How to pass Claude Certified Associate - Foundations (CCAO-F)
27 min read7 domains coveredFree practice, no sign-up
The Claude Certified Associate - Foundations (CCAO-F) is a foundational credential for people who use Claude as an everyday work tool rather than build software with it. It covers how to ask Claude for what you need, how to check what comes back before anyone relies on it, which feature and model suit a given job, how to set up Projects so recurring work starts from the right context, and how to use Claude responsibly with real organisational data.
It suits professionals in operations, marketing, project management, education, communications, HR and consulting who already use Claude in the apps and want to use it with more confidence and less rework. There is no coding and no technical set-up. The questions are short workplace scenarios: a council officer drafting a press release, a recruiter shortlisting applicants, a team lead setting up a shared Project. You are asked what a careful, accountable professional would do next.
The exam rewards judgement, not feature trivia. Several options in most questions describe something a reasonable person might try, and only one fits the situation as written. The heaviest theme is checking output: knowing that fluent, confident writing is no evidence of accuracy, and knowing which claims to verify first. Practise on scenario questions that explain every option, so you learn why the tempting answers fail as well as which one is right.
CCAO-F rewards the professional who checks before sharing, picks the right feature for the job, protects sensitive data before it goes in, and stays accountable for whatever Claude helped produce.
Difficulty
Foundational
Best for
Business professionals who use Claude in the Claude apps and set up Projects: operations, marketing, project management, education, communications, HR and consulting staff, with little or no technical background.
Prerequisites
None. Some hands-on time using Claude for real work tasks, including at least one Project, makes the scenarios feel familiar.
60
Questions
120 min
Time allowed
720 / 1000
Pass mark
$99
Exam cost (USD)
261
Practice questions
How this exam thinks
Four habits separate a pass from a fail on CCAO-F, and none of them is about knowing more features. The questions are workplace scenarios where several answers sound sensible. The right one is almost always the answer a careful, accountable professional would give.
First, verify before you share. Claude can state an invented citation, a wrong date or a misattributed quote as fluently as a true fact, so polish and confidence tell you nothing about accuracy. The exam's answer is to check the claim against an authoritative source: the approved policy, the original report, the supplier's own document, the person who is being quoted. Asking Claude to rate its own confidence, or to check its own work, is the classic distractor, because a self-rated score is not evidence. When time is short, check the specifics first: names, numbers, dates, quotations and anything going to an external or compliance audience.
Second, pick the feature and model that fit the task. A one-off question belongs in a plain chat; recurring work that reuses the same context belongs in a Project; a question that needs sources gathered and cited calls for research; content you will edit, reuse and share belongs in an artifact. Models follow the same logic: the fastest, lowest-cost model for simple high-volume work, the most capable model for complex reasoning, and the balanced middle model for varied everyday work. Reaching for the most powerful option every time is wrong, and so is reaching for a bigger model to fix a problem that is really missing information.
Third, protect sensitive data before it goes in. If personal or confidential information should not reach Claude, the control is to remove it, anonymise it or keep it out entirely, using the approved account and any approval process the policy requires. An instruction telling Claude not to retain or repeat the data is not a control, because the data has already been shared. Watch for indirect identifiers, free-text fields and a lookup key that turns codes back into names.
Fourth, escalate when the job outgrows the Claude apps, and only then. Most poor results are fixed inside the apps: a clearer prompt, the missing context, the current document, a Project, a different model. Escalation is right when the work needs to run unattended between business systems (a job for the technical team), or when the decision itself belongs to someone else, such as legal counsel, a safeguarding lead or a hiring manager. Escalating routine, low-risk work that you can check yourself is as wrong as failing to escalate a decision that is not yours to make.
What each domain tests and how to study it
The CCAO-F blueprint is split across 7 domains. Weights are the official share of the exam; see the official exam guide for the authoritative breakdown.
What you must be able to do. Diagnose which missing element of a prompt explains a weak response, match the prompting approach to the task type, and break large deliverables into steps you review before building on them.
In one sentenceGive Claude the goal, audience, source material, constraints and format it needs, split big jobs into reviewed steps, and fix a weak draft with specific feedback instead of asking again unchanged.
Recall check: answer these from memory first
Name the five elements of an effective request, and say which one is usually missing when an accurate draft is pitched at the wrong reader.
Explain why listing numbered steps in one prompt is not the same as running separate, reviewed steps.
Say what a good follow-up correction contains beyond the change you want.
For analysis, research, drafting and brainstorming, give the one thing each prompt should ask for.
What it tests. Whether you can write a request Claude can do well: stating the purpose and the decision the output supports, naming the audience and reading level, supplying the source material and asking Claude to work only from it, setting constraints such as word limits, and saying what format you want back. It tests breaking a long, multi-part deliverable into an outline and sections reviewed in turn, iterating with targeted feedback that says what to change and what to keep, and adapting the approach to analysis, research, drafting and brainstorming.
How to study it. Learn the five ingredients of a good request (goal, audience, context, constraints, format) well enough that, given a vague or off-target result, you can name the one that was missing. Then practise the follow-up: a useful correction states the change, the reason behind it and what must stay untouched. For decomposition, remember the order of operations: settle and check anything a later step depends on before the later step uses it. Finally, link each task type to its signature request: evidence and reasoning for analysis, scope and citations for research, tone and audience for drafting, range without ranking for brainstorming.
Easy to confuse
Regenerating the same request versus revising it with specific feedback. Regenerating an unchanged request repeats the same gap, so every new attempt misses in the same way. When the problem is missing information such as the audience or the funder's priority, the fix is to add it in a follow-up or a revised request.
Numbered steps in one prompt versus separate, reviewed steps. Numbered steps in one prompt only organise the reply; you see step one after steps two and three already rely on it. Separate steps let you check and correct early output, such as extracted figures, before anything is built on it.
Giving Claude a role versus giving Claude the facts. A role such as 'act as a fundraising expert' sets a perspective but supplies no knowledge of your organisation, so Claude fills the gaps with typical assumptions. Output that fits your situation needs your actual facts, figures and constraints.
A persuasive draft versus a balanced analysis. Asking Claude to make the case for one option produces advocacy, which is wrong while a decision is still open. When people have not yet decided, ask for a balanced comparison against the decision-makers' own criteria, drawbacks included.
Worked example from the CCAO-F bank
lock_openFree 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 sentencescheck_circle 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.
What you must be able to do. Check Claude's output for accuracy, completeness, fabrication and bias against authoritative sources, decide when it needs expert review, and choose the right format and shape for its audience.
In one sentenceThe most heavily weighted domain: confident is not correct, so verify the specifics against the real source, match review to the stakes, and shape the output for the reader who will use it.
Recall check: answer these from memory first
List the kinds of claim you would check first when there is no time to verify every sentence.
Explain why checking each sentence against the source confirms accuracy but cannot confirm completeness.
Name three situations where Claude's output needs review by someone with specific expertise before it is used.
Say when you would ask for an inline reply, an artifact, or a table.
What it tests. Whether you can tell a usable output from a plausible one. It covers checking facts against the source material and the request, spotting invented citations, statistics and quotes, internal contradictions and one-sided framing, and choosing which claims to verify first. It tests when output must go to a person with the right expertise before use, such as legal, financial, medical or compliance content, decisions about people and external publication. It also covers adapting a draft for its audience, comparing alternatives against that audience's needs, and choosing between an inline reply, an artifact and a table.
How to study it. Spend the most time here. Build a mental priority list for verification: quotations under a person's name, names, numbers, dates, citations and prices come first because errors there cause public harm. Learn the subtle failures as separate ideas: a summary can be accurate but incomplete, consistent but wrong, faithful to an uploaded file that is itself out of date, or correct figure by figure while comparing monthly and quarterly charges as equals. For review, practise matching it to the stakes: your own check for routine internal work, a specialist for regulated or public content. For formats, link the shape of the information to the reader's task.
