Updated September 4, 2026. Checked against the Claude Certified Associate - Foundations Exam Guide, Version 1.0, effective July 2026, the Anthropic Partner Academy certification FAQ, and the Claude Help Center articles for Projects, Artifacts, and Research.
You have read the seven CCAO-F domains and still cannot picture the questions. Anthropic publishes three sample items in the official exam guide and, per the certification FAQ, retired the practice exam that lived on its previous platform. Three samples for a 60-item exam is not a rehearsal.
This article is the rehearsal: how a CCAO-F item is built, the four distractor families that recur, twelve original worked questions in weight order, and the mechanics of multiple-response items, pacing, and the 720 score.
How a CCAO-F item is actually built
Every CCAO-F item is a small workplace situation, not a definition check. The official samples share one shape: a person in a business role, a task already underway, a constraint sitting in the background, and a question asking for the most appropriate next action.
Four parts do the work.
- The actor and their role. A compliance analyst or a recruiter. The role tells you which consequences count.
- The trigger. Claude produced an output, a chat drifted, a file needs uploading.
- The binding constraint. Policy, contract, audience, deadline, cost, or data sensitivity. Usually one clause, and it decides the answer.
- The asked-for judgment. "Most appropriate," "best," or "first" means the key is the strongest option, not the only defensible one.
Domain 2 also covers output format. Claude produces an Artifact when content is "significant and self-contained, typically over 15 lines" and likely to be edited or reused, and an inline reply otherwise.
The four distractor families
The rationales in Section 8 of the exam guide reveal a repeatable pattern. Learn these four and you can eliminate two options before finishing the stem.
| Family | What it looks like | Why it fails |
|---|---|---|
| The confident non-fix | Reformat the text, or ask Claude to rate its own confidence | Changes presentation, not correctness |
| The over-correction | Abandon the task or do it all by hand | Throws away a workable path when a proportionate control exists |
| The verbal patch | Instruct the model not to retain the data | An instruction is not a policy control |
| The wrong altitude | Escalate a prompt problem to a developer or switch platforms | Solves a bigger problem than the one described |
Twelve worked CCAO-F questions, in blueprint weight order
Sit these cold. Cover the answers, allow two minutes each, and mark which domain you missed.
Output Evaluation and Validation - 21%
1. A grants officer asks Claude to summarize a 90-page funder report stored in project knowledge. The summary quotes a dollar figure and a page number, and is headed for a board packet. A. Ask Claude to restate the figure with more confidence. B. Open the cited page in the source report and confirm the figure. C. Ask for the summary in a cleaner format. D. Accept it, because the source file was in the Project.
Answer: B. D is the sharpest distractor: a file in project knowledge means Claude could reach the source, not that the figure was read correctly. A and C are confident non-fixes. Pattern: provenance is not verification.
2. An analyst has Claude compare three vendor contracts uploaded to a Project. The output is fluent, addresses all three, and never mentions the termination clause that appears in only one. A. Claude cannot read contracts; do it manually. B. Treat this as an incompleteness problem: re-prompt with an explicit clause checklist, then verify. C. Accept it, since nothing stated is false. D. Escalate to a Claude Developer for an integration.
Answer: B. C traps anyone screening only for false statements. Domain 2 tests accuracy and completeness, and this omission changes a vendor decision. Pattern: completeness is a separate axis from accuracy.
3. An HR partner has Claude summarize promotion candidates. The draft calls two candidates "assertive" and two "supportive," and the split follows gender lines in the source notes. A. Publish it; the wording came from the notes. B. Ask Claude to make the language more positive for everyone. C. Flag it as a possible bias artifact, test each characterization against the evidence, and route the decision to a human owner. D. Delete the adjectives and publish.
Answer: C. B and D edit the copy and leave the reasoning intact. A treats the source as exonerating when the source is where the bias entered. Pattern: bias checks belong before a consequential decision, not in the copy edit.
Workflow Integration and Solution Design - 16%
4. A marketing team drafts every campaign brief in a fresh chat, pasting the brand voice guide each time. Quality swings week to week. A. Write longer prompts. B. Move the voice guide and briefing steps into a shared Project with instructions and knowledge, so every brief starts from the same context. C. Rotate models weekly and keep the best result. D. Let one person do all the prompting.
Answer: B. D hides the variance behind one operator and dies when they leave. Pattern: variance from re-pasting context by hand is a configuration problem, not a prompting-effort one.
5. A team wants Claude in the monthly compliance report a regulator reads. Which design choice most determines whether the pilot is defensible? A. The fastest model, to shorten the cycle. B. A named human owner who reviews and signs before filing, with the review recorded. C. Instructions telling Claude never to make mistakes. D. Keeping the AI assistance undisclosed so the regulator does not object.
Answer: B. D adds a transparency failure to a governance one. Pattern: accountability lives in the workflow, never in the prompt.
Governance, Risk, and Responsible Use - 15%
6. A consultant's client contract forbids sending client data to third-party services without written approval. Approval exists for one tool, not for Claude. A. Use Claude anyway, since a human reviews the output. B. Paraphrase the client's data so it is technically different. C. Keep client data out of Claude until approval is obtained, and use Claude on non-client work such as structuring the deliverable outline. D. Get verbal permission from a junior client contact.
Answer: C. B is a verbal patch and D is authority shopping. C is the option candidates miss: it neither uses the data nor abandons the tool. Pattern: a contractual data restriction usually still leaves a compliant partial use.
7. A team lead asks whether Claude may draft performance reviews. What decides the answer first? A. Whether Claude writes well enough. B. Whether the organization's AI policy and employment rules permit AI in employment decisions, and what disclosure and human review they require. C. Whether reviews get shorter. D. Whether the lead has a paid plan.
Answer: B. A is the capability answer to a permission question. Pattern: appropriate-use items are settled by policy and consequence, not by what the tool can do.
Prompting and Task Execution - 14%
8. An operations manager prompts, "Write something about our Q3 numbers," and gets a generic draft. A. Add "be more detailed." B. Specify the audience, the decision it supports, which figures to use, the format, and the length. C. Ask three times and pick the best. D. Switch to the most capable model.
Answer: B. D is the expensive wrong answer that also shows up in Domain 3. Pattern: an under-specified prompt is repaired with constraints, not with effort or a bigger model.
9. A project manager needs 40 pages of meeting notes turned into a risk register with owners and dates. The first attempt blends risks, actions, and decisions. A. Re-ask with a longer prompt. B. Decompose: have Claude label each note line as risk, action, or decision; confirm the labels; then build the register from the confirmed risks. C. Put the notes in a Project and ask again. D. Summarize the notes first, then discard them.
Answer: B. C changes where the notes live, not how the task is structured, and D destroys the evidence you verify against. Pattern: decompose when the failure is category confusion.
Product and Model Selection - 12%
10. A communications lead needs a competitive landscape briefing with sources they can open and check before tomorrow's partner meeting. A. One chat message asking for the briefing from the model's own knowledge. B. Research, which runs multiple linked searches and returns citations the lead can verify. C. The fastest, lowest-cost model, because the meeting is tomorrow. D. A Project with no knowledge sources.
Answer: B. C is the trap: speed is a real constraint here, but the binding one is verifiable sourcing. Research runs on paid plans. Pattern: match the feature to the evidence the task requires, then optimize cost inside that choice.
Configuration and Knowledge Management - 12%
11. A Project's instructions cite the 2025 pricing sheet. Project knowledge holds both the 2025 and 2026 sheets. Sales drafts keep quoting last year's prices. A. Tell every user to remind Claude about the 2026 sheet in each chat. B. Update the project instructions to name the current sheet, and remove or clearly supersede the outdated file. C. Start a new chat each time. D. Upload the 2026 sheet again.
Answer: B. A is tempting: it works in the short run while pushing a system defect onto every user. Pattern: stale configuration is repaired at the configuration layer.
Troubleshooting and Optimization - 10%
12. A long chat that started well now contradicts decisions made earlier in the same thread and re-explains settled points. A. The model degraded; switch platforms. B. Summarize the decisions so far into a short brief, open a fresh chat or Project seeded with it, and continue. C. Tell Claude to try harder. D. Repeat the question until the answer improves.
Answer: B. Pattern: drift inside one long thread is a context-management symptom; the guide's Domain 3 objectives name restarting, summarizing, and persisting as the levers.
Multiple-response items: what "select two" changes
The exam guide is explicit that CCAO-F uses "multiple-choice and multiple-response items" and that "each item states how many responses to select." You never guess how many answers an item wants, and that instruction is itself information.
Use it three ways. The count is a constraint: if an item asks for two and you can defend only one, you have not finished reading. "Select two" usually tests a pair that works together, such as a control plus a verification step, so near-duplicate options are rarely both keyed. And the strongest single option is not automatically in the pair.
Anthropic does not publish whether these items award partial credit, so select exactly the number requested. Candidates on Reddit report "select two" and matching-style items on recent sittings.
Pacing 60 items in 120 minutes
Two minutes per item is the average, and the blueprint says where they go. The guide calls the weights approximate proportions of scored items, so this is planning arithmetic.
| Domain | Weight | Approx. items | Approx. minutes |
|---|---|---|---|
| Output Evaluation and Validation | 21% | 13 | 25 |
| Workflow Integration and Solution Design | 16% | 10 | 19 |
| Governance, Risk, and Responsible Use | 15% | 9 | 18 |
| Prompting and Task Execution | 14% | 8 | 17 |
| Product and Model Selection | 12% | 7 | 14 |
| Configuration and Knowledge Management | 12% | 7 | 14 |
| Troubleshooting and Optimization | 10% | 6 | 12 |
Passing candidates report about 105 minutes for a first pass and 15 minutes on flagged items. Strike out options as you eliminate them, flag anything past 150 seconds, and keep moving. Half the exam sits in evaluation, workflow, and governance, so a slow start in Domain 1 costs you time where the items are.
What 720 out of 1,000 does and does not mean
It is not a percentage. The exam guide states that CCAO-F is criterion-referenced: "each candidate is measured against a fixed performance standard, not against other candidates." The cut score came from a formal standard-setting study in which subject matter experts judged the performance expected of a minimally qualified candidate. Answering 72% of items correctly is not the same as scoring 720, and it is not a curve either.
Your score report shows percent-correct within each domain, but the guide is clear those figures "are not used to determine your pass or fail result, which is based on your total scaled score." Domain percentages are a study map for a retake, not a second hurdle. No practice-set percentage can promise a 720.
A readiness self-check
Score the twelve by domain, the way the real report does. Then check what a percentage cannot tell you.
- For every item you got right, can you name the binding constraint? If you picked what felt professional, you were pattern-matching.
- For every item you got wrong, which distractor family caught you? Most people have one repeat offender.
- Can you explain why the second-best option is defensible but weaker?
Do all three on fresh questions at a two-minute pace and you are ready to schedule.
Why "real exam questions" sets are the wrong trade
Several vendors sell CCAO-F sets described as taken from previous real exams. Section 13 of the exam guide puts "all exam content, including questions, answer options, and scenarios" under a confidentiality agreement, and Section 12 says disclosure "may result in invalidation of your result, revocation of your credential, and a ban from future exams."
The accuracy risk is practical too. One widely linked free CCAO-F question page advertises the exam as 65 questions in 130 minutes; the guide says 60 in 120. A set that gets the published format wrong is not one whose rationales you should trust.
Your next step
For the blueprint, study plan, and registration and renewal rules, use the CCAO-F exam guide for 2026. Still choosing? Compare all four in which Claude certification to take in 2026. Register through the Anthropic Partner Academy.
Buy on Amazon