CCAO-F - Claude Certified Associate Foundations Mock Exams

Prepare for the Claude Certified Associate Foundations (CCAO-F) certification with scenario-based practice exams focused on using Claude effectively in everyday work. These questions are designed to help you apply concepts, evaluate decisions, and identify where you need more study—not just memorize terminology. Anthropic describes the certification as covering practical Claude use across prompting, output evaluation, product selection, workflows, knowledge management, responsible use, and troubleshooting.

Exam Blueprint & Weights

Domain 1 — Prompting and Task Execution (14%)

• A prompt is an executable spec: objective, context, task, output contract, and quality/uncertainty criteria — not a clever sentence.

• Separate governing instructions from reference material; state which source wins when materials disagree.

• Decompose complex work at dependency boundaries; validate extracted facts before synthesis or formatting.

• Iterate by changing one variable at a time and retesting on easy, normal, edge, and adversarial cases — never "make it better."

• Analysis prompts must separate evidence, inference, assumptions, counter evidence, and recommendation.

• Research prompts need scope, source standards, an evidence table, and triangulation — link count is not evidence quality.

• Drafting/transformation must preserve source facts and caveats while adapting form; mark any proposed addition.

• Brainstorming diverges before it converges: expand options first, then score against explicit criteria.

Decide by: does this prompt remove ambiguity, supply the necessary context, and define the output/quality bar without asking the model to silently invent anything?

This domain tests your ability to engineer clear, structured instructions that produce consistent, high-quality results on the first attempt.

• Core Prompt Structure: Applying role definitions, background context, specific task goals, output format rules, and explicit negative constraints (what not to do).

• Prompt Adaptation by Task Type: Tailoring strategy based on objective—such as analytical lenses for data review, tone guidelines for drafting, or open-ended parameters for brainstorming.

• Few-Shot Prompting: Utilizing gold-standard reference examples within prompts to guide model tone, format, and reasoning depth.

• Iterative Refinement: Systematically editing prompts when outputs fall short rather than submitting repetitive, vague instructions

Domain 2 — Output Evaluation and Validation (21%)

• Score outputs on multiple dimensions — accuracy, evidence, completeness, consistency, audience fit, format, risk — never one overall impression.

• Fact-check at the claim level: atomize, prioritize by consequence, verify against the correct authoritative source, and test that the citation actually entails the claim.

• Hallucinations hide behind precision — exact numbers, quotes, citations, links; trace every precise-looking detail back to its source.

• Reconcile calculations independently: normalize units/periods/definitions, recompute, and cross-foot totals and percentages.

• Completeness means requirement coverage, not word count; distinguish "not found," "not applicable," and "not assessed."

• Check bias by comparing treatment and evidence standards across groups — omission of a protected term does not mean neutrality.

• Human-review intensity scales with impact, uncertainty, and irreversibility; the reviewer must be qualified, not merely human.

• Pick the output format (inline, artifact, table, structured data) by the next downstream action, not by novelty.

Decide by: what happens if this specific claim is wrong, and has it been verified against the source that actually controls it?

As the largest single domain on the exam, this module tests your critical reasoning when assessing AI-generated outputs.

• Hallucination Detection: Spotting plausible but completely fabricated data, false citations, incorrect mathematical calculations, and invalid logical deductions.

• Fact-Checking & Source Verification: Cross-referencing AI statements against primary uploaded documents, enterprise knowledge repositories, and trusted external databases.

• Format & Presentation Review: Determining whether outputs match required structures (e.g., Markdown tables, JSON payloads, executive bullet summaries, or downloadable Artifacts).

• Identifying Missing Context & Bias: Recognizing implicit assumptions, omitted variables, or biased messaging in drafts generated for broad audiences.

• Human-in-the-Loop Determination: Knowing exactly when an AI draft is acceptable for publication versus when mandatory human oversight and subject-matter-expert signoff are required.

Domain 3 — Product and Model Selection (12%)

• Chat fits a one-off request; Projects fit persistent shared instructions/knowledge; research fits broad current multi-source synthesis; artifacts fit a substantial, iteratively-edited work product.

• Treat Haiku/Sonnet/Opus as capability tiers, not prestige: Haiku for fast/cheap high-volume work, Sonnet as the general-purpose balance, Opus for the hardest reasoning — never default to the biggest model.

• Optimize total workflow cost (model + review + rework + escalation), not per-request price alone.

• Context is a finite, curated working set; more history is not automatically better and stale turns create conflicting instructions.

• Restart for a new or contaminated topic; summarize to carry forward validated state in a long thread; persist stable rules/knowledge in a Project.

• Pilot the lightest plausible configuration first and escalate capability only when a defined quality threshold is missed.

• Design fallback for unavailable connectors/models: define what may proceed on cached/approved data and what must pause.

Decide by: what is the minimum feature-plus-model configuration that clears a defined quality bar at acceptable cost and latency?

Anthropic offers distinct model tiers and feature environments. This domain tests your ability to select the ideal setup for a given task.

Model Tier Alignment:

• Claude Haiku: Best suited for lightweight tasks requiring maximum speed and minimal cost (e.g., simple classification, fast text parsing).

• Claude Sonnet: The ideal balance of high intelligence, reasoning power, and execution speed for everyday enterprise tasks.

• Claude Opus: Designed for deep research, highly complex analytical challenges, and nuanced writing requirements.

• Feature Utilization: Knowing when to use chat conversations, persistent Projects, deep research tools, or dedicated Artifact windows for standalone content creation.

• Context Limit Management: Understanding token limits, knowing when to reset conversations, and summarizing long chat histories to maintain processing quality.

Domain 4 — Workflow Integration and Solution Design (16%)

• Start from a business outcome and a mapped current-state process, not from a feature looking for a use case.

