Medium CCAO-F practice questions

Your startup is preparing to launch a consumer-facing wellness app built on Claude that offers general lifestyle and stress-management tips. During a pre-launch review, the product manager suggests marketing it as a tool that can 'diagnose mental health conditions and recommend prescription treatments' to differentiate from competitors. As the responsible-deployment lead, what is the most appropriate action aligned with Anthropic's usage policies and a responsible-deployment mindset?

A team is building an agent that manages a customer's calendar. The agent can create, move, and delete events by calling tools. During testing, the agent frequently double-books slots because it decides which action to take next without ever incorporating the results of its previous tool calls. What is the most important design change to make this agent work reliably?

A product team wants to build a feature that investigates customer complaints. Each complaint requires an unpredictable number of steps: reading the ticket, searching internal knowledge bases, querying order histories, and sometimes escalating to check shipping partner APIs — but the exact sequence and which tools are needed vary widely from case to case and cannot be known in advance. The team asks whether to implement this as a fixed workflow or as an agent. Which approach best fits this task, and why?

You are building an agent that manages calendar events. Early testing shows the agent frequently picks the wrong action because you exposed 14 narrowly-scoped tools (e.g., create_morning_event, create_afternoon_event, create_recurring_event, create_all_day_event, and similar variants). The model often confuses which tool applies. Following best practices for giving agents tools, what is the most effective change to improve reliability?

Your team is deploying an autonomous agent that manages cloud infrastructure by calling tools to provision, modify, and delete resources. During internal testing the agent occasionally deletes production databases when a user's request is ambiguous. The team wants the agent to remain capable of independent multi-step action but reduce the risk of catastrophic, irreversible operations. Which change best addresses this while preserving the agent's autonomy?

A logistics company is building a system to reroute shipments during weather disruptions. Each disruption is different: the number of affected shipments, the alternate carriers available, and the sequence of checks needed vary widely and cannot be predicted in advance. The engineering team wants the system to figure out its own sequence of steps at runtime, calling tools to check carrier availability, inventory, and costs as needed until it produces a viable rerouting plan. Which implementation approach best fits these requirements?

You are building an agent that books meeting rooms by calling a scheduling tool. Currently, when a requested room is unavailable, the tool returns only 'false'. The agent frequently loops, re-requesting the same unavailable room, wasting tokens and time. Without changing the underlying booking logic, what is the most effective way to improve the agent's reliability?

A team is building a coding assistant agent. They give it a file-search tool, a code-linter tool, and a git-commit tool. The lead asks the team to categorize each tool by whether it primarily lets the agent take actions that change the external environment or primarily gives the agent information to reason with. Which categorization is correct?

Your team is building an internal assistant to handle IT support requests. Some requests are simple and always follow the same steps (e.g., 'reset my password' → verify identity → trigger reset → confirm). Others are open-ended troubleshooting problems where the resolution path is unknown up front and may require querying logs, checking system status, and deciding which tool to call next based on what earlier calls return. Which design best fits these two categories of requests?

A team is building an automated system to resolve customer refund requests. Some requests are simple (one lookup, apply refund), while others require an unpredictable number of steps: checking order history, verifying return shipments, escalating to fraud checks, or requesting more information from the customer. The team cannot predetermine how many steps or which order they will occur for any given request. Which implementation approach best fits this situation?

A team builds a Claude-powered internal help desk that answers questions from a company knowledge base. During testing, they notice that when a user asks about a topic the knowledge base does not cover, Claude confidently invents a plausible-sounding but incorrect policy. The retrieval step is working correctly and only returns relevant passages. Which change to the prompt would most directly reduce these fabricated answers?

A developer is building a feature that asks Claude to classify support tickets and return only a single category word (e.g., 'Billing'). They notice Claude sometimes prepends conversational text like 'Sure, the category is...' before the answer. Using only the Messages API structure, what is an effective technique to make Claude begin its response with the category directly?

