CertSafari

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

    35 free sample questions from our bank of 664+, covering every exam domain, with answers and detailed explanations. Updated August 2026.

    Practice all 664 questions in quiz mode

    Domain 1: Prompting and Task Execution

    Subdomain 1.1: Create effective prompts for business and technical tasks

    1.A developer wants Claude to reliably follow a specific tone and structure when generating customer emails. Plain-language instructions alone haven't produced consistent results. What technique should they add to the prompt?

    1. A.Provide three to five diverse examples wrapped in <example> tags to distinguish them from instructions
    2. B.Provide one lengthy example that repeats the exact same pattern several times within a single block
    3. C.Describe the desired tone and structure in another paragraph without giving any example text
    4. D.List every possible edge case output in one long comma-separated sentence at the prompt's end
    Show answer & explanation

    Correct answer: A — Provide three to five diverse examples wrapped in <example> tags to distinguish them from instructions

    • A. Correct. Multishot prompting with three to five relevant, diverse examples wrapped in <example> tags is one of the most reliable ways to steer output format, tone, and structure.
    • B. A single repeated example lacks diversity and risks Claude picking up an unintended narrow pattern instead of generalizing the desired style.
    • C. More prose description without concrete examples is exactly what has already been tried and has not produced consistent results.
    • D. A compressed list of edge cases is not the same as worked examples and does not show Claude the actual tone and structure to imitate.

    Subdomain 1.2: Apply task decomposition techniques to structure complex requests

    2.A prompt combines background context, several examples, and detailed instructions for a multi-part task. What technique helps Claude distinguish these different parts of the request?

    1. A.Wrap each part of the prompt, such as context, examples, and instructions, in its own XML tags
    2. B.Combine all the context, examples, and instructions together into one single unbroken paragraph
    3. C.Place the instructions at the very end with no separation from the examples that precede them
    4. D.Repeat the instructions twice, once before and once after the examples, without using any tags
    Show answer & explanation

    Correct answer: A — Wrap each part of the prompt, such as context, examples, and instructions, in its own XML tags

    • A. Correct. Structuring a prompt with XML tags around distinct parts, such as context, examples, and instructions, helps Claude clearly distinguish between the different components of a complex request.
    • B. Incorrect. Merging everything into one unbroken paragraph makes it harder for Claude to tell where context ends and instructions begin.
    • C. Incorrect. Placing instructions at the end without any separation still leaves the different components ambiguously blended together.
    • D. Incorrect. Repeating instructions without tags adds redundancy but does not clearly separate the distinct parts of the prompt.

    Subdomain 1.2: Apply task decomposition techniques to structure complex requests

    3.During an agentic coding task, Claude creates several temporary Python scripts as a scratchpad while iterating on a solution. The team wants to minimize net new files left behind. What should the prompt instruct?

    1. A.Instruct Claude to clean up the temporary scratch files it created once the final output is saved
    2. B.Instruct Claude to never create any temporary files, even for testing and iteration
    3. C.Instruct Claude to commit every temporary scratch file to version control for auditing
    4. D.Instruct Claude to rename its temporary scratch files as the final deliverable instead of removing them
    Show answer & explanation

    Correct answer: A — Instruct Claude to clean up the temporary scratch files it created once the final output is saved

    • A. Correct. To minimize net new file creation, Claude can be instructed to clean up after itself, removing temporary scratch files once the final output has been saved.
    • B. Incorrect. Forbidding temporary files entirely removes a useful scratchpad technique that can improve outcomes in agentic coding.
    • C. Incorrect. Committing every temporary scratch file increases net new files rather than minimizing them.
    • D. Incorrect. Renaming a temporary scratch file as the final deliverable risks shipping unfinished scratch work instead of a properly produced final file.

    Subdomain 1.4: Adapt prompting strategies based on task type (analysis, research, drafting, brainstorming)

    4.A researcher asks Claude to answer a factual question that could be verified from several independent sources before being treated as reliable. Which instruction best supports getting a trustworthy research answer?

    1. A.Ask Claude to cross-check the claim against several sources before calling it settled
    2. B.Ask Claude to answer only from the first source it finds so the reply comes back fastest
    3. C.Ask Claude to state the answer confidently no matter how many sources were checked
    4. D.Ask Claude to leave out which sources it consulted so the answer stays concise
    Show answer & explanation

    Correct answer: A — Ask Claude to cross-check the claim against several sources before calling it settled

    • A. Correct. Encouraging source verification, checking a claim across multiple sources, is a recommended practice for getting reliable results on research tasks.
    • B. Incorrect. Relying on the first source found skips the verification step needed to establish that a research finding is reliable.
    • C. Incorrect. Confident phrasing unrelated to actual verification can misrepresent how well-supported the claim really is.
    • D. Incorrect. Hiding which sources were consulted removes transparency and makes it harder to judge how well the claim was actually verified.

    Subdomain 1.4: Adapt prompting strategies based on task type (analysis, research, drafting, brainstorming)

    5.A knowledge base tool asks Claude to answer users' questions using only information retrieved from company documents, and the team wants every claim traceable to the exact retrieved passage it came from. Which technique best supports this analysis and answering task?

