Chapter 15 Quiz
Test your understanding of Claude's design philosophy, strengths, limitations, and best practices.
Question 1
What is Constitutional AI, and why does it matter for how Claude behaves?
A) A system that makes Claude follow the laws of whichever country it is used in B) A training approach where Claude evaluates its own outputs against a set of principles, producing behaviors like reduced sycophancy and calibrated uncertainty C) A constitutional framework that OpenAI and Anthropic jointly developed for AI safety D) A technique for making Claude more politically neutral
Answer
**B — A training approach where Claude evaluates its own outputs against a set of principles, producing behaviors like reduced sycophancy and calibrated uncertainty** Constitutional AI (CA) is Anthropic's published approach to training Claude. Rather than relying solely on human rater feedback (as in RLHF), CA gives the model a set of principles and trains it to evaluate and improve its own outputs against those principles. This produces observable behaviors: Claude is more likely to express genuine uncertainty, more likely to disagree and push back, and less prone to sycophantic validation than models trained primarily through user approval feedback.Question 2
You need to analyze an 80-page contract. Which Claude model should you use for the best results?
A) Claude Haiku, because it is fastest and the task is mostly reading B) Claude Sonnet, which balances capability and cost for complex professional tasks C) Claude Opus, because long document analysis with nuanced judgment benefits from maximum capability D) Any model works equally well since they all have the same context window
Answer
**C — Claude Opus, because long document analysis with nuanced judgment benefits from maximum capability** For complex analytical tasks requiring careful reading, judgment about which clauses are unusual, and nuanced professional assessment, Opus's maximum capability is worth the additional cost. Haiku is designed for simpler, high-volume tasks and is not appropriate for nuanced contract analysis. Sonnet is a reasonable choice for budget-constrained professional use, but Opus produces noticeably better results on tasks requiring deep, careful judgment. All models share similar context window sizes, but capability in using that context effectively differs.Question 3
What is the primary advantage of using XML tags in Claude prompts?
A) Claude reads XML faster, making responses quicker B) XML tags prevent Claude from refusing requests C) XML tags eliminate ambiguity about which parts of the prompt are data, instructions, context, and format requirements D) XML tags are required by the Claude API
Answer
**C — XML tags eliminate ambiguity about which parts of the prompt are data, instructions, context, and format requirements** Claude's training makes it particularly responsive to XML structure because it uses the tags to understand what type of content each section represents. Without tags, