Anthropic free practice

Free CCAO-F practice questions

21 real CCAO-F sample questions, each with an explanation of why every option is right or wrong. No account, no card. This is the reasoning the CCAO-F tests: knowing why the tempting answer is wrong, not just spotting the right one.

The real CCAO-F is 60 questions in 120 minutes, pass mark 720 / 1000. For a domain-by-domain breakdown and a study plan, read the CCAO-F study guide. The full bank has 261 questions.

Output Evaluation and Validation (21% of the exam)

Free sampleOutput Evaluation and Validationmedium

A communications officer at a local council has asked Claude to draft a press release announcing a new food-waste collection service. It goes to regional newspapers in one hour, so there is not enough time to check every sentence against the approved scheme paper. Which TWO parts of the draft should she verify first against an authoritative source? Select TWO.

  • AThe start date of collections and the number of households covered, checked against the approved scheme paper Correct
  • BThe opening paragraph's general claim that food waste is a large share of what households throw away
  • CThe quotation attributed to the council's environment lead, checked against the statement that person approved Correct
  • DThe closing sentence encouraging residents to take part and to tell their neighbours about the scheme
  • EThe tone and reading level of the release, so that it suits a general newspaper readership
When time is short, verify the specific names, numbers, dates and quotations in a draft first, because errors there cause the most harm. Claude can produce fluent text that contains a wrong date, figure or quotation, and those specific claims are what readers act on and what is hardest to retract once published. Checking them against the approved primary document gives the most protection for the limited time available, while general background and stylistic choices carry far less risk.

Why A is correct: Dates and figures are specific, checkable claims that readers and residents will act on, and an error would need a public correction, so they are the first priority when time is short.

Why B is wrong: It is tempting because it is a factual statement, but it is general background that carries little consequence if loosely worded, so it ranks below the specific dates, figures and quotations.

Why C is correct: Words attributed to a named person are high-stakes because Claude can invent or reshape a quotation, and publishing a misquote damages trust with the press and the person quoted.

Why D is wrong: Closing lines are visible and feel important, but this one makes no factual claim, so there is nothing in it to verify against a source.

Why E is wrong: Polishing the style is a sensible editing task, but it is not verification and leaves any wrong date, figure or quotation in place.

Free sampleOutput Evaluation and Validationmedium

A training coordinator in a hospital trust's admin team uses Claude with web search to find the deadline for this year's mandatory fire safety refresher. Claude gives a date and cites a news article and a staff forum post. The trust publishes its own training policy on its intranet. What should she treat as the authoritative source for the deadline?

  • AThe trust's current published training policy, since it is the primary document that actually sets the deadline. Correct
  • BThe news article Claude cited, since it is the most recent public source and comes from an established outlet.
  • CA second web search, since the date is confirmed if several more websites repeat the same deadline independently.
  • DThe answer from Claude's most capable model, since stronger reasoning makes it the most reliable on dates.
Verify claims against the primary source that actually governs the fact, not against secondary reports or another AI answer. An authoritative source is the one that creates or officially records the fact, here the trust's own current training policy. News articles, forum posts and repeated web results are secondary and can be stale or wrong, and no choice of model makes Claude's answer a substitute for checking the primary document.

Why A is correct: Correct. The organisation's own current policy is the primary source for its internal deadline, so Claude's answer and the cited secondary sources should be checked against it before the date is shared with staff.

Why B is wrong: Tempting because recency and a reputable outlet suggest reliability. It is wrong because a news report is a secondary account; it does not set the trust's deadline and may be out of date or misreported.

Why C is wrong: Tempting because agreement across sites feels like corroboration. It is wrong because secondary sites often copy each other, so repetition does not make them authoritative over the document that sets the rule.

Why D is wrong: Tempting because the most capable model handles complex reasoning better. It is wrong because model choice does not turn an answer into a primary source; the date still has to be checked against the trust's own policy.

Free sampleOutput Evaluation and Validationmedium

A procurement support analyst in a hospital trust's admin team used Claude to compare three supplier tenders for cleaning services. Claude produced a table of prices and delivery terms and, when asked, said it was 95 percent confident the table was accurate. The analyst planned to quote that rating in the covering paper. The procurement committee will use the table to award the contract. What should the analyst do before submitting it? Select TWO.

  • AAsk Claude to list the figures it feels least sure about and check those entries against the tenders
  • BCheck each price and delivery term in the table against the suppliers' original tender documents Correct
  • CRun the same comparison again in a new chat and submit the table once both versions agree
  • DDrop the 95 percent rating from the covering paper, since it gives the committee no evidence of accuracy Correct
  • EReformat the table with clearer headings and totals so the committee can compare suppliers quickly
Claude's self-rated confidence is not evidence of accuracy, so verify figures against the source documents and do not present the rating as assurance. When Claude states a confidence percentage it is generating a plausible-sounding answer, not measuring how often its table matches the tenders. The only way to know the prices and terms are right is to compare them with the original tender documents, and passing the rating on to decision makers would replace that evidence with a number that means little.

Why A is wrong: It looks efficient because it targets checking effort, but it relies on Claude's own sense of uncertainty, which does not reliably identify where its errors are.

Why B is correct: The tenders are the authoritative source, and a contract award rests on these figures, so comparing them directly is the verification that actually establishes accuracy.

Why C is wrong: Agreement between two runs feels like confirmation, but a second Claude pass is not independent evidence and can repeat the same misreading of a tender.

Why D is correct: A self-reported confidence figure is not an accuracy measure, and quoting it could give the committee false assurance about a decision with financial consequences.

Why E is wrong: A clearer layout helps the committee read the table, but it does nothing to confirm the figures and may make wrong numbers look more trustworthy.

Workflow Integration and Solution Design (16% of the exam)

Free sampleWorkflow Integration and Solution Designmedium

An operations lead at a logistics company must turn notes from three depot managers into requirements for a new parcel-returns process by Friday. The managers' notes disagree on who approves a return and how quickly refunds are paid. She plans to use Claude to analyse the notes before writing anything up. Which actions should she take? Select TWO.

  • AAsk Claude to list the conflicts and gaps across the three sets of notes, then confirm each with the managers Correct
  • BAsk Claude to settle each disagreement by picking the most sensible option and send that to the managers
  • CAsk Claude to restate each requirement with the note it came from, so she can check it against the original Correct
  • DAsk Claude to rate how complete the requirements are and treat a high score as ready for sign-off
  • EAsk the technical team to build an automated requirements tool before she analyses any of the notes
When analysing requirements with Claude, use it to surface conflicts and trace each requirement to its source, while stakeholders settle the decisions. Claude adds value in requirements work by comparing many inputs at speed and exposing contradictions, gaps and assumptions. It cannot know which competing requirement the business wants, so the stakeholders must resolve conflicts. Linking each restated requirement to its source note makes the output checkable, which is a stronger safeguard than any rating Claude gives itself.

Why A is correct: Correct. Claude is good at spotting contradictions and missing details across several documents, but only the managers can settle what the process should be, so surfacing the conflicts and confirming them with the owners is the right use.

Why B is wrong: Tempting under a Friday deadline, but it hands a business decision to Claude. The disagreements are about who approves returns and refund timing, which the managers own, so presenting Claude's choices as settled skips the stakeholders who must agree them.

Why C is correct: Correct. Tying every restated requirement to its source lets her verify quickly that nothing was invented, dropped or reworded in a way that changes its meaning, which matters when the notes already conflict.

Why D is wrong: A self-reported score feels like a quality check, but Claude rating its own output is not evidence of completeness. Only checking against the notes and the managers' answers shows whether the requirements are right.

Why E is wrong: It sounds like a scalable fix, but this is a one-off analysis of three sets of notes with a Friday deadline. A well-structured chat in the Claude apps handles it, so escalating to a build is unnecessary and too slow.

Free sampleWorkflow Integration and Solution Designmedium

A customer service team lead at an energy retailer asked Claude, 'How can we make our complaints process faster?' and received generic advice about chatbots and staff training. Complaints currently pass through seven steps, and the regulator requires a written acknowledgement within two working days. The lead wants changes the team can act on this quarter. What should the lead do next?

  • AGive Claude the seven steps with their average times and the regulator's rule, and ask which steps add delay Correct
  • BRegenerate the same question a few times and combine the most useful suggestions from each answer
  • CSwitch to the most capable model and ask the same question again for a more detailed answer
  • DAsk Claude to research industry best practice and adopt the top-ranked complaints process in full
To find inefficiencies in a process, give Claude the actual steps, measurements and constraints rather than a general question. Claude can only analyse the process it is shown. A general question produces general advice, whereas the step list, average times and the acknowledgement rule give it something concrete to compare and test proposed changes against. The lead then judges which suggested changes are workable and compliant before acting.

Why A is correct: The advice was generic because the prompt contained no detail about the actual process. Supplying the real steps, their timings and the regulatory constraint lets Claude point to specific bottlenecks the team can change without breaching the two-day rule.

Why B is wrong: Collecting several answers can feel like widening the options. The prompt is unchanged, so each answer is still generic advice about complaints in general, not about this team's seven steps.

Why C is wrong: A more capable model can reason more deeply, so this seems like an upgrade. The problem is missing context, not reasoning power, and a stronger model given the same vague question still cannot see where this process loses time.

Why D is wrong: Looking at how others do it is a sensible input. Adopting an outside process wholesale skips the analysis of the team's own steps and hands the design decision to a ranking instead of to the lead's judgment about the regulator's rule.

Free sampleWorkflow Integration and Solution Designmedium

A policy officer at a local council used Claude's research feature to prepare a briefing on how five neighbouring councils run food waste collections. The briefing will be quoted in a public committee paper next week, and its recommendation rests on two cost-per-household figures that Claude attributed to council reports in its citations. What should the officer do before the briefing goes to the committee?

  • AAsk Claude to review the briefing again and rate how confident it is in each figure
  • BReformat the briefing into a one-page summary and label it as AI-assisted research
  • CAsk Claude to add further citations for each figure so the evidence looks stronger
  • DOpen the cited council reports and confirm the two cost figures match the source text Correct
When research output will inform a public decision, verify the figures it rests on against the cited primary sources before sharing. Research with citations makes checking faster, but a citation only shows where Claude says a fact came from, not that the source says it. Self-rated confidence and extra citations are produced by the same process and add no independent evidence. Opening the source for the figures the recommendation depends on is the human verification step that keeps the officer accountable for what the committee reads.

Why A is wrong: A second look by the same tool feels like a quality check, and a confidence rating sounds reassuring. It is not independent evidence, though: Claude's stated confidence does not tell the officer whether the figures actually appear in the council reports.

Why B is wrong: Being open about AI involvement and making the paper easy to read are both reasonable. Neither checks the two figures the recommendation depends on, so a wrong number would still reach a public committee paper.

Why C is wrong: More citations can look like more rigour, which is why this tempts. Listing extra sources without opening them does not confirm that any of them contains the figures, and it adds more unverified claims to check.

Why D is correct: The recommendation rests on two specific figures and the paper will be public, so the officer should check those figures against the primary sources Claude cited. Opening the reports is the step that turns a plausible citation into a verified one.

Governance, Risk, and Responsible Use (15% of the exam)

Free sampleGovernance, Risk, and Responsible Usemedium

An HR coordinator at a logistics company has 300 applications for warehouse supervisor roles and a shortlist due in three days. The company's recruitment policy states that every shortlisting decision must be made by a named recruiter. The hiring manager suggests asking Claude to rank the applicants and reject the bottom half automatically. What should the coordinator do?

  • AUse Claude to summarise each application against the published criteria, and have the recruiter make every shortlisting decision Correct
  • BLet Claude reject the bottom half, because the published criteria are objective and it applies them more consistently
  • CAsk Claude to rank the applicants and write a reason for each rejection, then send the rejections as drafted
  • DDrop Claude from the process entirely, since any use of AI in recruitment breaches the company's recruitment policy
Claude can prepare and summarise information for a hiring decision, but the decision about each candidate stays with an accountable person. Shortlisting is a decision about a person, so it needs a human who can be held accountable and who can spot when the output misreads someone. Summarising applications against stated criteria is drafting and analysis work that Claude does well, and it speeds up the recruiter without transferring the decision. The policy fixes who decides, which separates an appropriate supporting use from an inappropriate automated one.

Why A is correct: This uses Claude for what it suits, condensing 300 applications against the criteria so the recruiter can work faster, while the shortlisting decision about each person stays with the named recruiter as the policy requires.

Why B is wrong: Consistency sounds like fairness, and the deadline makes automation tempting. But a rejection is a decision about a person, the policy requires a named recruiter to make it, and Claude can misread or unevenly weigh applications in ways nobody would catch if no human looks.

Why C is wrong: Written reasons look like accountability, so this can feel more responsible than a bare ranking. Claude is still making the decision, the reasons are generated after the fact rather than checked, and no recruiter has decided anything as the policy requires.

Why D is wrong: Caution about AI in hiring is sensible, so stepping back can feel safe. But the policy governs who makes the decision, not whether a recruiter may use help to read applications, so abandoning Claude over-corrects and puts the deadline at risk for no gain.

Free sampleGovernance, Risk, and Responsible Usemedium

A training manager in a software firm's customer support department must produce a generic onboarding module on the team's ticket escalation process by the end of the month. The firm's AI-use policy allows Claude for drafting internal material provided a subject-matter expert reviews it before release. The module contains no personal data. A colleague says AI should not be used for anything involving staff. What should the manager do?

  • AWrite the module by hand, because training material that staff will rely on is too important to draft with an AI tool
  • BUse Claude to draft the module from the existing process guide, and have a subject-matter expert review it before release Correct
  • CUse Claude to draft the module and release it straight away, since generic content with no personal data needs no review
  • DAsk the technical team to build an automated tool that writes and updates every onboarding module for the department
Recognise an appropriate use of Claude and apply the review it requires, rather than refusing it over a general worry about AI. Drafting internal, non-personal material with human review is the kind of work Claude suits, and here the firm's policy names it as permitted on that condition. Telling appropriate from inappropriate use depends on the task, the data and the policy, not on a blanket rule about anything touching staff. The review step is what makes the draft safe to release, so it is kept rather than skipped.

Why A is wrong: Taking the colleague's concern seriously feels prudent. But the policy permits exactly this use with review, so refusing it over-corrects and costs time against the deadline without reducing any real risk.

Why B is correct: Drafting generic training material is a use the policy explicitly permits, the content holds no personal data, and the expert review the policy requires catches any step Claude has described wrongly before staff rely on it.

