealed to the manager within seven days.
    4. D.Claude checks each request against the leave policy and recommends a decision with reasons, and the manager decides and sends the email.
    Show answer & explanation

    Correct answer: D — Claude checks each request against the leave policy and recommends a decision with reasons, and the manager decides and sends the email.

    • A. A calendar shows who is booked off, but the coverage judgment rests with the manager, and Claude would still send decisions on its own. A weekly summary reaches managers after the emails are out, so a wrong denial that costs someone a booked event has already happened.
    • B. A second Claude session only repeats a policy check. It has no access to the coverage knowledge that only the manager holds, and no accountable person weighs in before a denial that cannot be undone is emailed.
    • C. An appeal window opens only after the employee has received the denial. A denial can already have caused a missed event that an appeal cannot undo, so a human decision has to come before the email, not after.
    • D. Claude keeps the routine policy check and the reasoning, which preserves most of the time savings. The coverage judgment and the final decision move to the manager, who holds the missing knowledge and acts before any irreversible email is sent.

    5: Configuration and Knowledge Management

    5.1: Project Configuration

    26.An operations manager has a Project for monthly vendor reviews with instructions and a standing scorecard. This week only, she must review one unusual supplier contract that will not be needed again. Where should the contract go?

    1. A.Package the contract's terms as a Skill, so Claude can load them when a review starts in the Project
    2. B.Attach the contract to the chat where the review happens, so Claude can read it while answering there
    3. C.Add the contract's key clauses to the Project instructions, so Claude applies them in its Project replies
    4. D.Upload the contract to the Project knowledge, so Claude can search it when answering in the Project
    Show answer & explanation

    Correct answer: B — Attach the contract to the chat where the review happens, so Claude can read it while answering there

    • A. Incorrect. A Skill packages a reusable procedure that Claude loads when a matching task comes up. One supplier's contract is a single document, not a repeatable procedure, and it will not be needed after this review.
    • B. Correct. Attaching the file to the chat where the review happens gives Claude the contract for that conversation only. The Project's instructions and knowledge stay limited to the material the monthly reviews reuse.
    • C. Incorrect. Project instructions set standing behavior for the Project's replies. Clauses from one supplier's contract are reference material, and placing them there would shape later monthly reviews that have nothing to do with that supplier.
    • D. Incorrect. Project knowledge persists and can be searched in later chats of the Project. Storing a contract that will not be needed again leaves it there for future reviews, where it adds clutter and can be pulled into unrelated answers.

    5.2: Connector and Knowledge Management

    27.A consultant has one signed vendor agreement, final and never to be edited, and needs Claude to help answer questions about its termination clauses during a single review. The agreement sits in a Drive folder that also holds many unrelated client files. Which approach fits the need best?

    1. A.Connect Google Drive and ask Claude to search the folder holding the agreement, then pull the termination clauses from the live Drive file.
    2. B.Upload the signed PDF to the chat or a Project, then ask Claude to answer the termination questions from the text of that uploaded copy.
    3. C.Paste the termination clauses into a new Google Doc and connect Drive, then have Claude re-read that doc as its text is revised.
    4. D.Connect Google Drive and Gmail, then have Claude look for the agreement across both and answer from whichever copy it finds first.
    Show answer & explanation

    Correct answer: B — Upload the signed PDF to the chat or a Project, then ask Claude to answer the termination questions from the text of that uploaded copy.

    • A. Incorrect: a Drive connector works against the folder, so the many unrelated client files become searchable for a task that needs one document. A live link also adds nothing for a file that never changes.
    • B. Correct: a single, finished document needed for one review is best supplied as an upload. It needs no live connection and gives Claude only the one file, leaving the other client files in the folder out of reach.
    • C. Incorrect: the agreement is final and will not be edited, so a Doc that Claude re-reads after changes gives no benefit. It also adds a copy step and a Drive connection when a plain upload would do.
    • D. Incorrect: the agreement is one known file, so linking Gmail opens a whole mailbox to a task that needs a single document. Answering from whichever copy turns up first can also draw on a draft instead of the signed final version.

    5.3: Persistent-Instruction Design

    28.A marketing Project's instructions say both 'Always keep every answer under 100 words' and 'Always include a full step-by-step explanation of the reasoning'. Teammates report that some answers are terse and others run long, with no pattern. What is the best correction?