    1. A.Have Claude quote the specific passages from the retrieved documents before answering
    2. B.Have Claude answer directly from memory instead of using the retrieved documents at all
    3. C.Have Claude paraphrase only the general topic of the documents, not any passage
    4. D.Have Claude answer first and only fetch supporting documents afterward instead
    Show answer & explanation

    Correct answer: A — Have Claude quote the specific passages from the retrieved documents before answering

    • A. Correct. Grounding responses in quotes, having Claude quote relevant passages before answering, helps produce traceable, source-grounded answers, which fits the requirement that every claim be traceable to a passage.
    • B. Incorrect. Answering from memory instead of the retrieved documents does not produce claims traceable to specific retrieved passages.
    • C. Incorrect. A general paraphrase without quoting specific passages does not give the traceability to exact source text that the team needs.
    • D. Incorrect. Answering before retrieving documents reverses the order needed to ground the answer in the actual retrieved passages.

    Subdomain 1.3: Iterate prompts to improve output quality

    6.A prompt mixes background context, formatting instructions, few-shot examples, and a variable user query all as plain, unlabeled paragraphs. Claude sometimes treats an example as if it were part of the actual input to process. Which iteration would most directly fix this misinterpretation?

    1. A.Shorten the prompt overall so there is less text for Claude to potentially misread, removing non-essential parts.
    2. B.Wrap each type of content in its own XML tag, such as <context>, <instructions>, <examples>, and <input>.
    3. C.Move the few-shot examples to a separate follow-up message sent immediately after the first response.
    4. D.Rewrite all of the content into a single unbroken paragraph so there are no section boundaries to confuse.
    Show answer & explanation

    Correct answer: B — Wrap each type of content in its own XML tag, such as <context>, <instructions>, <examples>, and <input>.

    • A. Incorrect. Shortening the prompt reduces total text but does not address the core issue of structural ambiguity between content types. The model will still confuse examples with input because the boundaries remain unclear.
    • B. Correct. Wrapping each content type in its own XML tag provides explicit structure that helps Claude distinguish instructions, context, examples, and user input. This is the recommended technique to prevent misinterpretation when a prompt mixes multiple content types.
    • C. Incorrect. Sending examples in a follow-up message leaves the original prompt ambiguous, so Claude may still mistake examples for input during the first turn. This approach alters the conversation flow without directly resolving the misinterpretation.
    • D. Incorrect. Combining all content into a single unbroken paragraph removes any visual or structural cues that help Claude separate examples from the actual query. This would likely increase confusion, not reduce it.

    Domain 2: Output Evaluation and Validation

    Subdomain 2.1: Evaluate Claude-generated outputs for accuracy and completeness

    7.A team asks Claude to read a scanned receipt image and extract the total amount, and they want to check whether the extracted total actually matches what appears in the image. Which capability of Claude is directly relevant to producing this output in the first place?

    1. A.Extended thinking, since it is required for any task involving numbers or currency
    2. B.Batch processing, since it allows many receipts to be queued for later manual review
    3. C.Prompt caching, since it stores the image so it does not need to be re-uploaded
    4. D.Vision, since it allows Claude to process and analyze the visual content of the image
    Show answer & explanation

    Correct answer: D — Vision, since it allows Claude to process and analyze the visual content of the image

    • A. Extended thinking provides deeper reasoning transparency, but it is not required for every task involving numbers, and it is not what enables Claude to read an image in the first place.
    • B. Batch processing is about asynchronous request throughput and does not affect whether Claude can read and extract data from the image itself.
    • C. Prompt caching reduces the need to resend repeated content across requests but does not enable Claude to interpret the visual content of an image.
    • D. Correct. Vision is the capability that lets Claude process and analyze visual input, such as a scanned receipt image, which is what makes extracting the total amount from the image possible at all.

    Subdomain 2.2: Identify hallucinations, inconsistencies, and biases in responses

    8.A developer wants to grade thousands of Claude outputs for quality and plans to have the same model that generated the outputs also score them. A colleague flags this as a risky design. What is the main concern with using the same model to both generate and grade its own outputs?

    1. A.The model cannot process its own prior outputs as input, so grading requests will fail
    2. B.The model will exceed its context window whenever it grades more than a few responses
    3. C.The model may favor its own outputs, so a separate model or human should grade instead
    4. D.The model's pricing doubles whenever it is reused for both generation and grading tasks
    Show answer & explanation

    Correct answer: C — The model may favor its own outputs, so a separate model or human should grade instead

    • A. Incorrect. A model can process its own prior text as input to a new request; there is no technical restriction preventing this.
    • B. Incorrect. Context window limits depend on total input size, not on whether the grading model is the same one that generated the text.
    • C. Correct. A recommended best practice is to use a separate model or human reviewers for grading, since a model grading its own generations can be biased toward judging them favorably.
    • D. Incorrect. Pricing is based on token usage per request, not on whether the same model is reused across generation and grading steps.

    Subdomain 2.4: Determine when human review or additional verification is required

    9.A team enables extended thinking with a summarized display so they can review Claude's reasoning steps before accepting the answer to a complex multi-step business problem. What is the primary benefit of reviewing this reasoning?

    1. A.It removes the need to check the model's final answer against any outside data source for the problem
    2. B.It lets reviewers catch flawed logic or a wrong assumption before the final answer is trusted and acted on
    3. C.It proves mathematically that the final answer is correct before anyone reads the generated response
    4. D.It automatically corrects any flawed reasoning it detects before the final answer is ever shown to reviewers
    Show answer & explanation

    Correct answer: B — It lets reviewers catch flawed logic or a wrong assumption before the final answer is trusted and acted on

    • A. Incorrect. Reviewing the reasoning steps does not replace checking the conclusion against external data when the conclusion depends on facts outside the model.
    • B. Correct. Visible reasoning steps let a reviewer spot a flawed assumption or logic error before the final answer is trusted, which is the main verification benefit of extended thinking.
    • C. Incorrect. Showing the reasoning process is not a mathematical proof of correctness; it is a transparency aid that still requires human judgment.
    • D. Incorrect. Extended thinking displays the reasoning trace; it does not automatically fix errors found within it, a human must still catch and correct them.