Why C is wrong: The absence of personal data does make the task low-risk, which is why skipping review is tempting. But the policy makes expert review a condition of use, and an unchecked draft could teach new staff an escalation step that is wrong.

Why D is wrong: Automation can look like the scalable answer to recurring training work. But this is one module that a manager can draft in the Claude apps, so escalating to developers adds cost and delay for a need a reviewed draft already meets.

Free sampleGovernance, Risk, and Responsible Usemedium

An administrator on a hospital trust's outpatient booking team uses the trust's approved Claude workspace to draft replies to the team's general inbox. A patient emails asking whether they should halve their medication dose before next week's appointment because of side effects. The team's guidance says administrators book, move and cancel appointments and pass any clinical question to the clinic's nursing staff. What should the administrator do?

  • AAsk Claude to draft a careful answer about the dose, adding a line that suggests the patient confirms it with their GP
  • BAsk Claude to research the medication's side effects and send the patient a summary with the cited sources attached
  • CUse Claude to draft a reply acknowledging the email and saying it has been passed to the nursing staff, then forward it to them Correct
  • DReply that the team cannot help with this, and suggest the patient looks up advice about the medication online
Use Claude for administrative drafting, and route any medical decision about a person to a qualified professional instead of answering it. Whether a patient should change a dose is a medical decision about an individual, which is an inappropriate use for Claude and outside the administrator's role whatever the output says. Claude remains useful for the part of the task that is administrative: a clear acknowledgement and a handover. Hedging or citing sources does not turn clinical advice into an appropriate use, because the problem is who decides, not how carefully it is worded.

Why A is wrong: A cautious answer with a referral line can feel helpful and safe. But it still gives medical advice about one patient's treatment from someone not qualified to give it, and the patient may act on the dose advice before seeing anyone.

Why B is wrong: Cited sources look like the verification step that makes output trustworthy. But general information about a drug is not advice on this patient's dose, and sending it still oversteps the administrator's role on a clinical question.

Why C is correct: Drafting an administrative reply is an appropriate use, and routing the clinical question to the nursing staff follows the team's guidance so a qualified person decides about the patient's medication.

Why D is wrong: Declining to give clinical advice is the right instinct. But pointing the patient to the internet abandons them, ignores the guidance to pass clinical questions to the nursing staff, and leaves the side effects unreviewed by anyone qualified.

Prompting and Task Execution (14% of the exam)

Free samplePrompting and Task Executioneasy

An HR adviser at a logistics company asks Claude to 'Summarise this parental leave policy' and attaches the policy. The summary is accurate but full of legal phrasing and long sentences. It will be posted in the staff room for warehouse employees, many of whom read English as a second language. What should the adviser change in the prompt?

  • AAsk for a more detailed summary so that no part of the policy is left out
  • BSwitch to the most capable model so the summary is written more clearly
  • CState who will read it and ask for plain English with short sentences Correct
  • DRun the same prompt again and choose whichever version reads best
When an accurate response is pitched wrongly, state the intended audience and the reading level they need in the prompt. Claude infers tone and complexity from whatever the prompt gives it; with no audience named, it tends to mirror the source material, here a legal policy. Telling Claude who will read the output and how plainly it must be written changes the register directly, which a different model or a rerun of the same prompt cannot do.

Why A is wrong: Completeness feels like the safe choice for a policy document. But the problem is readability for this audience, and adding more detail would make the summary longer and harder for these readers, not easier.

Why B is wrong: A more capable model sounds like a fix for any quality problem. But no model can guess an audience the prompt leaves out, so this costs more without addressing the missing information.

Why C is correct: The prompt never said who the readers were, so Claude matched the register of the source policy. Naming the audience and the reading level gives Claude the information it needs to pitch the language correctly.

Why D is wrong: Regenerating can produce a slightly different draft, which makes it tempting. With the same prompt Claude still has no reason to change its register, so the adviser is relying on luck instead of a clear instruction.

Free samplePrompting and Task Executioneasy

A project officer at a local council asks Claude to 'Write a progress update on the Mill Lane resurfacing scheme for councillors.' The draft looks professional but includes completion dates and costs that do not match the project, because Claude had no information about it. Councillors will use the update to approve the next stage of funding. What should the officer do?

  • AProvide the latest progress report and ask Claude to use only those facts Correct
  • BAsk Claude to search the web for details of the Mill Lane scheme first
  • CAsk Claude to word the figures more cautiously so they sound less certain
  • DAdd a note asking councillors to treat any figures as rough estimates
When a task depends on specific facts, give Claude the source material and ask it to work only from that material. Without source material, Claude produces text that fits the request in shape, including realistic-looking dates and costs that it has no way of knowing. Attaching the authoritative document and restricting Claude to it changes the job from inventing plausible content to summarising known facts, which is what a funding decision needs.