    1. A.Add a third instruction asking Claude to balance brevity and explanation consistently, and keep both existing rules in the instructions unchanged.
    2. B.Remove the 100-word limit from the instructions, and keep the explanation rule as the single standing guidance on how long replies should be.
    3. C.Move the 100-word limit into Project knowledge, and keep the explanation rule in the instructions so each rule is stored in a separate place.
    4. D.Replace both rules with one instruction that makes brief replies the default, and asks for a worked explanation only when the user requests reasoning.
    Show answer & explanation

    Correct answer: D — Replace both rules with one instruction that makes brief replies the default, and asks for a worked explanation only when the user requests reasoning.

    • A. Incorrect: the two contradictory standing rules stay in place, and a general request to balance them does not say which one wins for a given request. Claude keeps guessing, so answers stay uneven.
    • B. Incorrect: this discards the brevity requirement the team wrote instead of reconciling it with the explanation rule. Answers would become uniformly long rather than short or detailed according to the type of request.
    • C. Incorrect: Project knowledge holds reference material, not standing behavior rules, and both rules would still apply to the same requests. Changing where a rule is stored does not tell Claude which one takes precedence.
    • D. Correct: with two unconditional rules that cannot both be met, Claude has to guess which one to follow, which produces the uneven answers. One rule that states when each behavior applies removes the contradiction and keeps both goals.

    5.4: Configuration Lifecycle and Degradation

    29.A finance team shares a Claude Project that drafts client payment-terms emails. Six months ago the Project instructions were written to say "offer net-30 terms and a 2% early-payment discount." The company has since moved to net-45 with no discount, and the Project knowledge files were updated to match. Drafts now mix the two policies. The team lead must attribute the problem to the right layer. Which layer is most likely responsible?

    1. A.The Project memory, which keeps summaries of earlier chats where net-30 and the discount came up and reuses them in new drafts
    2. B.The knowledge files, which Claude retrieves as excerpts from several documents and blends into one policy summary for each draft
    3. C.The Project instructions, which were written for the net-30 policy and were not revised when the knowledge files were refreshed
    4. D.The underlying Claude model, which leans on net-30 conventions from its training data and overrides the supplied files in drafts
    Show answer & explanation

    Correct answer: C — The Project instructions, which were written for the net-30 policy and were not revised when the knowledge files were refreshed

    • A. Incorrect. Nothing in the scenario points to remembered conversations as the source of the old terms. The net-30 and 2% discount wording sits in the Project instructions, which are applied to each draft the team generates, so memory is not needed to explain the mix.
    • B. Incorrect. The knowledge files were updated to match the net-45, no-discount policy, so excerpts retrieved from them and blended together would still describe the current terms. They cannot be the source of a 2% discount or net-30 wording.
    • C. Correct. The instructions are the one layer that was never updated, so they still tell Claude to offer net-30 and a 2% early-payment discount while the knowledge files describe net-45 with no discount. Claude combines the two sources, and editing the instructions to match current policy removes the conflict.
    • D. Incorrect. The drafts mix the company's own two specific policies, including a 2% early-payment discount, which matches the text in the Project instructions rather than general training-data conventions. The model follows supplied context, so two configuration layers disagreeing explains the symptom.

    5.4: Configuration Lifecycle and Degradation

    30.A product manager notices that Claude's competitor-comparison drafts began citing a pricing tier that was retired in the spring. The manager's Project instructions are accurate, and the Project's knowledge files contain no pricing for competitors. Memory is turned on for the account. Which step is the most direct way to test whether accumulated memory is the cause?

    1. A.Tell Claude in the chat to disregard its remembered competitor details and tier information, then rerun the same request
    2. B.Open the saved memory entries and delete or correct the outdated competitor pricing note, then rerun the same request
    3. C.Rewrite the Project instructions from scratch and add the current competitor pricing tiers, then rerun the same request
    4. D.Upload the latest competitor price sheets and tier summaries to the Project knowledge files, then rerun the same request
    Show answer & explanation

    Correct answer: B — Open the saved memory entries and delete or correct the outdated competitor pricing note, then rerun the same request

    • A. Incorrect. A chat message only adds a request-level instruction. The stored memory entries stay untouched, so the result cannot isolate memory as the source, and following the instruction cannot be verified.
    • B. Correct. Memory can keep a fact after it stops being true. Opening the entries, fixing or removing the outdated note and rerunning the identical request directly confirms or rules out memory as the cause.
    • C. Incorrect. The Project instructions are already accurate and contain no competitor pricing, so they are not implicated. Rewriting them leaves any stored memory entry in place and tests nothing about memory.
    • D. Incorrect. The knowledge files hold no competitor pricing, so that layer is not the source. Adding documents may push the drafts toward newer figures, but it masks the problem and does not show whether memory caused it.

    6: Responsible Use, Data Protection, and Governance

    6.3: Policy, Ethics, and Diligence

    31.An internal review of AI-use incidents sorts each observation into 'policy gap' or 'training gap'. Which two observations most clearly point to a policy gap rather than a training gap? Choose two.(Select 2)

    1. A.Staff who completed the latest training used Claude to transcribe client calls, and no section of the policy mentions or restricts transcription tools
    2. B.Two departments applied the AI-disclosure rule in opposite ways, and the policy wording on when notice is required supports each department's reading
    3. C.Employees hired after the last training session pasted customer names into Claude, and the policy's data-handling section prohibits sharing customer data with AI tools
    4. D.Contractors who skipped the policy briefing used an unapproved AI workspace, and the policy's tool register lists that workspace as prohibited
    5. E.About a third of staff could not recall the data-handling rule six months after onboarding, and the policy sets out that rule in clear, unambiguous terms
    Show answer & explanation

    Correct answers: A, B — Staff who completed the latest training used Claude to transcribe client calls, and no section of the policy mentions or restricts transcription tools; Two departments applied the AI-disclosure rule in opposite ways, and the policy wording on when notice is required supports each department's reading

    • A. Correct. Staff who completed the training still used Claude for transcription, and the policy says nothing about that use. The written rules have a hole, so better training on the existing text could not have prevented it.
    • B. Correct. Two departments each defended an opposite reading with the same policy text, so the wording on when notice is required is ambiguous. Clarifying the policy fixes this; more training on it would not.
    • C. Incorrect. The policy already prohibits sharing customer data with AI tools. The employees joined after the last session and never learned the rule, so this is a training gap.
    • D. Incorrect. The policy already lists the workspace as prohibited. The contractors skipped the briefing and never learned the restriction, so this is a training and onboarding gap, not a gap in the written rules.
    • E. Incorrect. The rule is written clearly, and staff simply forgot it over six months. That points to refresher training, not to changing the policy.

    6.3: Policy, Ethics, and Diligence

    32.Usage review at a media company shows that every policy violation involving unapproved Claude accounts comes from freelance contractors. The policy is clear, and employees who completed the annual AI training show no violations. Contractors are not included in policy distribution or the training schedule. What is the most appropriate corrective action?

    1. A.Send the AI usage policy to contractors and enrol them in the training, with a signed acknowledgement required before they receive company material
    2. B.Prohibit contractors from using AI tools on company projects and limit the approved accounts to employees, with the rule written into their contracts
    3. C.Add stricter penalties for unapproved accounts to the policy and apply them to contractors, with repeat violations escalated to their managers
    4. D.Rewrite the policy in plainer language and publish it on the staff intranet, with a one-page summary and worked examples for each team
    Show answer & explanation

    Correct answer: A — Send the AI usage policy to contractors and enrol them in the training, with a signed acknowledgement required before they receive company material

    • A. Correct. Employees who receive the policy and training comply, and contractors are the only group never given either. Giving contractors the same policy and training, with a recorded acknowledgement, closes that gap.
    • B. Incorrect. Contractors have broken the policy because it never reached them, not because they cannot follow it. A prohibition removes a legitimate tool when the same policy and training that worked for employees can be extended to them.
    • C. Incorrect. The policy is already clear, and penalties do not fix the fact that contractors were never told its rules. Punishing people who were never given the policy treats a coverage gap as an enforcement problem.
    • D. Incorrect. Employees comply with the current wording, so complexity is not the cause. Publishing a clearer version on the staff intranet also fails to reach contractors, who are outside policy distribution.

    6.1: Use-Case Appropriateness

    33.An office manager wants to enable a third-party Skill for her team. Which sequence of steps best prepares her to adopt it safely?

    1. A.Identify the Skill's publisher and download source → Read the Skill's instructions and bundled scripts → Check the files and connections it reaches → Trial the Skill on non-sensitive sample files → Roll out to team with a usage note
    2. B.Read the Skill's instructions and bundled scripts → Identify the Skill's publisher and download source → Check the files and connections it reaches → Trial the Skill on non-sensitive sample files → Roll out to team with a usage note
    3. C.Identify the Skill's publisher and download source → Check the files and connections it reaches → Read the Skill's instructions and bundled scripts → Trial the Skill on non-sensitive sample files → Roll out to team with a usage note
    4. D.Identify the Skill's publisher and download source → Read the Skill's instructions and bundled scripts → Trial the Skill on non-sensitive sample files → Check the files and connections it reaches → Roll out to team with a usage note
    Show answer & explanation