    Subdomain 2.5: Edit, adapt, refine, and compare outputs for the intended audience

    10.A nonprofit needs a grant report for both a technical evaluator and a general public donor newsletter, but the first draft uses terms like 'confidence interval' and 'effect size' throughout. What is the best way to adapt the newsletter version?

    1. A.Restate the statistical findings in plain language describing what the results mean practically
    2. B.Keep the statistical terms but define each one in a footnote at the end of the newsletter
    3. C.Remove all numbers from the newsletter so no explanation of findings is needed
    4. D.Keep the report identical for both audiences since the underlying data does not change
    Show answer & explanation

    Correct answer: A — Restate the statistical findings in plain language describing what the results mean practically

    • A. Correct. Translating statistical jargon into plain, practical statements is exactly what adapting content for a general donor audience requires while preserving the actual findings.
    • B. Incorrect. Footnoted definitions still leave jargon embedded in the main text, requiring donors to cross-reference terms instead of reading naturally.
    • C. Incorrect. Removing all numbers strips out meaningful impact information donors would want, rather than simply making it understandable.
    • D. Incorrect. Using the identical technical report for donors ignores that the intended audience has a different background and needs different phrasing.

    Subdomain 2.5: Edit, adapt, refine, and compare outputs for the intended audience

    11.In general, why is telling Claude about the intended audience's role, background, or purpose for reading useful when requesting a written output?

    1. A.It helps Claude choose vocabulary, detail, and framing that will be clear to that specific reader
    2. B.It guarantees the output will always be shorter than a version without audience information
    3. C.It removes the need to review the output afterward since it becomes automatically correct
    4. D.It allows Claude to skip explaining any technical concepts regardless of the topic
    Show answer & explanation

    Correct answer: A — It helps Claude choose vocabulary, detail, and framing that will be clear to that specific reader

    • A. Correct. Audience details let Claude calibrate vocabulary, depth, and framing so the response is understandable and relevant to the reader it is written for.
    • B. Incorrect. Audience information affects clarity and framing, not a guaranteed reduction in output length.
    • C. Incorrect. Providing audience context does not eliminate the need to review outputs for accuracy before use.
    • D. Incorrect. Some audiences still need technical concepts explained; audience information changes how much explanation is needed, not whether any is included.

    Subdomain 2.6: Organize and curate information and select appropriate output formats (artifacts, inline, structured data)

    12.A user asks Claude to write a 400-word case study about a customer's product rollout, noting they plan to revise the wording several times before sharing it with stakeholders. How should Claude format the response?

    1. A.As a Markdown artifact, because the case study is long-form, self-contained, and meant for iterative editing before external sharing.
    2. B.Inline in the chat, because inline responses are easier to scroll back to than artifacts when a user wants to review prior wording.
    3. C.As a strict-tool-validated response, because tool schemas automatically enforce the output to match the requested 400-word length exactly.
    4. D.As a CSV export, because tabular formats are often the standard way to prepare narrative content for stakeholder distribution.
    Show answer & explanation

    Correct answer: A — As a Markdown artifact, because the case study is long-form, self-contained, and meant for iterative editing before external sharing.

    • A. Correct. The case study is a long-form, self-contained document meant for multiple rounds of revision. A Markdown artifact provides an editable, standalone file that the user can refine and later share with stakeholders, perfectly matching the described use case.
    • B. Incorrect. Inline responses are embedded in the chat, making them harder to isolate and edit across multiple rounds of revision. An artifact offers a cleaner, more reusable format for iterative work, especially for longer content.
    • C. Incorrect. Strict tool validation ensures that tool call parameters meet a schema, not that prose output meets a specific word count. It does not provide any formatting advantage for a case study that needs to be revised.
    • D. Incorrect. CSV is designed for tabular data, not narrative prose. A case study is a written narrative, and a tabular format is inappropriate for stakeholder distribution of such content.

    Domain 3: Product and Model Selection

    Subdomain 3.3: Align model selection with task requirements (cost, speed, quality)

    13.A team is designing a workflow that runs for long stretches of time asynchronously and would prefer a pre-built, configurable agent harness running in managed infrastructure rather than building and maintaining a custom agent loop. Which building approach fits this need?

    1. A.The Messages API
    2. B.The Batch API
    3. C.The Files API
    4. D.Claude Managed Agents
    Show answer & explanation

    Correct answer: D — Claude Managed Agents

    • A. The Messages API provides direct model prompting access and is best for custom agent loops and fine-grained control, which is the opposite of the pre-built harness this team wants.
    • B. The Batch API processes large volumes of requests asynchronously at a discount, but it is not the configurable agent harness described for long-running agent workflows.
    • C. The Files API manages uploaded documents and data for reuse across requests; it is not an agent-running harness for long-running asynchronous workflows.
    • D. Claude Managed Agents are described as a pre-built, configurable agent harness that runs in managed infrastructure, best suited for long-running tasks and asynchronous work.

    Subdomain 3.1: Select appropriate Claude product features (Projects, research mode, chat, artifacts)

    14.A game designer asks Claude to build a playable HTML5 prototype of a simple puzzle game with score tracking, and wants to keep tweaking the mechanics over the next hour based on playtesting. Which feature keeps the running prototype visible and editable throughout that session?

    1. A.The artifact panel keeps the prototype rendered separately while accepting live iterative edits.
    2. B.A Project is needed since project knowledge alone renders the HTML5 prototype and supports iterative tweaks.
    3. C.Research mode is required because its engine renders and updates interactive content during iterative tweaks.
    4. D.A shared team workspace is necessary because individual accounts cannot render and tweak HTML5 prototypes.
    Show answer & explanation

    Correct answer: A — The artifact panel keeps the prototype rendered separately while accepting live iterative edits.

    • A. Correct. The artifact panel is designed to render self-contained content like an HTML5 game prototype separately from the chat, allowing the designer to see a live view while requesting iterative edits to the mechanics. It explicitly supports this interactive, visible feedback loop.
    • B. Incorrect. Project knowledge bases store reference documents and do not render live interactive content. Artifacts—not a Project—are required to render the HTML5 prototype and support iterative tweaks within the conversation.
    • C. Incorrect. Research mode performs web searches and has no engine for rendering or updating interactive HTML5 content. That rendering and iterative edit capability comes from artifacts, not research mode.
    • D. Incorrect. Individual accounts can render artifacts including HTML5 prototypes without a shared team workspace. The artifact panel works within a single user’s session to display and iteratively edit the prototype.

    Subdomain 3.4: Understand and manage context limitations and memory considerations (when to restart, summarize, or persist)

    15.A developer is about to start a completely unrelated task in the same terminal and does not want any of the current conversation's context, files read, or discussion to influence Claude's next response, but wants to be able to return to the original conversation later if needed. What should they do?

    1. A.Run /clear to start with an empty context, as the conversation is saved for later resumption.
    2. B.Run /compact to summarize the current conversation, then type the new task in that context.
    3. C.Run /rewind to restore to the first message, eliminating all later context for a fresh start.
    4. D.Open a second terminal window in a different directory to keep the current session clean.
    Show answer & explanation

    Correct answer: A — Run /clear to start with an empty context, as the conversation is saved for later resumption.

    • A. Correct. /clear starts a completely new, empty context, ensuring no residual information from the previous conversation influences the new task. The original conversation is preserved and can be resumed later, exactly meeting the developer's requirements.
    • B. Incorrect. /compact creates a summary of the current conversation and keeps it in context, so details of the old task would still affect the response to the new task. It does not provide a clean slate.
    • C. Incorrect. /rewind restores the conversation to its first message, but the original context is not fully cleared; the first message and any associated files or discussion remain active. This does not provide an entirely fresh start and may not preserve the full conversation for easy resumption.
    • D. Incorrect. Opening a second terminal window in a different directory starts a new session, leaving the original terminal's conversation intact but not clearing it. This does not address the need to reuse the same terminal with a fresh context while retaining the ability to return to the original conversation.

    Subdomain 3.3: Align model selection with task requirements (cost, speed, quality)

    16.A developer is using Claude Opus 4.8 for advanced coding and high-autonomy agentic work. According to Anthropic's documentation, which effort level is recommended for such tasks?

    1. A.low
    2. B.high
    3. C.xhigh
    4. D.max
    Show answer & explanation

    Correct answer: C — xhigh

    • A. Incorrect. The 'low' effort level is designed to conserve tokens and is not appropriate for advanced coding and high-autonomy agentic work. Anthropic recommends 'xhigh' for such tasks.
    • B. Incorrect. While 'high' is a valid effort level suitable for complex debugging, it is not the recommended level for high-autonomy agentic work. The documentation reserves 'xhigh' for long autonomous coding sessions and the most demanding agentic tasks.
    • C. Correct. Anthropic's documentation recommends the 'xhigh' effort level for high-autonomy agentic work. It increases thinking depth, tool call appetite, response length, and agentic persistence, making it ideal for long autonomous coding sessions and demanding agentic tasks with Claude Opus 4.8.
    • D. Incorrect. 'max' is not a documented effort level for Claude Opus 4.8. The supported effort levels include low, medium, high, and xhigh, with 'xhigh' being the recommended setting for high-autonomy tasks.

    Domain 4: Workflow Integration and Solution Design

    Subdomain 4.1: Apply Claude to analyze requirements and use cases

    17.A team wants to trade off response thoroughness against token cost and latency within the same model, without switching to a different model family. Which parameter should they tune?

    1. A.Effort parameter
    2. B.Context window size
    3. C.Batch discount rate
    4. D.Zero data retention flag
    Show answer & explanation

    Correct answer: A — Effort parameter

    • A. Correct. The effort parameter controls how many tokens Claude uses when responding, trading off response thoroughness against token efficiency and latency within a single model, and is often a better lever than switching models entirely.
    • B. Incorrect. Context window size controls how much input a model can process; it is not the lever for trading thoroughness against cost and latency within one model.
    • C. Incorrect. The batch discount rate is a fixed pricing benefit of asynchronous batch processing, not a tunable parameter for response depth.
    • D. Incorrect. The zero data retention flag governs data handling and storage policy; it has no effect on response thoroughness, cost, or latency tradeoffs.

    Subdomain 4.5: Communicate Claude's value and limitations to stakeholders

    18.An HR stakeholder proposes letting Claude autonomously make final hiring decisions from resumes with no human involvement, to speed up recruiting. What limitation should be raised, and what should be recommended instead?

    1. A.Raise that model outputs can reflect biases in input data, and recommend keeping a human reviewer in the final decision.
    2. B.Raise no concerns, since removing human involvement from hiring eliminates all risk of biased outcomes.
    3. C.Raise that Claude cannot read resume text at all, so the proposal is technically impossible to build.
    4. D.Raise that hiring speed is the only relevant factor, so fairness and bias considerations can be dropped entirely.
    Show answer & explanation

    Correct answer: A — Raise that model outputs can reflect biases in input data, and recommend keeping a human reviewer in the final decision.

    • A. Correct. Like any model trained on real-world data, Claude's outputs can reflect biases present in that data, so high-stakes decisions like hiring should keep a human reviewer involved rather than being fully automated.
    • B. Incorrect. Removing human oversight does not eliminate bias risk; unsupervised automation can actually amplify it if left unchecked.
    • C. Incorrect. Claude can process and reason about resume text, so this claim is factually wrong and not the relevant concern here.
    • D. Incorrect. Fairness and legal compliance are material concerns in hiring and should not be deprioritized in favor of speed alone.

    Subdomain 4.3: Use Claude to support solution design, development, and iteration

    19.Which approach best leverages Claude to improve a basic database schema design?

    1. A.Accept the basic schema as given, and use Claude to write efficient queries and mappings during coding.
    2. B.Ask Claude to automatically normalize the schema to fifth normal form based solely on the table structures provided.
    3. C.Provide Claude with business rules and ask for an enhanced schema with relationships and constraints.
    4. D.Request that Claude propose a document-based NoSQL schema to replace the relational design for better scalability.
    Show answer & explanation

    Correct answer: C — Provide Claude with business rules and ask for an enhanced schema with relationships and constraints.

    • A. Incorrect. Accepting the basic schema without refinement and moving straight to query optimization risks building on a flawed foundation. Claude should be used to improve the schema design first, which leads to better long-term outcomes.
    • B. Incorrect. Automatically normalizing to fifth normal form based only on table structures lacks the business context needed for a practical design. Without understanding real-world rules, the resulting schema may be overly complex and misaligned with actual needs.
    • C. Correct. Providing business rules allows Claude to generate a schema with appropriate relationships and constraints, resulting in a design that accurately reflects the domain. This targeted approach produces a more usable and maintainable database.
    • D. Incorrect. Swapping to a document-based NoSQL schema without analyzing data access patterns does not inherently solve relational design challenges. The focus should be on defining clear relationships and constraints within the given context, not prematurely changing paradigms.

    Subdomain 4.2: Leverage Claude for research, planning, and process optimization

    20.A research assistant repeatedly sends Claude the same lengthy background dossier with only small follow-up questions changing each turn, and wants to cut both cost and latency across the session. Which feature directly addresses this?

    1. A.Structured outputs enforce a fixed JSON schema on every generated response consistently.
    2. B.The advisor tool supplies mid-generation strategic guidance from a second model.
    3. C.Prompt caching reduces cost and latency by reusing the background dossier across turns.
    4. D.The tool search tool loads relevant tools on demand from a large tool catalog.
    Show answer & explanation

    Correct answer: C — Prompt caching reduces cost and latency by reusing the background dossier across turns.

    • A. Incorrect. Structured outputs ensure responses adhere to a predefined JSON schema, but they do not address the repeated transmission and processing of the same background dossier. Therefore, they do not reduce cost or latency across multiple turns with reused long context.
    • B. Incorrect. The advisor tool provides mid-generation strategic guidance from a second model, which can improve response quality but does not tackle the repeated processing of the same lengthy context. It does not directly lower costs or latency from re-sending the dossier.
    • C. Correct. Prompt caching allows Claude to store and reuse the large, static background dossier across turns, so only the small follow-up questions need to be processed in subsequent requests. This directly reduces both cost (by avoiding reprocessing the entire context each time) and latency for the session.
    • D. Incorrect. The tool search tool helps discover and load relevant tools from a large catalog on demand, but it does not involve reusing previously sent background content. It is unrelated to cutting cost and latency from repeated long context.

    Domain 5: Configuration and Knowledge Management

    Subdomain 5.2: Manage uploaded knowledge and connectors (e.g., Google Drive, Gmail)

    21.Which of the following file types is supported for upload to a Claude project's knowledge base?

    1. A.DOCX
    2. B.EXE file
    3. C.ZIP archive
    4. D.MP4 video
    Show answer & explanation

    Correct answer: A — DOCX

    • A. Correct. Project knowledge bases support formats including PDF, DOCX, CSV, TXT, HTML, ODT, RTF, and EPUB.
    • B. Incorrect. EXE is an executable binary format and is not a supported knowledge upload type.
    • C. Incorrect. ZIP archives are not among the supported project knowledge file types.
    • D. Incorrect. MP4 video files are not a supported format for project knowledge uploads.

    Subdomain 5.2: Manage uploaded knowledge and connectors (e.g., Google Drive, Gmail)

    22.A user asks Claude, connected via the Drive connector, to save a report Claude just generated into a specific Drive folder. What can Claude do?

    1. A.Claude can only save generated files to the root of Drive, not a folder.
    2. B.Claude cannot save any files to Drive, only read files already there.
    3. C.Claude can create folders and save Claude-generated files directly into Drive itself.
    4. D.Claude can save files to Drive only when code execution is separately turned on first.
    Show answer & explanation

    Correct answer: C — Claude can create folders and save Claude-generated files directly into Drive itself.

    • A. Incorrect. Folder creation and organized saving are supported, not just root-level file placement.
    • B. Incorrect. Saving generated files to Drive is an explicit connector capability beyond just reading existing files.
    • C. Correct. The Drive connector supports creating folders and saving Claude-generated files, so the report can go into the requested folder.
    • D. Incorrect. Code execution is required for XLSX chat uploads, not for the Drive connector's save and folder features.

    Subdomain 5.4: Inform, maintain, and update Claude configurations, knowledge sources, and instructions

    23.A developer suspects Claude is ignoring a path-scoped rule in .claude/rules/api.md and wants to confirm exactly when the rule file loads and why. Which tool gives them that visibility?

    1. A.The InstructionsLoaded hook, which logs which instruction files load, when, and why
    2. B.The claude mcp reset-project-choices command, which clears prior MCP approval decisions
    3. C.The claudeMdExcludes setting, which is used to skip specific files from loading at all
    4. D.The --add-dir flag, which grants Claude access to directories outside the working directory
    Show answer & explanation

    Correct answer: A — The InstructionsLoaded hook, which logs which instruction files load, when, and why

    • A. Correct. The InstructionsLoaded hook is designed for exactly this: logging which instruction files are loaded, when they load, and why, which helps debug path-specific rules or lazily loaded files.
    • B. Incorrect. claude mcp reset-project-choices clears MCP server approval decisions and has nothing to do with rule file loading visibility.
    • C. Incorrect. claudeMdExcludes is used to prevent specific files from loading, not to report when and why a rule file did load.
    • D. Incorrect. --add-dir extends which directories Claude can access; it does not report load timing or reasoning for instruction files.

    Subdomain 5.1: Configure Claude Projects with instructions and knowledge sources

    24.An Enterprise admin adds an entire department to a project in bulk rather than sharing it with each employee individually. Which project capability does this scenario describe?

    1. A.Bulk user addition and organization-wide distribution is a Team/Enterprise collaboration feature.
    2. B.Automatic project duplication creates a completely separate standalone copy for each added employee.
    3. C.A default sharing behavior is equally available on free accounts with no plan restriction.
    4. D.A knowledge base file setting has no relation to project sharing or membership controls.
    Show answer & explanation

    Correct answer: A — Bulk user addition and organization-wide distribution is a Team/Enterprise collaboration feature.

    • A. Correct. The scenario describes bulk user addition and organization-wide distribution, a collaboration feature specifically available in Team and Enterprise plans. This capability allows admins to share a project with many users efficiently, rather than adding each member individually.
    • B. Incorrect. Bulk user addition shares the same existing project with multiple members; it does not automatically duplicate the project or create separate standalone copies. Each added employee accesses the shared project, not an individual duplicate.
    • C. Incorrect. Bulk, organization-wide sharing is a Team/Enterprise collaboration feature with plan restrictions, not a default behavior equally available on free accounts. Free accounts lack dedicated bulk sharing capabilities and generally require individual user invitations.
    • D. Incorrect. The described scenario concerns project sharing and membership controls, not knowledge base file settings. Knowledge base settings govern document uploads for Claude’s context, which is unrelated to the distribution method for granting project access.

    Subdomain 5.4: Inform, maintain, and update Claude configurations, knowledge sources, and instructions

    25.An administrator adds two configuration files, `10-telemetry.json` and `20-security.json`, to the `managed-settings.d/` directory used by Claude Code. Both files contain the same setting with different values. According to Claude Code's rules for loading files in this directory, which value will be used?

    1. A.The value from `20-security.json`, because files are processed in alphabetical order and the last file's value takes priority.
    2. B.The value from `10-telemetry.json`, because files are processed based on their file system creation timestamps.
    3. C.The value from the base `managed-settings.json` file, because files in the directory cannot override the main configuration.
    4. D.Neither value; when two files define the same setting, the setting is discarded.
    Show answer & explanation

    Correct answer: A — The value from `20-security.json`, because files are processed in alphabetical order and the last file's value takes priority.

    • A. Claude Code processes files in the directory alphabetically. Since `20-security.json` sorts after `10-telemetry.json`, its value overrides the earlier one.
    • B. Claude Code uses alphabetical order by file name, not creation time. Renaming a file changes its priority, regardless of its age.
    • C. Files in the directory are meant to extend or override the base settings. When a file in the directory sets the same key, it overrides the base value.
    • D. Claude Code does not discard conflicting settings. It uses the value from the file that is processed last (the one with the highest alphabetical name).

    Domain 6: Governance, Risk, and Responsible Use

    Subdomain 6.1: Identify appropriate and inappropriate use cases

    26.A software team uses Claude to review pull requests, suggest bug fixes, and generate unit tests as part of its normal development workflow, with engineers reviewing all suggested code before merging. Is this an appropriate use of Claude?

    1. A.Yes, because code review and test generation are common development tasks with engineer sign-off
    2. B.No, because the Usage Policy restricts AI-assisted code review to Anthropic-certified engineering teams
    3. C.Yes, but only if the team disables Claude for any file that touches production credentials
    4. D.No, because generated unit tests are excluded from the Usage Policy's permitted development activities
    Show answer & explanation

    Correct answer: A — Yes, because code review and test generation are common development tasks with engineer sign-off

    • A. Correct. Code review, bug-fix suggestions, and test generation are standard, supported development use cases, and human engineers retaining final approval fits normal responsible practice.
    • B. Incorrect. There is no requirement that engineering teams be specially certified by Anthropic to use Claude for code review or development assistance.
    • C. Incorrect. There is no blanket requirement to disable Claude on credential-related files for this use case to be appropriate; the scenario as described is already acceptable.
    • D. Incorrect. Generating unit tests is a routine and supported coding use case, not an excluded activity.

    Subdomain 6.3: Follow organizational AI policies and governance standards

    27.A new engineer needs to use both the Workbench and the organization's Claude Code workspace for a pilot project, but should not be able to manage API keys or webhook endpoints. Which Console role fits this need?

    1. A.Developer
    2. B.Limited Developer
    3. C.Admin
    4. D.Claude Code User
    Show answer & explanation

    Correct answer: D — Claude Code User

    • A. Incorrect. Developer includes API key and webhook management along with session traces and file downloads, exceeding this need.
    • B. Incorrect. Limited Developer explicitly grants API key and webhook management, which the engineer should not have.
    • C. Incorrect. Admin grants full organization-wide administrative control, far more than a pilot project role requires.
    • D. Correct. Claude Code User grants access to both the Workbench and the organization's Claude Code workspace without key or webhook management capabilities.

    Subdomain 6.3: Follow organizational AI policies and governance standards

    28.The finance team needs to update payment methods and monitor usage and cost data across the organization, but should have no access to the Claude Code workspace or Workbench. Which Console role matches this requirement?

    1. A.Admin
    2. B.Developer
    3. C.Billing
    4. D.User
    Show answer & explanation

    Correct answer: C — Billing

    • A. Incorrect. Admin grants far broader access, including Workbench, Claude Code, and user management, beyond what finance needs.
    • B. Incorrect. Developer grants Workbench and Claude Code access along with technical capabilities that finance does not need.
    • C. Correct. Billing can manage billing details and view usage and cost data without any Claude Code workspace access.
    • D. Incorrect. User grants Workbench access but no billing management capability at all.

    Subdomain 6.2: Apply data sensitivity, regulatory, and privacy considerations

    29.A European customer wants a specific API request's model inference to run only on US-based infrastructure due to a regulatory requirement, without changing where their workspace data is stored at rest. Which setting should they use?

    1. A.Set the inference_geo parameter to "us" on that Messages API request
    2. B.Configure the workspace geo setting to "us" during workspace creation instead
    3. C.Enable HIPAA readiness, which restricts all inference to US infrastructure by default
    4. D.Set the default_inference_geo workspace setting to "global" to force US-only routing
    Show answer & explanation

    Correct answer: A — Set the inference_geo parameter to "us" on that Messages API request

    • A. Correct. The inference_geo parameter controls where inference runs on a per-request basis, letting a request specify "us" without touching where data is stored at rest.
    • B. Incorrect. Workspace geo controls where data is stored at rest, not per-request inference location, and it can't be changed after workspace creation.
    • C. Incorrect. HIPAA readiness governs privacy and security safeguards for PHI, not geographic routing of inference by default.
    • D. Incorrect. Setting default_inference_geo to "global" does the opposite of forcing US-only routing; "us" is the value that restricts inference geography.

    Subdomain 6.2: Apply data sensitivity, regulatory, and privacy considerations

    30.A compliance officer at a regulated enterprise needs a 6-year audit trail of administrative and usage activity across their Claude organization for a regulatory review. Which Anthropic capability is designed to retain this kind of data for that duration?

    1. A.The Activity Feed, which is designed to retain organizational activity data for a full 6 years
    2. B.The Compliance API, since it follows the same 30-day retention model applied to Covered Models
    3. C.Zero data retention, since enabling it for an organization extends retention out to 6 years
    4. D.The Usage and Cost API, since it stores billing and token usage records for 6 years
    Show answer & explanation

    Correct answer: A — The Activity Feed, which is designed to retain organizational activity data for a full 6 years

    • A. Correct. The Activity Feed retains organizational activity data for 6 years, sitting outside the standard ZDR and HIPAA retention models.
    • B. Incorrect. The Compliance API follows its own retention model, distinct from the 30-day Covered Model requirement, and is not the source of the 6-year audit trail.
    • C. Incorrect. Zero data retention is designed to minimize storage, not extend it to 6 years; the 6-year figure applies specifically to the Activity Feed.
    • D. Incorrect. The Usage and Cost API tracks usage and billing metrics but is not documented as the 6-year activity audit trail described here.

    Subdomain 6.4: Understand the ethical implications of AI usage

    31.Which of the following best describes the purpose of Anthropic's Transparency Hub?

    1. A.To provide developers with real-time uptime and latency metrics for the Claude API
    2. B.To publish periodic reports on Anthropic's safety practices and enforcement measures
    3. C.To list the internal source code and model weights used to train Claude models
    4. D.To serve as the primary customer support channel for billing and account disputes only
    Show answer & explanation

    Correct answer: B — To publish periodic reports on Anthropic's safety practices and enforcement measures

    • A. Incorrect. Real-time uptime and latency data is provided through Anthropic's status page, not the Transparency Hub.
    • B. Correct. The Transparency Hub publishes information about Anthropic's safety processes, programs, and enforcement metrics to enable public understanding of how Claude is governed responsibly.
    • C. Incorrect. The Transparency Hub does not disclose proprietary source code or model weights.
    • D. Incorrect. Billing and account support is handled through the Help Center, not the Transparency Hub.

    Domain 7: Troubleshooting and Optimization

    Subdomain 7.2: Adjust approach based on feedback and results

    32.A developer wants to evaluate whether Claude's paraphrased answers to the same underlying question remain semantically consistent with each other. Which evaluation approach best fits this goal?

    1. A.Compute cosine similarity between the embeddings of the multiple paraphrased responses
    2. B.Use exact match scoring to check that each paraphrased response is character-for-character identical
    3. C.Apply ROUGE-L scoring against a single fixed reference answer written by the development team
    4. D.Run a binary LLM grader that outputs only 'yes' or 'no' for whether the response is grammatically correct
    Show answer & explanation

    Correct answer: A — Compute cosine similarity between the embeddings of the multiple paraphrased responses

    • A. Correct. Cosine similarity between output embeddings is the documented method for measuring consistency across paraphrased questions, capturing semantic closeness even when wording differs.
    • B. Incorrect. Exact match requires identical text, which is not expected or desired when responses are paraphrased with different wording.
    • C. Incorrect. ROUGE-L against a single fixed reference measures overlap with one specific answer rather than consistency across multiple varied responses.
    • D. Incorrect. A grammar-focused binary grader evaluates a different quality dimension and does not measure semantic consistency between responses.

    Subdomain 7.1: Identify, diagnose, and resolve issues with underperforming prompts or poor outputs

    33.A prompt for a tone-sensitive customer email generator gives Claude no persona or role, just the raw task instructions. Emails come back with an inconsistent, generic tone that doesn't match the brand voice. What is a lightweight change likely to help?

    1. A.Ask Claude to write each email in a different language and then translate it back, hoping tone carries over more consistently.
    2. B.Remove the existing task instructions entirely and rely only on examples to communicate the desired tone.
    3. C.Increase the max_tokens limit so Claude has more space to naturally settle into a consistent tone over a longer response.
    4. D.Add a system prompt sentence establishing Claude's role, such as being a customer service specialist for the brand.
    Show answer & explanation

    Correct answer: D — Add a system prompt sentence establishing Claude's role, such as being a customer service specialist for the brand.

    • A. Incorrect. Round-tripping through translation is not a documented technique for tone control and introduces unnecessary risk of meaning drift.
    • B. Incorrect. Removing the task instructions would leave Claude without clear direction on what to actually write, which is a separate problem from tone and likely to make output worse.
    • C. Incorrect. A larger token budget affects available response length, not tone consistency, and is not the documented lever for brand voice.
    • D. Correct. Anthropic notes that setting a role in the system prompt focuses Claude's behavior and tone for a given use case, and that even a single sentence establishing a role makes a noticeable difference in output consistency.

    Subdomain 7.3: Optimize workflows for efficiency and effectiveness

    34.An engineering team is planning a multi-hour autonomous agent run that will refactor a large, complex codebase with intricate cross-module dependencies. Accuracy on this advanced coding task matters far more than minimizing cost. Which model should the team start with?

    1. A.Claude Opus 4.8, since it is best suited to complex agentic coding and enterprise-scale engineering work
    2. B.Claude Haiku 4.5, since its low cost lets the team run many parallel attempts at the refactor
    3. C.Claude Sonnet 5 configured with the lowest available effort level to keep the session responsive
    4. D.Any model paired with the Batch API so the multi-hour refactor runs at the discounted rate
    Show answer & explanation

    Correct answer: A — Claude Opus 4.8, since it is best suited to complex agentic coding and enterprise-scale engineering work

    • A. Correct. Opus 4.8 is positioned for complex agentic coding, large-scale refactoring, and advanced systems engineering where accuracy on long, autonomous tasks outweighs cost.
    • B. Haiku 4.5 is optimized for fast, economical, high-volume tasks; running many cheap parallel attempts does not substitute for the deeper reasoning needed on an intricate refactor.
    • C. Lowering effort trades intelligence for latency and cost, which works against the stated priority of accuracy on a complex, dependency-heavy refactor.
    • D. Batch processing suits independent, non-urgent bulk requests; a single long-running interactive autonomous agent session is not the asynchronous bulk workload batch is designed for.

    Subdomain 7.2: Adjust approach based on feedback and results

    35.In the context of adjusting a Claude application's approach based on evaluation results, why is it important to measure with the same metrics and test set consistently throughout iteration?

    1. A.It allows performance changes across prompt versions to be attributed to the prompt changes themselves rather than to differences in how each version is measured.
    2. B.It guarantees that every prompt version will achieve a higher score than the previous version by removing evaluation variability, allowing even small prompt changes to appear as improvements.
    3. C.It removes the need to define success criteria in advance because consistent comparison against a fixed test set means any score above the baseline signals success.
    4. D.It ensures the model provider automatically improves the underlying model weights each time the same test set is reused, as consistent feedback refines the model’s parameters.
    Show answer & explanation

    Correct answer: A — It allows performance changes across prompt versions to be attributed to the prompt changes themselves rather than to differences in how each version is measured.

    • A. Correct. Using the same metrics and test set each time ensures that any changes in performance can be traced back to the prompt modifications themselves, not to variations in measurement approach. This isolation of variables is essential for valid iteration and comparison.
    • B. Incorrect. Consistent measurement only guarantees a fair comparison; it does not ensure each version will score higher than the last. Scores can still decrease if the prompt changes hurt performance, so removing evaluation variability does not guarantee improvement.
    • C. Incorrect. Even with a fixed test set and metrics, you still need predefined success criteria to interpret scores, because a score above baseline does not automatically indicate meaningful success. Consistent measurement complements, but does not replace, the need for clear success definitions.
    • D. Incorrect. Reusing a test set provides consistent evaluation, but it does not trigger any changes to the underlying model weights. Model parameters are not updated by evaluation runs; consistent feedback alone does not refine the model's parameters automatically.

    Want the full experience?

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

    Practice all 664 questions in quiz mode