Why A is correct: Claude filled the gaps with plausible but invented details because it had no source material. Supplying the actual progress report and limiting Claude to it grounds the update in real figures, which matters when councillors will approve funding on it.

Why B is wrong: Web search does bring in outside information, so it looks like a way to fill the gap. But current internal project figures are unlikely to be published, and the officer already holds the authoritative report.

Why C is wrong: Softer wording can feel like a responsible hedge. But the figures are still made up, so councillors would be deciding on wrong numbers that merely sound tentative.

Why D is wrong: A caveat seems transparent. But it passes the problem to the decision makers instead of fixing it, and the officer has the real figures available to use.

Free samplePrompting and Task Executioneasy

A marketing officer at an animal welfare charity asks Claude to 'Write a fundraising email for our winter appeal.' The draft is 400 words and leans on guilt-heavy phrases. The charity's style guide sets a 150-word limit for appeal emails and bans guilt-based language, and every email is checked against it before sending. What is the best change to the prompt?

  • AAsk Claude to make the email more persuasive and emotionally moving
  • BAsk Claude to rate how closely its draft follows fundraising best practice
  • CKeep the prompt as it is and cut the draft down by hand each time
  • DAdd the word limit and the ban on guilt-based language as constraints Correct
When output must meet known rules such as a word limit or banned language, state those rules in the prompt as explicit constraints. Claude cannot follow an organisation's style guide it has not been shown, so it falls back on general conventions for the genre. Writing the limits and prohibitions into the prompt turns unstated expectations into instructions, which is far more reliable than asking for a vaguer quality like persuasiveness or fixing each draft afterwards.

Why A is wrong: Fundraising emails should persuade, so this sounds on target. But it would push the draft further towards emotional pressure, which is exactly what the style guide forbids, and it says nothing about length.

Why B is wrong: A self-assessment can feel like a quality check. But Claude has not been told this charity's rules, so its rating measures generic practice, and a self-rating is not evidence that the draft meets the style guide.

Why C is wrong: Hand-editing does get a compliant email out. But it repeats the same rework on every appeal because the prompt still lacks the rules, when stating them once would prevent the problem.

Why D is correct: The draft broke two specific rules that the prompt never mentioned. Stating the 150-word limit and the language ban as explicit constraints lets Claude write within the style guide from the first draft.

Product and Model Selection (12% of the exam)

Free sampleProduct and Model Selectionmedium

An HR adviser at a logistics company answers about 30 questions a week from line managers about leave and absence rules. Every answer must follow the 120-page staff handbook and use a plain, neutral tone for managers who are not HR specialists. The handbook is reissued each quarter with changes. Which TWO actions should the adviser take? Select TWO.

  • ACreate a Project whose instructions set the plain, neutral tone and whose knowledge holds the handbook Correct
  • BPaste the full handbook text at the start of each new chat so every answer starts from the same rules
  • CWhen the handbook is reissued, replace the old file in project knowledge with the new version Correct
  • DPaste the whole handbook into the project instructions so Claude reads every rule before answering
  • EKeep each quarter's handbook in project knowledge side by side so Claude can see how rules changed
For recurring work, put behaviour in project instructions, reference documents in project knowledge, and replace outdated knowledge files rather than adding contradicting versions. A Project separates how Claude should behave (instructions) from what it should draw on (knowledge), and both persist across every chat in the Project. That suits a steady stream of similar questions answered against one document. Because the handbook changes quarterly, accuracy depends on knowledge holding only the current version, since two conflicting files give Claude two answers to choose from.

Why A is correct: This is the recurring, high-volume work a Project is designed for: the instructions carry how Claude should write for every answer, and the handbook sits in project knowledge as the reference material Claude draws on in each chat.

Why B is wrong: It is tempting because it does put the rules in front of Claude. With 30 questions a week it repeats the same setup every time and invites mistakes when an older copy is pasted, which is the problem a Project removes.

Why C is correct: The quarterly reissue is the deciding fact. Swapping the outdated file for the current one keeps a single source of truth, so Claude cannot draw on a superseded rule when answering a manager.

Why D is wrong: It sounds thorough to put the rules where Claude always sees them. Instructions are for behaviour such as tone, audience and format; a long reference document belongs in project knowledge, and cramming it into instructions buries the tone guidance.

Why E is wrong: Keeping history feels careful. Leaving contradictory versions together means Claude may quote a withdrawn rule to a manager; the outdated file should be replaced, with any history kept outside the Project.

Free sampleProduct and Model Selectioneasy

A marketing manager at a software firm has two Claude tasks this week. The first is a one-off positioning analysis for the leadership team that weighs three long competitor reports with conflicting figures and reasons through several pricing scenarios. The second is drafting 200 short variants of a social media post, where speed and cost matter more than depth. She currently uses one model for everything. Which approach best fits these two tasks?