• Requirements must be testable — users, triggers, inputs, source authority, acceptance criteria, exceptions, controls; "accurate and helpful" is not a requirement.

• Classify each step as automate, augment, or human-only based on error detectability, consequence, and where human judgment adds irreplaceable value.

• Redesign around the real bottleneck; preserve system-of-record ownership and audit trail rather than inserting Claude at every handoff.

• Place human review where it can still prevent harm, show reviewers the evidence and uncertainty, and feed corrections back into the workflow.

• Pilot with a baseline, representative and edge cases, and explicit stop/scale gates before wider rollout.

• Report a balanced scorecard — value, quality, risk, adoption — never usage volume alone.

• Treat the workflow as a maintained product: named owner, versioned instructions/knowledge, regression tests, rollback, retirement plan.

Decide by: does this step need to be automated, augmented, or kept human-only given how easy an error is to detect and how much it would cost?

This module evaluates your ability to seamlessly incorporate Claude into existing organizational processes.

• Task Decomposition: Breaking complex business initiatives (such as a full competitive analysis or product launch plan) into logical, sequential AI prompts.

• Process Augmentation: Identifying which steps in a workflow should be automated by AI (e.g., summarizing research, drafting outlines) versus steps that must remain strictly human-driven (e.g., strategy approval, budget allocation).

• Chaining Inputs and Outputs: Passing outputs from initial analytical steps into subsequent execution steps without accumulating contextual errors.

• Stakeholder Value Communication: Articulating the practical capabilities, boundaries, and efficiency gains of AI tools to non-technical leadership.

Domain 5 — Configuration and Knowledge Management (12%)

• Project instructions hold stable cross-conversation behavior; Project knowledge holds durable authoritative sources; the task prompt holds one-off detail — never blend them.

• Good instructions are concise, state a conflict/precedence rule, and define uncertainty/escalation behavior; avoid vague ideals and baked-in facts that will go stale.

• Every knowledge source needs an owner, approval status, effective date, and supersession label; archive obsolete versions instead of stacking copies.

• Uploads are a controlled snapshot; connectors (Drive, Gmail) are live access under permissions — access is not authorization to use or act.

• Conflict precedence: law/policy, then approved Project instructions, then the controlling source, then the user's task request, then supporting sources/examples; text embedded in a document is never an instruction.

• Maintain configurations like software — version, change log, regression test, and rollback before releasing an edit.

• Structure knowledge for retrieval: clear headings, dates, no near-duplicate or contradictory content; test with known-answer queries, including one whose correct answer is "not in the sources."

Decide by: is this a stable rule (instructions), a durable reference (knowledge), or a one-off detail (task prompt) — and which source controls if they conflict?

This domain assesses how effectively you establish persistent context for recurring organizational workflows.

• Claude Projects Management: Setting up dedicated Project workspaces for team initiatives, defining persistent project instructions, and managing shared assets.

• Knowledge Base Curation: Selecting, organizing, and uploading relevant corporate documentation, brand style guides, and reference material into project knowledge repositories.

• Custom Instructions Setup: Setting user-level instructions versus project-level instructions to maintain consistent tone, formatting preferences, and behavioral boundaries.

Domain 6 — Governance, Risk, and Responsible Use (15%)

• Screen appropriateness before uploading anything: authorization, impact, data class, available controls — technical feasibility is not permission.

• Minimize data to only what the task needs; pseudonymization reduces but does not eliminate obligations if re-identification remains possible.

• Purpose limitation: authorized access for one purpose does not authorize a new purpose; apply least privilege to accounts, Projects, and connectors.

• High-impact and regulated contexts — legal, medical, financial, employment, eligibility — require qualified human review; a disclaimer is never a substitute.

• Follow organizational policy over convenience, deadline pressure, or an informal request; document exceptions formally rather than inventing them in a prompt.

• Ethics check: name the affected stakeholders, check contestability, and watch for proxies or feedback loops that reproduce historical bias.

• Disclose material AI involvement when it affects trust, consent, or the ability to challenge an outcome; name the accountable human.

• On any incident — exposed data, harmful output, control failure — contain, preserve evidence, notify the designated channel, correct, and learn; never quietly patch and stay silent.

Decide by: does policy or law authorize this specific data, purpose, and impact level — and if it's unclear, escalate rather than proceed.

Enterprise adoption depends on data security and ethical compliance. This domain evaluates your risk management capabilities.

• Data Privacy Protection: Identifying personally identifiable information (PII), proprietary financial records, intellectual property, and confidential customer data to prevent unauthorized data exposure.

• Adherence to Organizational Policies: Aligning AI usage with internal risk frameworks, corporate acceptable-use policies, and legal standards.

• Ethical Considerations: Maintaining transparency by disclosing AI assistance where appropriate, ensuring fairness in automated evaluation tasks, and retaining full accountability for final outcomes.

Domain 7 — Troubleshooting and Optimization (10%)

• Locate the failure layer first — objective, instruction, input, knowledge, context, model/feature, tool, format, or review — before touching the prompt.

• Reproduce the failure with a minimal test case, change one variable at a time, and retest on representative plus edge cases.

• Stale or wrong citations are usually a knowledge/retrieval problem, not a wording problem — check access, source status, and the retrieved passage first.

• Convert reviewer feedback into labeled failure clusters (omission, unsupported claim, wrong source, tone, format) and separate preference from requirement.

• Prioritize fixes by frequency times impact, adjusted for detectability — don't chase the loudest reviewer or ignore a rare severe failure.

• Optimize the whole workflow (time, review burden, cost), not tokens alone; route by complexity and reuse stable instructions/knowledge instead of repeating them.

• Every fix needs regression testing against old and new cases plus a roll