CAF

Free CAF practice questions

10 questions from our CAF practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.

Back to the CAF practice exams

Question 1 of 10, Prompting and Task Execution

An HR specialist pastes a 12-page policy and a set of employee questions into one prompt. Claude sometimes treats lines from the policy as instructions. What is the BEST way to structure the prompt?

  1. AWrap the policy and the questions in separate XML tags and refer to those tags in the instructions Correct answer
  2. BShorten the policy to a one-paragraph summary written from memory so there is less text to confuse Claude
  3. CPut the instructions in capital letters so they stand out from the policy
  4. DSplit the policy into several separate messages and send the questions only after Claude confirms each part

Answer and explanation: XML-style tags clearly separate reference material from instructions, and naming the tags in the instructions tells Claude exactly what to do with each part. Capital letters do not create a structural boundary. Sending many messages adds turns without delimiting content. A summary written from memory discards the source details the answers depend on.

Question 2 of 10, Output Evaluation and Validation

A support lead plans to let Claude draft refund decisions. Drafts for refunds under $50 are low risk; larger refunds have caused costly mistakes. Which review design BEST balances speed and risk?

  1. AAsk Claude to flag drafts it feels unsure about and send only those flagged drafts to a human reviewer
  2. BLet Claude approve all refunds and audit a random sample of decisions at the end of each month
  3. CAuto-approve low-value drafts after spot checks; route drafts above the threshold to a human Correct answer
  4. DHave a human review every draft regardless of value so that no decision is ever made without oversight

Answer and explanation: A value-based threshold sends high-impact decisions to people while keeping low-risk work fast, which is the purpose of a human-in-the-loop threshold. Reviewing everything is safe but discards the efficiency gain. Monthly sampling finds costly mistakes too late. Model self-flagging is not a reliable trigger because confident output can still be wrong.

Question 3 of 10, Output Evaluation and Validation

A policy team gives Claude three internal documents and asks for a summary that uses only those documents. Which instruction BEST supports checking that the summary is grounded?

  1. AAsk Claude to add any helpful background it knows so the summary feels more complete
  2. BAsk Claude to quote the supporting passage for each key claim Correct answer
  3. CAsk Claude to keep the summary under one page so there is less text to review
  4. DAsk Claude to write in a neutral tone so the summary does not sound promotional

Answer and explanation: Requiring a supporting quote for each claim lets reviewers trace every statement back to the provided documents and spot anything ungrounded. Adding outside background works against the only-these-documents requirement. Length and tone limits make review easier or nicer but do not show where claims come from.

Question 4 of 10, Product and Model Selection

An analyst needs a multi-source briefing on a competitor that draws on many web pages and internal Google Drive files, with citations. Which Claude capability is the BEST fit?

  1. AResearch, which searches the web and connected sources and returns a cited report Correct answer
  2. BA custom style called "Briefing" that makes Claude's responses longer and more formal
  3. CExtended thinking with no sources attached, so Claude reasons more deeply from memory
  4. DA single chat reply with web search off, using the knowledge Claude already has from training

Answer and explanation: Research is designed to investigate a question across many searches, including connected sources such as Google Workspace, and produce a report with citations. Training knowledge alone is stale and uncited. A style changes presentation, not sources. Extended thinking without sources cannot retrieve current information.

Question 5 of 10, Workflow Integration and Solution Design

A claims department classifies tasks as Automate, Augment, or Human-only. Which task belongs in Human-only?

  1. ASorting incoming emails into claim, complaint or general inquiry
  2. BMaking the final decision to deny a disputed claim Correct answer
  3. CDrafting a first summary of a claimant's submitted documents
  4. DTranslating a standard acknowledgment letter into Spanish

Answer and explanation: A final denial of a disputed claim is a consequential, accountable decision that should rest with a person. Summarizing documents is augmentation with review. Translating a standard letter and sorting emails are good candidates for automation or light review.

Question 6 of 10, Workflow Integration and Solution Design

A content team wants Claude to turn interview transcripts into blog posts. Results are inconsistent when they ask for everything in one prompt. Which workflow design BEST addresses this?

  1. APut all the steps into a single, much longer prompt with every rule written in capital letters
  2. BDrop the transcripts and write from the topic alone
  3. CAsk Claude to write five complete posts at once and let the team choose the best one each time
  4. DChain the work: extract quotes, then outline, then draft from the approved outline Correct answer

Answer and explanation: Prompt chaining breaks a complex task into steps whose outputs feed the next, and gives the team checkpoints to review. A longer single prompt keeps the same overloaded structure. Generating many full drafts wastes effort without fixing consistency. Removing the transcripts removes the source material the posts depend on.

Question 7 of 10, Configuration and Knowledge Management

A project's knowledge contains the 2024 and the 2026 versions of the same travel policy, and Claude sometimes quotes outdated limits. What is the BEST fix?

  1. ARemove or archive the 2024 version so only the current policy remains Correct answer
  2. BUpload a third copy of the 2026 policy to outweigh the old one
  3. CRename the 2024 file with a longer name so that Claude reads it after the newer policy document
  4. DAdd an instruction asking Claude to pick whichever version it thinks is more likely to be correct

Answer and explanation: Conflicting versions in project knowledge cause stale answers; curating the knowledge so only the current version is present removes the conflict at its source. Letting Claude guess keeps the risk. Duplicate copies and file renaming do not reliably stop the old policy being used.

Question 8 of 10, Governance, Risk, and Responsible Use

An employee asks Claude to reproduce a full paywalled industry article so the team can avoid paying for a subscription. What is the MOST appropriate response under responsible use?

  1. AReproduce it, as long as the team removes the author's name afterward
  2. BDo not reproduce it; summarize lawfully obtained material or buy access Correct answer
  3. CReproduce it, provided the team keeps the text inside internal chats
  4. DReproduce it, because content generated by Claude is free of copyright

Answer and explanation: Reproducing copyrighted articles in full to avoid paying infringes the publisher's rights; lawful options are summarizing material you have access to or licensing it. AI output of protected text does not remove copyright. Removing the byline or keeping it internal does not make the copy lawful.

Question 9 of 10, Governance, Risk, and Responsible Use

A communications officer wants to paste an unreleased earnings announcement into Claude to polish the wording. What is the MOST appropriate action?

  1. ARemove the company name but keep all figures, then paste it into a free personal account
  2. BPaste it into any personal Claude account because the text will be public in a few days anyway
  3. CAsk a colleague to paste it instead so the communications officer's account is not involved
  4. DUse only an approved organization workspace and follow the company's MNPI rules Correct answer

Answer and explanation: Material non-public information must stay within approved tools and follow disclosure rules until it is released. Upcoming publication does not make early sharing acceptable. Removing the name leaves the sensitive figures. Using someone else's account does not change the policy breach.

Question 10 of 10, Troubleshooting and Optimization

In a long chat, a user corrected Claude's use of the wrong product name several times, but the error keeps returning. What is the BEST approach?

  1. ARename the product so it matches the name that Claude keeps using
  2. BStart a fresh chat with the correct name stated clearly at the start Correct answer
  3. CAdd more messages so the early wrong usage is pushed further back
  4. DKeep correcting it in the same chat until Claude finally stops repeating the wrong name

Answer and explanation: Early errors in a long conversation keep influencing later replies; a fresh chat with the correct term stated up front, or set in project instructions, removes the mixed signals. Repeated corrections keep both names in context. Renaming the product is not a fix. Adding messages grows the problem.