  • AUse the most capable model for both tasks so the quality of every output is as high as it can be
  • BUse the most capable model for the positioning analysis and the fastest model for the post variants Correct
  • CUse the fastest, lowest-cost model for both tasks so the team's AI spend stays as low as possible
  • DUse the balanced middle model for both tasks as a compromise between output quality and spend
Choose the model type per task, matching complex reasoning to Opus and simple high-volume drafting to Haiku, rather than using one model for everything. Model choice is a per-task decision, not a single account-wide setting. Complex analysis with conflicting sources benefits from the most capable model, while short, repetitive drafting gains little from it and costs more, so splitting the work by need gives better results for the same effort.

Why A is wrong: Defaulting to the top model feels like a quality guarantee. For 200 short post variants it adds cost and wait time without a meaningful gain, which is exactly what the stem says matters for that task.

Why B is correct: The two tasks have opposite needs. Opus suits the complex, multi-step analysis for leadership, and Haiku suits the high-volume short drafts where the stem says speed and cost matter more than depth.

Why C is wrong: Keeping spend down is a real concern, so this is tempting. The positioning analysis involves conflicting figures and multi-step reasoning for leadership, where the lowest-cost model is the most likely to miss connections.

Why D is wrong: A single compromise sounds tidy and avoids switching. It under-serves the complex analysis and over-spends on the simple variants, when each task has a clearly better fit.

Free sampleProduct and Model Selectioneasy

A communications officer at a local council uses Claude throughout the day for routine but varied drafting: newsletter items for residents, short press releases and replies to councillors' queries. The work needs good writing and some judgement about tone, but none of it involves deep, multi-step analysis, and she wants drafts back quickly between meetings. Which model type should she choose as her everyday default?

  • AOpus, because public-facing council writing deserves the most capable model on every draft
  • BHaiku, because quick turnaround matters more than quality for routine council writing
  • CSonnet, because it balances capability with speed for varied everyday drafting work like hers Correct
  • DOpus for drafting and Haiku for checking, so a second model confirms each draft is accurate
Sonnet is a sensible everyday default for varied work that needs good quality and quick turnaround but not deep, multi-step reasoning. Sonnet sits between Haiku and Opus, offering strong writing and judgement while staying responsive. For a steady stream of varied, moderately demanding drafts, that balance avoids both the extra cost and wait of the top model and the quality risk of the fastest one.

Why A is wrong: Public-facing writing feels important enough to justify the top model. The stem rules out deep multi-step analysis and asks for quick drafts, so Opus adds cost and wait time she does not need as a daily default.

Why B is wrong: Speed is one of her needs, which makes the fastest model appealing. The stem also says the work needs good writing and judgement about tone, so trading quality away for speed is the wrong balance.

Why C is correct: Her work needs solid writing quality and tone judgement, plus quick turnaround, without deep analysis. Sonnet is the model type that balances capability and speed, which matches that mix.

Why D is wrong: A second pass sounds like a safety step. Another model reviewing the draft is not independent evidence of accuracy, and pairing models adds cost and delay for routine drafting she wants back quickly.

Configuration and Knowledge Management (12% of the exam)

Free sampleConfiguration and Knowledge Managementmedium

A customer service team lead at a local council drafts replies to residents' questions about bin collections, parking permits and council tax. For three weeks she has pasted each new query into one long conversation, and recent drafts have started quoting a parking rule she corrected in the first week. The council's current policy documents and reply style apply to each query. What should she do?

  • AKeep the long conversation going and restate the corrected parking rule above each new query she pastes in.
  • BSet up a Project holding the policy documents and reply-style instructions, then start a fresh chat per query. Correct
  • CAsk Claude in the current conversation to list the rules it is applying, then check that list for mistakes.
  • DTurn on Memory so the corrected parking rule carries over, then carry on using the same long conversation.
For recurring work with stable context, a Project gives each fresh chat the same instructions and documents instead of relying on one ever-growing conversation. Very long conversations can lose or blur earlier detail, so a correction made weeks ago may stop being applied. A Project stores the instructions and reference documents once and supplies them to every chat inside it, so each query starts clean with the same, current context.

Why A is wrong: Restating the correction is tempting because it fixes the symptom she noticed. It leaves her in a conversation long enough to blur earlier detail, so other corrections and the reply style can drift in the same way.

Why B is correct: This is recurring work with stable context. A Project gives each new chat the same instructions and the current policy documents from the start, so no single conversation grows long enough to lose earlier detail.

Why C is wrong: Reviewing Claude's own list feels like a verification step, but it asks the same degraded conversation to report on itself. It does nothing to stop the next drafts drifting and adds a manual check to each query.

Why D is wrong: Memory can carry some context between chats, which makes this sound like a fix. It is not where reference policy belongs, and keeping one very long conversation is the cause of the problem rather than a cure for it.

Free sampleConfiguration and Knowledge Managementmedium

A marketing coordinator at a homeware retailer set up a Project for product descriptions. She pasted the full 35-page brand guide into the project instructions, followed by four lines saying who the descriptions are for and how they must be formatted. Drafts often miss those formatting lines, and colleagues find the instructions hard to edit. The team consults the brand guide section by section. What change should she make?

  • ASplit the brand guide across three new Projects so each one has a shorter set of instructions to follow.
  • BReplace the brand guide in the instructions with a one-paragraph summary that Claude writes of it.
  • CMove the brand guide into the project knowledge and keep the instructions to audience, tone and format. Correct
  • DLeave the instructions as they are and repeat the four formatting lines at the start of each new chat.
Project instructions should say how Claude behaves, while long reference documents belong in project knowledge where Claude can draw on them. Instructions and knowledge do different jobs. Instructions set role, audience, tone and format for every chat, so they work best when short and clear; knowledge holds reference material Claude draws on as needed. Burying four behaviour lines under a long document makes them easy to miss and hard to maintain.

Why A is wrong: Shorter instructions sound like progress, but splitting the work across three Projects scatters one recurring task and still puts reference material in the instructions. The team would have to guess which Project to use.

Why B is wrong: This shortens the instructions, which is tempting. It throws away the detail the team consults section by section, so drafts lose access to the specific brand rules they need.

Why C is correct: Long reference documents belong in project knowledge, where Claude can draw on the relevant sections. Short instructions about audience, tone and format are easier to edit and stand out instead of being buried in 35 pages.

Why D is wrong: Repeating the lines may improve formatting, but it reintroduces the copy-and-paste routine a Project exists to remove and leaves the instructions just as hard for colleagues to maintain.

Free sampleConfiguration and Knowledge Managementmedium

The head of geography at a secondary school uses a Project to build lesson resources. At the start of each chat she types the same paragraph: pitch it for Year 9, use a starter, main activity and plenary structure, and include one idea for pupils who need extra support. Two colleagues are about to start using the same shared Project and should get resources in that format without being told. What should she change?

  • APut the paragraph into the project instructions so it shapes each chat in the Project for whoever uses it. Correct
  • BUpload the paragraph to the project knowledge as a short document titled "House style".
  • CSave the paragraph to her Memory so Claude applies it whenever she builds lesson resources.
  • DKeep the paragraph in a shared department document for each teacher to paste in at the start.
Standing directions about audience, structure and format belong in project instructions, which apply to every chat in a shared Project. Project instructions are the Project's standing brief: they apply to each chat started inside it, whoever starts it. Memory follows an individual user and knowledge holds reference documents, so neither guarantees that colleagues in a shared Project get the same format without being told.

Why A is correct: Project instructions set how Claude behaves in every chat in the Project, including audience and structure. Because they belong to the shared Project, her colleagues get the same format without typing anything.

Why B is wrong: Putting it in the Project is the right instinct, but knowledge is reference material Claude draws on when relevant. A standing rule about audience and structure belongs in the instructions, which shape each chat.

Why C is wrong: Memory can carry her preferences between her own chats, which makes this tempting. It is tied to her account, so her two colleagues would not get the format.

Why D is wrong: This shares the wording, but it keeps the paste-every-time routine that a Project exists to remove, and relies on each colleague remembering to do it.

Troubleshooting and Optimization (10% of the exam)

Free sampleTroubleshooting and Optimizationmedium

A marketing coordinator at a regional chain of garden centres types 'Write a spring promotion email for our customers' into a new chat. The draft is fluent but features barbecues the chain does not sell and a 20 per cent discount nobody approved. The commercial manager has fixed the offer: 15 per cent off bedding plants for loyalty-card holders, valid for two weekends. What should the coordinator do next?

  • ARegenerate the same request several times and keep the draft that comes closest to the real offer.
  • BRewrite the request with the approved offer, product range, audience and tone, then check the draft against it. Correct
  • CSwitch to the most capable model, since a stronger model is less likely to invent product details.
  • DAsk Claude to review its own draft for invented details and confirm that every claim is accurate.
When Claude invents specifics, the usual cause is missing context, so supply the real facts, audience and tone in the prompt and check the result against them. Claude produces plausible content to fill whatever the request leaves unspecified, so a short prompt with no offer details invites invented products and discounts. Giving the approved terms, product range, audience and tone turns guessing into drafting from facts, and checking the draft against the commercial manager's terms confirms nothing was altered.

Why A is wrong: Regenerating feels quick and sometimes produces a better draft by chance. It is wrong because the request still lacks the offer and product details, so every version is a guess and the coordinator would be choosing the least wrong one rather than fixing the cause.

Why B is correct: The invented barbecues and discount show that Claude filled gaps the prompt left open. Supplying the fixed offer, what the chain sells, who the email is for and the tone removes those gaps, and checking the draft against the approved terms protects the commercial manager's decision.

Why C is wrong: A more capable model is a natural reach when output disappoints. It is wrong because no model can know the chain's approved offer or stock range unless it is told; the problem is missing context, not reasoning power.