Easy to confuse
Verifying against a source versus asking Claude to check itself. Only comparison with an authoritative source, such as the approved policy, the original report or the person quoted, is evidence. Claude's self-rated confidence, or a second pass where it confirms its own work, is not, and presenting that rating as assurance misleads the reader.
Accurate versus complete. Checking each statement in the output against the source tests accuracy only. Completeness is checked the other way round, from the source or the original request towards the output, item by item, so nothing that should be there is missing.
Internally consistent versus accurate. A document with no contradictions agrees with itself, which says nothing about whether it matches the ledger or the policy. Accuracy needs comparison with the source data, and a contradiction is resolved against the source, not by picking one version.
A second pair of eyes versus the right expert review. Review only counts when the reviewer can detect the kind of error the content risks. A proofread catches wording; a compliance claim needs a compliance specialist. A disclaimer about AI involvement replaces neither, because it flags possible errors without finding them.
Worked example from the CCAO-F bank
lock_openFree 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 papercheck_circle 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 approvedcheck_circle 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.
What you must be able to do. Match a task to the right Claude feature and model family, and choose correctly between starting a fresh chat, summarising, and persisting context in a Project or Memory.
In one sentenceChat for one-offs, Projects for recurring work, research for sourced facts, artifacts for reusable content; the fastest model for simple volume, the most capable for complex reasoning, and a fresh chat with a summary when a long one starts to blur.
Recall check: answer these from memory first
Match each of chat, Project, research and artifact to the kind of task it suits, in one line each.
Describe the trade-off between the fastest, the balanced and the most capable model families.
Say what to do when a long conversation starts mixing up earlier and current decisions.
Explain why upgrading the model does not fix an answer based on an outdated policy.
What it tests. Choosing the feature that fits the job: a plain chat, a Project, research or an artifact. It tests the relative profile of the Haiku, Sonnet and Opus families (fastest and lowest-cost, balanced, most capable) and matching them to cost, speed and quality needs, without version numbers or prices. It also covers context: why very long conversations lose earlier detail, why separate chats do not share what was agreed, and when to restart with a summary, persist information in a Project, or rely on Memory.
How to study it. Make two short decision tables and drill them until they are automatic. Feature: is this one-off or recurring, does it need outside sources, will the output be edited and reused? Model: is the work simple and high-volume, varied and everyday, or complex multi-step reasoning? Then learn the limits of each choice. A more capable model reasons better but does not know your private policies; model choice does not make sensitive data safe to upload; Memory carries context between chats but is not a store for reference documents or a guarantee that a fact is current.
Easy to confuse
Chat versus Project versus research versus artifact. Ask what the task needs. A one-off with its context supplied is a plain chat; recurring work reusing the same instructions and documents is a Project; gathering and citing outside sources is research; a standalone document to edit, reuse and share is an artifact. A document you already hold belongs in an upload, not research.
A faster, lower-cost model versus the most capable model. Simple, repetitive, high-volume work suits the fastest, lowest-cost model with human review as the quality check; complex reasoning with interacting constraints suits the most capable model, where slower replies are the expected trade-off. Defaulting to the top model for everything wastes time and cost.
Restarting with a summary versus persisting in a Project versus Memory. A fresh chat seeded with a short, checked summary of current decisions fixes a blurred long conversation; pasting the whole old transcript does not. Work that continues across sessions or people belongs in a Project's knowledge. Memory carries brief personal context between chats, not reference documents or a team standard.
A more capable model versus the missing information. A bigger model improves reasoning, not knowledge of your organisation. If Claude has not been given the returns policy, or repeats last year's figure from outdated reference material, the fix is to supply or replace the document, whichever model is used.
Worked example from the CCAO-F bank
lock_openFree 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 handbookcheck_circle 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 versioncheck_circle 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.
What you must be able to do. Use Claude to analyse needs, plan and design solutions, decide whether to augment a step or redesign a process, place human checkpoints where errors would be costly, and set realistic expectations with stakeholders.
In one sentenceClaude speeds up the language-heavy work in a process, people keep the decisions and the checks, and the checkpoint sits where an error would enter or become irreversible.
Recall check: answer these from memory first
Say when to augment a single step with Claude and when to redesign the whole process.
Name two places a human checkpoint belongs in a multi-step workflow, and who should own it.
List what a credible business case for Claude should include besides time saved.
Explain why a repeatable workflow that produces consistent output can still be inaccurate.
What it tests. Applying Claude to real business processes. It covers analysing requirements and use cases with Claude while stakeholders settle the decisions, using Claude for research, planning and finding inefficiencies when it is given the real steps and measurements, and iterating on prototypes, templates and drafts with real-user feedback. It tests whether to augment one slow step or redesign a process around Claude, where to place a human checkpoint and who should own it, and how to explain Claude's value and limits to colleagues with evidence rather than enthusiasm.
How to study it. Think like a process owner. For each scenario, ask where the time is actually lost, where errors would enter, and who is accountable for the result. Learn that Claude works only from what it is given: a written procedure shows how work should run, not where it really stalls, and requirements drafted without stakeholder input are plausible guesses. Practise placing checkpoints: just before an error becomes costly or before later steps build on extracted figures, owned by someone who checks facts against the source rather than how the draft reads. For stakeholder questions, prefer answers that pair time saved with the review that remains, backed by a pilot on real work.
Easy to confuse
Augmenting a step versus redesigning the workflow. When a working process has one slow, language-heavy step, augment that step and keep the existing checks. When the delay comes from handoffs between people, redesign around Claude, removing steps that only move or reformat information and keeping steps that control a risk.
Checking how a draft reads versus checking its facts. A checkpoint that only reads for tone and flow is not a control. An effective checkpoint compares the facts in Claude's output, such as prices, dates and figures, against the source record before the output is used.
Research versus working from your own material. Research gathers and cites outside sources. Analysing a client's interviews, a team's notes or an internal procedure means giving Claude that material directly; research will not find it and will fill the gap with general information.
An artifact prototype versus a connected system. An artifact is ideal for shaping and testing a form, template or calculator with real users. Connecting that design to business systems so it runs at scale is integration work for the technical team, not something the prototype becomes on its own.
Worked example from the CCAO-F bank
lock_openFree 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 managerscheck_circle 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 originalcheck_circle 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.
What you must be able to do. Set up and maintain a Project so every chat starts from the right instructions and current knowledge, choose between uploading and connecting a source, and keep shared Projects owned and up to date.
In one sentenceInstructions say how Claude should behave, knowledge holds the documents it draws on, outdated files are replaced rather than added to, and a shared Project needs a named owner.
Recall check: answer these from memory first
Say what belongs in project instructions and what belongs in project knowledge.
List the steps to take when a policy document in a Project's knowledge is revised.
Give one reason to upload a file and one reason to connect a live source instead.
Explain why a chat started outside a Project does not follow that Project's tone rules.
What it tests. Configuring Claude Projects for repeated work: what belongs in project instructions (audience, tone, format, boundaries) and what belongs in project knowledge (the authoritative reference documents). It tests uploading a file versus connecting a live source such as Google Drive or Gmail, including the access a connector grants and what a shared Project exposes to its members. It covers writing standing instructions with clear priorities and reasons, separating them from one-off requests, and maintaining a Project as policies and documents change.
How to study it. Fix the split first: instructions are behaviour, knowledge is reference material, and instructions cannot supply facts that are not in the knowledge. Then learn the maintenance habits the exam returns to: replace an outdated file instead of adding a second version, update any instructions that restate old details, test with questions whose answers changed, and correct decisions already made from wrong answers. Remember the scope rules: Project instructions and knowledge only shape chats inside that Project, decisions made in one chat do not become shared context, and everyone a Project is shared with can draw on its knowledge.
Easy to confuse
Project instructions versus project knowledge. Instructions tell Claude how to behave: audience, tone, format and boundaries. Knowledge holds the full documents Claude draws on, such as a brand guide or policy. Pasting a long reference document into the instructions, or pointing the instructions at a policy that was never added to knowledge, are the classic mistakes.
Standing instructions versus the one-off request. Guidance that should apply to every piece of recurring work goes in the project instructions. A requirement for one piece only, such as a single word limit or one urgent ask, goes in that request, so the standing instructions stay correct for everything else.
Uploading a file versus connecting a source. An uploaded file is a fixed copy that goes stale when the original changes; a connector keeps content current but grants Claude the same access as the connected account. Connect only what the task needs, and upload a stable document when connecting would grant access policy does not allow.
Replacing an outdated file versus adding the new version alongside it. Two contradicting versions in project knowledge produce inconsistent answers. Remove the superseded file so the current one is the only source, and keep the authoritative document rather than a summary that silently drops the detail answers depend on.
Worked example from the CCAO-F bank
lock_openFree 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.check_circle 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.
What you must be able to do. Tell appropriate uses of Claude from inappropriate ones, protect personal and confidential data before it is shared, follow the organisation's AI policy and approvals, and stay accountable for the output you use.
In one sentenceClaude drafts and summarises under human review; people make decisions about people; sensitive data is removed before it goes in; and the person who sends the output owns it.
Recall check: answer these from memory first
State the test that separates an appropriate use of Claude from an inappropriate one.
List four things to remove or check before uploading a spreadsheet that contains customer records.
Explain the difference between anonymised and pseudonymised data.
Say why disclosing AI involvement does not reduce your responsibility for the content.
What it tests. Responsible use in an organisation. It covers telling apart appropriate work, such as drafting, summarising and analysis with review, from inappropriate uses, such as unreviewed decisions about people or content that deceives. It tests handling personal, confidential and regulated data: anonymising before upload, removing indirect identifiers, keeping restricted material out entirely, and using the approved account. It covers following AI policies on approved tools, disclosure and approval processes, and the ethics of bias, fairness, transparency and accountability.
How to study it. Learn the central test for an appropriate use: a competent person reviews the output and makes any decision that affects someone, within policy. Then work through data protection as a sequence of practical moves: get any required approval first, delete identifying columns from a copy rather than hiding them, check free-text fields, generalise details that identify someone in a small group, and keep any lookup key outside the tool. Treat policy as written: an approved-tools list approves specific tools and accounts, an approval covers its stated purpose, and a client's contract can restrict AI use on top of your firm's approval.
Easy to confuse
Anonymising before upload versus telling Claude not to retain the data. An instruction shapes what Claude writes back, but the personal data has still been shared. The control is to remove or anonymise identifiers before anything goes in, or to keep that material out of Claude entirely if policy classes it that way.
Claude drafting versus Claude deciding about a person. Claude can summarise applications against published criteria or draft letters about decisions people have already made. It must not make the hiring, disciplinary or eligibility decision, and a sign-off that never examines the individual case is not meaningful review.
Anonymised versus pseudonymised data. Replacing names with codes while a key exists that links codes back to people is pseudonymisation, and the result is still personal data. It protects people in Claude only if the key stays outside the tool and the names are added back afterwards by a person.
Disclosing AI involvement versus reviewing the content. Disclosure and review are separate duties. A disclosure statement meets a transparency requirement, but it does not check the facts or move responsibility; the person who sends or signs off the output remains accountable for it.
Worked example from the CCAO-F bank
lock_openFree 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 decisioncheck_circle 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.
What you must be able to do. Work out why Claude's output is poor and apply the fix that addresses that cause, adjust the approach from feedback and results, and make recurring workflows faster without removing the checks that keep quality up.
In one sentenceFind the cause before changing anything: missing context, an unclear request, the wrong feature or model, outdated knowledge or an overlong chat each has its own fix, and quality checks are made faster, never removed.
Recall check: answer these from memory first
For invented details, generic output, an outdated answer and drift in a long chat, name the likely cause and fix of each.
Explain how to confirm that an adjustment to your request actually worked.
Say what you may remove from a recurring workflow to speed it up, and what you must keep.
What it tests. Diagnosing underperforming prompts and poor outputs: invented specifics from missing context, vague requests, stale Project knowledge, conflicting standing instructions, a task that needs research or a more capable model, and quality drift in a very long conversation. It tests using feedback from reviewers, recipients and results to change the request rather than repeating it, confirming a fix by rerunning the cases that failed, and optimising recurring work with Projects, templates, the right model for the volume and fewer manual steps.
How to study it. Build a cause-and-fix table and drill it: invented details mean missing facts; generic output means a vague request; an outdated answer means the knowledge needs replacing; failure on a complete, complex request means a more capable model; drift in a long chat means a fresh chat with a summary. Then learn the optimisation rule: remove copying, reformatting and re-explaining, keep the human check and make it faster. Finally, recognise the boundary of the apps: work that must run unattended between two business systems is a job for the technical team.
Easy to confuse
A better prompt versus escalating to a developer. Weak drafts, missing context and inconsistent tone are fixed inside the Claude apps with a clearer request, the right documents or a Project. Escalate to the technical team only when results show the job must run unattended between business systems, which a chat workflow cannot do.
A better prompt versus a more capable model. If Claude invents specifics or misses the audience, the request lacks information and a bigger model will not help. If simple tasks succeed but a complete request needing multi-step reasoning fails, the more capable model is the likely fix.
Speeding up a quality check versus removing it. Optimising a workflow means cutting manual copying and reformatting, or making a required check faster, for example by having Claude list each claim beside its source. Replacing the check with Claude's own assessment, or dropping it, is never the answer.
Personal Memory versus shared Project instructions. Memory is personal to one user, so it cannot hold a standard the whole team must follow. A layout or tone every colleague needs belongs in the instructions of a shared Project they all work in.
Worked example from the CCAO-F bank
lock_openFree 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.check_circle 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.
A study plan that works
Map the blueprint and set a date
Day 1
Read the official exam guide and the seven domains with their weights. Book a provisional exam date now: a fixed date turns open-ended reading into a plan. Note that output evaluation and validation carries the most weight, so it will get the most time.
Get hands-on with the Claude apps
Week 1
Use Claude for real tasks from your own job. Run one task in a plain chat, set up a Project with instructions and a knowledge document, try research on a question that needs sources, and ask for an artifact. Questions are far easier when you have seen each feature do its job.
Lock in prompting, features and models (Domains 1 and 3)
Week 1
Learn the five elements of a good request and the feature and model decision tables. Use the recall prompts in this guide: cover the summary, answer from memory, then reveal. Aim to say out loud which feature and model you would use for a task and why.
Go deep on checking output (Domain 2)
Week 2
This is the heaviest domain. Practise verifying the specifics first, telling accuracy from completeness and consistency, spotting invented citations and quotes, and matching review to the stakes. Work through scenario questions rather than reading alone.
Cover Projects, workflows and responsible use (Domains 4, 5 and 6)
Weeks 2-3
Learn the split between project instructions and knowledge, where checkpoints belong in a workflow, and the data protection moves: anonymise before upload, keep lookup keys out, follow the approved account and approval process. Then cover troubleshooting (Domain 7) as cause and fix pairs.
Practise on scenarios with every answer explained
Week 3
Move to full practice sets and read the explanation for every question, including the ones you got right. The tempting wrong answers repeat in families, such as asking Claude to check itself or relying on an instruction to protect data, so learn to name each family on sight.
Close weak domains and sit a timed mock
Week 4
Use your per-domain accuracy to drill the domains dragging you down rather than re-reading what you already know. Take at least one full timed mock to rehearse pacing, then review every missed question before you sit the real exam.
