Domain 2: Prompt Engineering Fundamentals

Claude Certified Associate - Foundations · 66 questions

• Which of the following BEST defines prompt engineering?
• What is the PRIMARY purpose of a system prompt in the Claude API?
• A developer sets temperature=0 when calling the Claude API. What effect does this have on Claude's responses?
• What is 'few-shot prompting'?
• What is chain-of-thought (CoT) prompting and when is it most beneficial?
• A developer wants Claude to always return responses in valid JSON format. What is the MOST reliable approach?
• The 'topp' parameter in the Claude API is used for nucleus sampling. What does setting topp=0.9 mean?
• What is 'role prompting' and how does it affect Claude's responses?
• If an operator instructs Claude in the system prompt to keep the system prompt confidential, how should Claude respond if a user asks to see it?
• What is 'grounding' or 'context stuffing' in prompt engineering?
• What is 'prefilling' Claude's response in the API?
• When should a developer consider fine-tuning a model versus using prompt engineering to achieve the desired behavior?
• What is the key difference between instructions placed in the system prompt versus the first user message?
• Which statement about prompt length is MOST accurate?
• Research suggests that telling Claude what NOT to do ('Don't include lengthy preambles') is generally less effective than what alternative approach?
• When using Claude with a very long context (e.g., a 100K token document), where should critical instructions be placed for maximum effectiveness?
• When presenting Claude with a complex, multi-part task, what prompting strategy generally produces the best results?
• For a sentiment classification task (positive/negative/neutral), which prompt structure gives the MOST reliable and parseable output?
• For a data extraction task (extracting names, dates, and amounts from invoices), what prompt approach MAXIMIZES reliability?
• When using prompt templates with variable substitution (e.g., filling in {{CUSTOMERNAME}} placeholders), what security concern should developers consider?
• A developer finds that zero-shot prompting (instructions only, no examples) produces inconsistent classification results. What is the BEST next step?
• To extract action items from a meeting transcript, which prompt approach is MOST effective?
• In a multi-turn conversation, Claude can reference information shared earlier. What enables this capability?
• To get the most diverse and creative brainstorm of product name ideas from Claude, which approach works BEST?
• What is the relationship between temperature and topk in language model sampling?
• A company is making 1 million Claude API calls per month with a 500-token system prompt. What would have the BIGGEST impact on reducing their monthly API costs?
• Claude tends to use markdown formatting (headers, bullet points, bold) by default in many responses. When is this PROBLEMATIC?
• What prompting technique is used when a developer includes 'Explain this to a non-technical executive with no programming background' in a prompt?
• A team wants to systematically improve their Claude-powered application over time. What evaluation practice should they establish?
• An enterprise wants Claude to always refuse to discuss internal salary bands. Where should this policy primarily live?
• When writing few-shot examples for classification, which practice improves robustness?
• Why do many Claude prompts use XML-like tags such as <document> and <instructions>?
• You need guaranteed JSON fields from Claude for a downstream parser. Which approach is MOST reliable?
• What is a valid use of assistant prefilling with Claude?
• How can you improve factual grounding when Claude answers from retrieved docs?
• After changing a production prompt, what is the BEST next step?
• When should you ask Claude to reason step-by-step?
• Your conversation history approaches the context limit. What is a sound strategy?
• What is a limitation of role prompting ('You are a senior lawyer...')?
• Why separate operator/system instructions from end-user messages?
• Why include enums and required fields in tool JSON Schemas?
• When is RAG preferable to stuffing many few-shot examples into every prompt?
• Why set maxtokens thoughtfully rather than always maximum?
• What is a downside of stuffing entire knowledge bases into every prompt?
• To encourage JSON object output, a developer prefills assistant content with '{'. This technique:
• To reduce sycophantic agreement with incorrect user claims, prompts should:
• Production prompt templates should be:
• Even with large context windows, best practice is to:
• For creative brainstorming, teams often:
• Token-efficient prompting means:
• Compared to 'please reply in JSON', tool use is often better because:
• Using clear delimiters around untrusted user content helps because:
• Best-practice document QA with Claude includes:
• Regarding chain-of-thought / extended thinking content, products should:
• The best way to compare two system prompts in production is:
• A 2s chat latency budget should account for:
• To improve tool argument reliability:
• Relying on a single accuracy number for LLM evals is risky because:
• Ideal content to place in a cacheable prefix includes:
• Even with structured outputs, production systems should:
• Before upgrading production from Sonnet N to Sonnet N+1:
• Assigning per-feature token budgets helps:
• Capturing user thumbs-down on Claude answers enables:
• Pinning specific model IDs in production is recommended to:
• Conversation summarization for long sessions should:
• To keep answers concise for mobile UX: