1/119
Looks like no tags are added yet.
Question 1: What should a privacy reviewer recommend when processing PHI caregiver notes with ZDR enabled?
Verify ZDR coverage and HIPAA readiness for the actual product path before processing the notes.
Question 1 (Explanation): Does Zero Data Retention apply uniformly across all Claude products in an organization?
No, ZDR is feature-specific and does not cover every product path.
Question 1 (Explanation): Can you rely on a prompt instruction asking Claude not to retain notes for PHI compliance?
No, privacy and retention requirements must be satisfied through the applicable product arrangement.
Question 2: Which two design choices best support a brokerage research workflow using Claude for listing notes?
Keep listing permissions in the system of record and connect Claude to the listing workflow only when availability is needed.
Question 2 (Explanation): Why is bulk exporting the full listing database into Claude a poor workflow design?
It bypasses the controlled source and exposes more information than necessary.
Question 3: What is the most appropriate action when refining a prompt for college financial aid email responses?
Define measurable audience criteria, add example model replies, and compare prompt versions on representative cases.
Question 3 (Explanation): Why should you avoid selecting a single polished draft as a permanent prompt template?
A single draft may not represent the broad range of applicant questions the office will receive.
Question 4: What is the most appropriate next step when prototyping a Claude review process for restaurant inspection notes?
Compare Claude's reviews across varied notes using measurable success criteria.
Question 4 (Explanation): Why is approving a process based on one good review inadequate?
A single case can miss edge cases, inconsistent reasoning, or tone problems.
Question 5: Which model-selection approach best fits a restaurant menu team needing frequent polished drafting with moderate constraints?
Use Claude Sonnet 5 as the default model for this balanced workload.
Question 5 (Explanation): Why is Claude Opus not recommended as the default for frequent polished drafting workloads?
It is positioned for complex agentic coding and enterprise work, adding unnecessary cost and latency.
Question 6: Which feature choice best fits a university advancement office coordinating grant reviews over several weeks across multiple chats?
Create a shared Project with materials in Project Knowledge and grant reviewers Can use access.
Question 6 (Explanation): Why is using a regular chat and re-uploading funder files for each new thread inefficient?
It creates avoidable setup work and does not address shared access control across conversations.
Question 7: What is the most appropriate next step when a medical device team's pilot memos miss evidence or use the wrong tone?
Define measurable success criteria and test representative cases before changing prompts or models.
Question 7 (Explanation): Why should you avoid asking Claude to score each memo to keep high self-ratings?
Self-ratings are not a reliable substitute for a tested evaluation process against external criteria.
Question 8: What is the most appropriate validation plan for a marine insurance claims team extracting loss dates and coverage rationale?
Use automated checks for objective fields and require a specialist to review coverage rationale before decisions.
Question 8 (Explanation): Why are citations alone insufficient for validating high-stakes insurance coverage decisions?
Citations help identify source passages, but they do not guarantee the conclusion or interpretation is correct.
Question 9: What prompting approach is most appropriate when a research group needs to understand where multiple market reports agree or conflict?
Ask Claude to extract key claims, note source conflicts, and organize findings by theme.
Question 9 (Explanation): Why is asking Claude to brainstorm taglines inappropriate for analyzing multiple uploaded market sources?
Brainstorming is suited for creative ideation, not for synthesizing and understanding evidence across sources.
Question 10: What is the most appropriate action when an animal welfare adoption team's pet profile prompts produce overly confident behavioral claims?
Define audience-specific success criteria, add clear tone guidance and examples, and compare prompt versions on representative cases.
Question 10 (Explanation): Why are negative instructions alone a weak way to steer style in public pet profiles?
Negative instructions are weaker for steering style than explicit guidance and examples of what the output should do.
Question 11: Which prompt change would most directly improve a public radio team's sponsor-read script review workflow?
Ask Claude to rewrite the script directly and provide the revised version in the required review format.
Question 11 (Explanation): Why do requests like 'suggest ways to improve this draft' fail to yield a ready-to-review script?
They invite feedback and critique rather than directing Claude to execute a direct rewrite.
Question 12: What is the most appropriate next step when a credit union prompt revision catches more escalation cues but increases speculation and latency?
Compare the revised prompt against the original on the same eval set with separate criteria for accuracy, speculation, and latency.
Question 12 (Explanation): Why is deploying a revised prompt immediately after a single recent improvement a risk?
It overweights recent feedback and ignores newly observed regressions like speculation and slower turnaround.
Question 13: What prompting approach is most appropriate when a coffee roaster needs a customer subscriber notice matching a warm voice without unapproved promises?
Provide approved sample notices, define the audience, specify constraints, and request the final format.
Question 13 (Explanation): Why should you avoid asking Claude to write a customer notice in its best style without examples?
It relies on vague style guidance instead of providing concrete examples, format constraints, and boundaries.
Question 14: Which TWO practices should a book publisher use when writing system-level instructions for author update emails?
Define the standing tone, role, boundaries, and output structure in instructions, and rely on current chat materials for changing facts.
Question 14 (Explanation): Why should fast-changing facts like launch dates or contract details be excluded from system instructions?
They become stale quickly and are a poor fit for standing instructions that should remain stable.
Question 15: What is the most appropriate next step when a Google Workspace Drive option is unavailable for a shared Claude Project?
Upload an exported copy of the document to the shared project knowledge as a workaround.
Question 19: What should the team do when prompt outputs still miss learning objectives or include questions that are too easy?
Define task-specific success criteria, build an eval set with edge cases, and compare prompt versions against it.
Question 19 Explanation: Why is defining task-specific success criteria better than informal prompt edit cycles?
Memory makes comparisons inconsistent, whereas explicit criteria and repeatable tests reveal failure modes systematically.
Question 19 Explanation: What is the recommended mental model for prompt troubleshooting?
First define what good output means, test against a stable set of cases, and revise based on observed failures.
Question 20: Which model selection approach best fits a recurring, business-critical claims review workflow needing a balance of reasoning, latency, and cost?
Use Claude Sonnet 5 as the default model for the daily claims workflow.
Question 20 Explanation: Why is Claude Sonnet 5 the best fit for recurring claims analytics?
It avoids treating the task as a simple lowest-cost batch job or a highest-complexity reasoning task, balancing speed and intelligence.
Question 20 Explanation: When should Claude Haiku 4.5 be used instead of Sonnet 5?
Haiku 4.5 is well suited to quick iteration, tight latency, cost-sensitive use, and high-volume straightforward tasks.
Question 21: Which model-selection approach best fits a telehealth scheduling team's repetitive message classification and short drafting?
Use Claude Haiku 4.5 for the routine classifications and short drafts.
Question 21 Explanation: Why is Haiku 4.5 chosen for repetitive telehealth scheduling messages?
Speed and cost matter more than deep reasoning for short, routine inputs, making the fastest near-frontier model ideal.
Question 22: What action would best optimize live actuarial pricing reviews crowded out by routine policy classifications?
Process routine classifications offline and reserve the live review for edge cases.
Question 22 Explanation: Why move routine policy classifications offline during actuarial reviews?
It reduces meeting load, removes bottlenecks, and leaves time for judgment-heavy discussion on edge cases.
Question 23: What is the most appropriate next step for a specialty pharmacy iterating medication counseling prompts across many transcript variations?
Create a structured evaluation set with varied examples and measurable success criteria.
Question 23 Explanation: Why test pharmacy prompt revisions across multiple examples instead of a single fixed one?
A single transcript cannot reveal whether a workflow handles edge cases, privacy constraints, or tone requirements consistently.
Question 24: What is the most appropriate next step when the Gmail connector finds email threads but needs to summarize terms inside PDF attachments?
Add the PDF attachments as uploaded files or Drive files before requesting the clause summary.
Question 24 Explanation: Why must PDF attachments be uploaded separately when using the Gmail connector?
The Gmail connector can find and read email content, but attachment contents are not directly accessible through Gmail.
Question 25: What is the most appropriate governance step before publishing an article drafted by Claude from interview notes and public records?
Have a qualified editorial professional review the draft and disclose the AI involvement to readers.
Question 25 Explanation: Why does published journalistic content require qualified human review and AI disclosure?
External publishing is a high-risk use case requiring editorial accountability and transparency for readers regarding AI involvement.
Question 26: What should an animal shelter's adoption team do when staff keep changing prompts based on the latest complaint without empirical proof?
Define measurable criteria and compare prompt versions on representative cases.
Question 26 Explanation: Why is directly applying every staff comment to a prompt considered an anti-pattern?
It can improve one example while making other cases worse and perpetuates an ad hoc cycle.
Question 27: What should a municipal transit agency do to reliably improve issue ticket quality from rider complaints?
Define measurable ticket-quality criteria, run representative eval cases, then compare prompt revisions against the same cases.
Question 27 Explanation: Why is rewriting a prompt around the latest complaint a weak approach?
It can overfit the prompt to one recent example without proving it handles real distribution or edge cases.
Question 28: What should a veterinary diagnostics lab do to improve client-facing summary outputs that miss caveats or use technical phrasing?
Define measurable success criteria, build representative eval cases, and compare prompt revisions against them.
Question 28 Explanation: What is the practical mental model for prompt iteration in the veterinary lab workflow?
Test, refine, and validate using the same eval set to reveal both gains and regressions.
Question 29: What should a municipal water utility team do to decide whether their outage notice and billing workflow is reliable for customer agents?
Create a representative eval set, score drafts against measurable criteria, and verify fee claims against source materials.
Question 29 Explanation: Why are citations alone insufficient for verifying customer service billing drafts?
Citations help ground responses, but critical details like fees still need direct verification against source material to prevent outdated info.
Question 30: What is the most appropriate action before using a volunteer shift feedback summary that makes an unsupported claim about student status?
Check the notes for quoted support, remove unsupported group claims, and revise the summary.
Question 30 Explanation: Why must unsupported group claims be removed from volunteer feedback summaries?
It prevents Claude from hallucinating or generalizing categories not established in the source data.
Question 31: What is the most appropriate way for a wildlife rescue center to evaluate treatment log release-readiness summaries before routine use?
Create a representative eval set with edge cases, then compare summaries against source logs using specific criteria for factual support and omissions.
Question 31 Explanation: Why are readability and tone scores insufficient for evaluating veterinary summaries?
They do not establish whether summaries preserve critical facts like medications, restrictions, and unresolved observations.
Question 32: What is the most appropriate recommendation for a digital publishing editor whose chat contains weeks of old drafts and resolved comments?
Use a Project for the stable style guide and start fresh reviews with a concise current summary.
Question 32 Explanation: Why should an editor move from a long-running chat to a Project with a concise summary?
Long chats accumulate clutter and outdated details that compete with current manuscript needs and exceed bounded context utility.
Question 1: Inappropriate mortgage AI use
Using Claude outputs at scale to train a separate mortgage-advice model without Anthropic's prior authorization.
Question 1: Prohibited AI training conduct
Model scraping or model distillation.
Question 1: Agentic outreach restriction
Impersonation to collect private offers is consistent with prohibited patterns like social engineering, fraud, or unauthorized data collection.
Question 1: High-risk mortgage summaries safeguard
Requires review by a qualified professional before finalization and required AI disclosure.
Question 2: Monthly synthesis model selection
Start on Claude Sonnet 5, then adjust effort upward if quality is insufficient.
Question 2: Daily ranger-note triage model selection
Select Claude Haiku 4.5 for short daily triage where speed and low cost matter most.
Question 2: Model selection tradeoff factors
Tradeoff among capability, latency, and cost.
Question 3: Regulated personal information deployment check
Confirm the documented retention controls for the approved Claude API deployment before submitting regulated reports.
Question 3: API data retention option
Zero Data Retention where applicable.
Question 3: Incorrect privacy check assumption
Relying on the model's privacy-oriented training without additional checks.
Question 4: Real estate advisory rollout message
Position Claude as a model-specific assistant for drafting and review, with validation before stakeholder use.
Question 4: Claude characteristic variations
Context windows, maximum output, latency, pricing, and knowledge cutoff.