Why D is wrong: Asking for a self-check sounds responsible. It is wrong because Claude has no access to the approved offer, so it cannot tell which details are invented, and its confirmation is not independent evidence of accuracy.

Free sampleTroubleshooting and Optimizationmedium

An HR adviser at a logistics company pastes a two-page absence policy into a chat and asks Claude to 'make this better'. The result is longer and more formal than the original. The summary is for warehouse shift workers who will read it on their phones during breaks, and many of them speak English as a second language. Which change to the request is most likely to fix the problem?

  • ARepeat the request with stronger wording, such as 'make this much better and far more professional'.
  • BUpload the company's other HR policies so Claude has more background on how the organisation writes.
  • CState the audience, ask for plain language under 200 words, and list the key actions as short bullet points. Correct
  • DAsk Claude to produce five alternative versions so the adviser can choose the one that reads best.
A vague request such as 'make this better' should be replaced with a specific audience, length, reading level and format so Claude knows what good looks like. Words like 'better' leave Claude to decide the goal, and it often defaults to fuller, more formal writing. Stating who will read the text, how much they can read and in what format turns an ambiguous request into a measurable one. The fix targets the cause, an unclear instruction, instead of adding effort or volume.

Why A is wrong: Emphasis can feel like a way to signal dissatisfaction. It is wrong because it is still vague, and 'more professional' pushes the text further towards the formal style that already failed this audience.

Why B is wrong: More background sometimes helps, so this looks like adding context. It is wrong because the missing information is about the readers and the format, and formal policy documents would reinforce the long, formal style rather than correct it.

Why C is correct: The request never said who the summary was for or what 'better' meant, so Claude read it as more polished and complete. Naming the audience, the reading level, a length and a scannable format gives Claude a clear target that matches shift workers reading on phones.

Why D is wrong: Choosing from several drafts seems efficient. It is wrong because all five versions start from the same unclear request, so the adviser is likely to get five variations on the wrong target.

Free sampleTroubleshooting and Optimizationmedium

An operations officer at a local council runs a Project that answers staff questions about procurement rules. The procurement policy was revised last month and the quote threshold changed. The officer uploaded the revised policy to project knowledge but left the old version in place, and Claude now gives different thresholds in different chats. Staff rely on these answers before placing orders. What should the officer do?

  • ARemove the superseded policy from project knowledge, then test the Project with a threshold question. Correct
  • BAdd a project instruction telling Claude to prefer the newer document whenever the two disagree.
  • CTell staff to state the new threshold at the start of their chats so Claude has the correct figure.
  • DAsk Claude to save the new threshold to Memory so it carries the figure forward into future chats.
When a Project's knowledge document is revised, replace the outdated file rather than keeping two contradicting versions, then test the Project's answers. Claude draws on whatever is in project knowledge, so two versions of the same policy give it two equally available answers and the result varies between chats. Replacing the old file removes the contradiction at its source. A quick test question then confirms the Project gives the current threshold before staff act on it.

Why A is correct: The inconsistent answers come from two contradicting versions of the same policy in knowledge. Removing the outdated version leaves one source of truth, and testing with a threshold question confirms the Project now gives the current figure before staff rely on it.

Why B is wrong: An instruction looks like a quick way to settle the conflict without deleting anything. It is wrong because the contradicting file stays in knowledge, so Claude can still draw on the old threshold, and the officer is relying on an instruction to manage a problem that removing the file would eliminate.

Why C is wrong: This would correct individual answers, so it can seem practical. It is wrong because it pushes the fix onto every user, fails whenever someone forgets, and leaves the outdated file in the Project to mislead future chats.

Why D is wrong: Memory can carry context between chats, which makes this tempting. It is wrong because reference documents belong in project knowledge, Memory does not remove the contradicting file, and it is not a dependable way to keep a policy figure current for every user.

Want the full bank?

261 CCAO-F questions, every one with an explanation of why every option is right or wrong. No sign-up to start.

Practise CCAO-F free

Frequently asked questions

Are these CCAO-F practice questions free?

Yes. Every CCAO-F question on this page is free to read with no sign-up, and each one explains why the right answer is right and why every other option is wrong. The full bank of 261 questions is on Examworthy.

Do the questions explain why the wrong answers are wrong?

Yes, and that is the point. Each option, correct or not, has its own rationale, so you learn to rule out the tempting wrong answer, not just recognise the right one. That is the reasoning the CCAO-F tests.

Are these real CCAO-F exam questions?

No. These are original, blueprint-aligned practice questions written to the public Anthropic content outline. We never reproduce live exam items. They mirror the format and difficulty of the real exam.

How many questions are on the real CCAO-F?

The CCAO-F is 60 questions in 120 minutes, with a pass mark of 720 / 1000. For the full domain-by-domain breakdown and a study plan, read the study guide.

Examworthy is not affiliated with or endorsed by Anthropic. All questions are original, blueprint-aligned practice material. We never reproduce live exam items. CCAO-F and related marks belong to their respective owners.