Know when you're ready
Readiness for CCAO-F is a score on questions you have not seen before, not a feeling that the material is familiar. If you already use Claude every day, this exam can feel easy on a first read, because every scenario looks like something you have done. That familiarity is the trap: the questions reward the careful habit, such as checking a quotation against the person who said it, over the quick one most people use at work.
The test is whether you can answer fresh scenario questions and explain why each tempting option fails. If you can say why asking Claude to rate its confidence is not verification, why an instruction not to retain data is not a control, and why a bigger model does not fix a missing policy document, you know the material. If you can only agree with an explanation once you read it, you do not yet.
Trust your measured per-domain accuracy over your instinct, and set the bar at clearing every domain comfortably across more than one session, with output evaluation solid because it carries the most weight. The practice bank explains why the right answer is right and every wrong one is wrong. Readiness scoring tells you when you are there. Not before.
Ready to put this into practice?
Free CCAO-F questions, every answer explained. No sign-up.
Read the last line of the question first. It tells you what is being asked, such as what to do first or which claim to check, so you read the scenario looking for the answer.
Check how many responses an item asks for. Some items need more than one answer, and the item states how many to select.
When an option asks Claude to check, rate or confirm its own work, treat it as suspect. The exam's verification answer points to an authoritative source outside Claude.
When an option protects data by telling Claude what not to do with it, treat it as suspect. The control is removing or anonymising the data before it goes in.
Prefer the proportionate answer. Escalating routine, low-risk work is as wrong as skipping review on legal, financial or public content, so match the response to the stakes.
Before choosing a bigger model or a new feature, ask whether the real problem is missing information, an outdated document or a vague request. The simpler fix is usually the right one.
Eliminate two options fast. Most questions have two clearly weaker choices; removing them leaves you comparing the two that a careful professional might actually consider.
Frequently asked questions
Is CCAO-F hard?
It is a foundational exam with no coding or technical set-up. The difficulty is that most options describe something a reasonable person might do, and only one reflects careful, accountable practice. Scenario practice that explains every option matters more than memorising feature names.
How long should I study for CCAO-F?
Regular Claude users are often ready in three to four weeks of focused study. Newer users should allow extra time to try Projects, research and artifacts on real tasks, because the scenarios assume you know what each feature is for.
Do I need technical or coding skills?
No. The exam is set in everyday business work in the Claude apps: writing requests, checking output, setting up Projects and using data responsibly. It does not cover programming or building software with Claude.
How does CCAO-F differ from the Claude Architect and Developer credentials?
CCAO-F is for non-developers who use Claude in the Claude apps for their own work. The Claude Certified Architect - Foundations (CCAR-F) and Claude Certified Developer - Foundations (CCDV-F) credentials cover building with Claude's developer platform, so they suit engineers and technical architects rather than business users.
What is the pass mark for CCAO-F?
The exam is scored on a scaled range and the published pass mark is in the facts panel above. Because scoring is scaled, your raw percentage and the scaled score are not the same thing; aim to clear every domain comfortably in practice rather than scraping a target.
Which domains should I focus on?
Output evaluation and validation carries the most weight, so checking Claude's work deserves the most time. The other domains are closer in size; governance and responsible use, and workflow integration, reward careful reading of who owns each decision.
Do I need to know model version numbers, prices or plan limits?
No. The exam tests the relative trade-off between the model families, fastest and lowest-cost, balanced, and most capable, and which suits a given task. Learn the profile of each family and the reasoning for choosing it.
Is CCAO-F worth it?
It is a practical credential for professionals who want to show they use Claude with good judgement: checking output before it is shared, choosing the right feature for the job and handling organisational data responsibly. Preparing for it also tends to cut rework in day-to-day use.
Examworthy is not affiliated with or endorsed by Anthropic. This guide is original study material based on the public exam blueprint. We never reproduce live exam items. CCAO-F and related marks belong to their respective owners.