Question 4: Reducing hallucination limitations
Grounding, quotes, citations, and uncertainty can reduce but not eliminate hallucinations.
Question 5: Claims analytics privacy requirements
Verify the required retention arrangement for the intended Claude deployment and use only the product surface approved for the required data controls.
Question 5: Anti-pattern for data retention
Entering notes first and requesting deletion in the prompt text.
Question 6: Coaching roaster prompt structure
Separate materials with descriptive tags, list review steps in order, and define coaching summary fields.
Question 6: Complex prompting guidance
Explicit instructions, numbered lists, and XML-style tags to separate instructions, context, examples, and inputs.
Question 7: Museum exhibit description refinement action
Add grade-level examples and score prompt versions against measurable tone and clarity criteria.
Question 8: Museum provenance recommendation action
Check the cited source passages and route the recommendation to a qualified provenance reviewer.
Question 8: Use of citations in Claude
Provide passages or locations that a reviewer can inspect, but do not make conclusions automatically reliable.
Question 9: Pediatric therapy intake notes action
Check the claim against the notes and revise unsupported group-level conclusions.
Question 10: Support chat governance requirement
Disclose at the beginning of each chat that users are interacting with AI.
Question 10: Consumer chatbot transparency obligation
Applies even when the chatbot is not otherwise a high-risk use case.
Question 11: Learner success model selection approach
Benchmark Claude Haiku 4.5 and Claude Sonnet 5 on representative posts, then choose based on quality, latency, and cost.
Question 12: Airport manual checklist prompt revisions
Place manual content before the task request, instructions, and examples, and ask Claude to quote relevant manual passages first.
Question 13: Renewable energy outage notes iteration action
Test each prompt version on representative notes against specific success criteria.
Question 14: Boutique hotel guest review tracker format
Ask Claude to return a compact table or CSV with one row per excerpt.
Question 15: Performing arts venue rollout message
Present Claude as a task-support tool, choose model and capabilities for the workflow, and review important facts before distribution.
Question 16: Agricultural cooperative vendor follow-up practice
Create a Skill for the follow-up note format, verify inventory details in the approved system, and send only after review.
Question 17: Maritime insurance project configuration actions
Add the policy wording and escalation checklist to the project knowledge base, and set project instructions that describe the required response format.
A freight railroad's safety office uses Claude to summarize incident reports. Before relying on the workflow, what should the office do?
Create an eval set with typical and edge cases, then score summaries against specific criteria for support and omissions.
Why is creating an eval set better than relying on citations or informal agreement for incident summaries?
Citations and polished writing do not prove that key facts were omitted or that inferences are correct.
Question 19: A museum archive wants Claude to create short accessibility descriptions for tens of thousands of scanned photos with quick turnaround and low cost. Which model approach fits best?
Select Claude Haiku 4.5 and validate outputs on a representative batch.
Why is Claude Haiku 4.5 chosen over Opus for high-volume museum archive image descriptions?
The workload is high volume, straightforward, and cost-sensitive, whereas Opus is reserved for complex reasoning.
Question 20: A nonprofit legal aid clinic wants consistent voice and response format while client facts change daily. Which approach is most appropriate for system-level instructions?
Define durable role, tone, boundaries, format, source use, and missing-information behavior, then provide case facts in the relevant chat.
Why should client-specific facts be kept out of global system-level instructions?
Global instructions can become stale and apply incorrectly to the wrong matter.
Question 21: A marketing team wants a range of campaign directions for a new rowing machine before final copy. Which prompting approach is appropriate?
Ask Claude for a varied idea list, with audience, constraints, and a simple format for comparing concepts.
Why is asking for a compliance checklist a poor prompting choice for early brainstorming?
It applies an analysis workflow before the team has developed enough ideas to assess.
Question 22: A renewable energy cooperative turns technician notes into customer outage explanations with mixed results. What two actions should the team take next?
Test the draft workflow on representative technician notes, including edge cases, and define measurable acceptance criteria.
Why should prompt revisions not be based solely on stakeholder preference of demo outputs?
Preferred wording from a few demos is not an empirical readiness test and misses factual or safety requirements.