    Correct answer: A — Identify the Skill's publisher and download source → Read the Skill's instructions and bundled scripts → Check the files and connections it reaches → Trial the Skill on non-sensitive sample files → Roll out to team with a usage note

    • A. Establishing who published the Skill comes first, then reading its contents reveals what it does, which lets its reach be compared with the task. Only then is it trialled on harmless sample files before the team gets it.
    • B. Reading the instructions and scripts before identifying the publisher means the contents are judged without knowing who wrote them or where they came from. Provenance is the baseline for deciding how much trust the contents deserve, so it should come first.
    • C. Checking what the Skill can reach before reading its instructions and scripts means the reach is assessed without knowing what the Skill actually does. Its behavior has to be known before its access can be compared with what the task needs.
    • D. Trialling the Skill on sample files before checking what files, data and connections it can reach means the test runs without knowing its exposure. Even sample files could be touched or sent somewhere unexpected, so reach should be checked before any run.

    6.1: Use-Case Appropriateness

    34.An analyst connects an email connector that can read and send mail so Claude can summarize supplier messages. One message from an unknown sender contains hidden text telling the assistant to forward the latest price list to an outside address. The analyst only wants summaries. Which setup is most appropriate?

    1. A.Add a keyword filter that flags messages containing phrases like “forward” or “send to”, and leave the connector's send access on.
    2. B.Limit the connector to read access, and require a person's approval before Claude can send anything on the analyst's behalf.
    3. C.Add a system prompt rule telling Claude to treat message bodies as data and not follow instructions in them, and leave send access on.
    4. D.Turn on logging of sent mail so the analyst can review the sent folder each morning and recall suspicious messages, and keep send access on.
    Show answer & explanation

    Correct answer: B — Limit the connector to read access, and require a person's approval before Claude can send anything on the analyst's behalf.

    • A. Incorrect. A keyword filter only catches the wording its author anticipated, and hidden text can be phrased or encoded differently. Any message that slips through still reaches a connector that can send mail, and the analyst's task never required that capability.
    • B. Correct. Summarizing needs only read access, so removing send capability takes away the means to act on injected text. Requiring a person's approval for sends keeps a human in the loop if sending is ever needed.
    • C. Incorrect. A system prompt rule is a soft control, and injected text can still persuade the model to act on it. Leaving send access on keeps a capability the summarizing task does not need, so one successful injection could still send the price list out.
    • D. Incorrect. Reviewing sent mail the next morning detects a leak only after it happened. A price list already delivered to an outside address cannot be reliably recalled, so this does not prevent the exposure.

    6.2: Data Protection in Practice

    35.A strategy analyst is preparing a week-long competitor analysis in Claude. The material is classified Internal, and company policy permits Claude for Internal information. Memory is currently on in the analyst's account. The analyst does not want the analysis details stored in memory, but needs to reopen the same conversation on later days to continue it. Which setup fits?

    1. A.Run the analysis in an incognito chat and reopen that same chat from the chat list each morning, keeping the details out of memory.
    2. B.Open a new incognito chat each morning and ask Claude to continue from the earlier days' incognito chats, keeping the details out of memory.
    3. C.Run the analysis in a regular saved chat with memory on and reopen it each morning, then delete the related memory entries once the week is done.
    4. D.Switch memory off in settings and run the analysis in a regular saved chat, reopening it from the chat list each morning to continue the work.
    Show answer & explanation

    Correct answer: D — Switch memory off in settings and run the analysis in a regular saved chat, reopening it from the chat list each morning to continue the work.

    • A. Incognito chats are not saved to the chat list, so there is no entry to reopen on later days. The analyst needs the same conversation to continue, so this setup cannot work.
    • B. Each incognito chat is separate and ends without being kept, so Claude has nothing from earlier days to continue from. The analysis would restart each morning instead of continuing the same conversation.
    • C. With memory on, the analysis details are stored in memory as the week goes on, and deleting entries afterwards does not undo that. The analyst does not want the details stored in memory at all.
    • D. With memory off, no analysis details are stored in memory. A regular saved chat remains in the chat list, so the analyst can reopen the same conversation each day, and Internal data is within company policy.

    Want the full experience?

    These are just samples. Practice the full Anthropic Claude Certified Associate – Foundations (CCAO-F) - December 7 onwards question bank in quiz mode — free, no signup, with domain practice and exam simulation.