You are building a customer support chatbot using the Messages API. To preserve context across turns, you send the full conversation history with each new request. In this history, what content should be placed in messages that use the 'assistant' role?

A support team deploys an agent that can call several tools (knowledge base search, ticket lookup, refund API) to resolve customer issues autonomously. During testing, a few conversations spiral: the agent repeatedly calls tools, racks up high token costs, and occasionally never returns a final answer. The team wants to keep the agent capable but reliable and bounded. Which change most directly addresses this problem?

A financial analyst at your firm is using Claude to evaluate whether loan applications meet a set of complex, interdependent eligibility rules. The model frequently returns a correct-sounding final verdict but occasionally reaches the wrong conclusion, and the analyst cannot tell where the reasoning went astray. Which prompt engineering technique should the analyst apply first to improve accuracy on this multi-step reasoning task?

A financial analyst wants Claude to evaluate whether loan applications meet a set of five specific eligibility rules. Early tests show Claude sometimes skips rules or gives a verdict without checking each one. The analyst wants to improve reasoning quality but keeps getting inconsistent results when simply saying 'think about it carefully.' Which chain-of-thought refinement is most likely to improve accuracy?

A financial analyst uses Claude to evaluate loan applications. She wants Claude to reason step-by-step through each application's risk factors before giving a final approval decision, but her downstream system only needs the final decision (APPROVE or DENY) so it can be parsed automatically. What is the best prompting approach to satisfy both needs?

Your team is building a legal document analysis feature that must reason through complex, multi-clause contracts and produce nuanced summaries. Accuracy on subtle legal reasoning is the top priority, and the feature runs asynchronously in the background, so response time is not a concern. Occasional per-request cost is acceptable given the low volume. Which model-selection approach best fits these requirements?

Your team runs a customer-support chat tool that classifies incoming messages into one of five intent categories. The task is simple and well-defined, but you process millions of messages per day and need responses in well under a second while keeping costs low. Accuracy on this narrow task is already excellent with smaller models. Which model-selection approach best fits these requirements?

A support engineer is writing a prompt for Claude to draft customer email replies. Their current prompt says: "Do not be too formal. Do not use jargon. Do not write more than three paragraphs." The replies are inconsistent — sometimes overly stiff, sometimes cluttered with technical terms. The engineer wants to make the instructions clearer and more reliable without adding examples yet. Applying the 'be clear and direct' technique, what is the single best change to the prompt?

A support engineer is building a prompt that asks Claude to process a customer refund request. The task involves verifying the order ID, checking the return window, calculating the refund amount, and drafting a customer email. In testing, Claude frequently drafts the email before performing the verification and calculation, and sometimes skips the return-window check entirely. The engineer wants to keep this as a single prompt. What is the most effective way to fix this behavior?

A product team asks Claude to "write documentation for our new API." The output is generic and pitched at a level that confuses their audience of non-technical business analysts. The team wants Claude to consistently produce docs the analysts can follow, without switching models or adding examples. Which single adjustment to the prompt best applies the clear-and-direct technique?

A support team asks Claude to "make this customer email better" and pastes a draft reply. The results are inconsistent — sometimes Claude shortens the email, sometimes it adds legal disclaimers, and sometimes it changes the tone entirely. The team wants more predictable output without adding examples or changing models. Applying the 'be clear and direct' technique, what is the best first change to the prompt?

A support team has finalized a prompt for summarizing customer tickets. They want to reuse the exact same prompt structure across hundreds of tickets, changing only the ticket text each time, and manage it centrally in the Anthropic Console. Which Console feature best supports this need?

A developer at a legal-tech company needs to analyze a single 280,000-token contract in one API call to answer questions about clauses scattered throughout the document. Their current integration truncates the document because it exceeds the model's input limit, causing incomplete answers. Which approach best addresses the root cause while keeping the analysis in a single request?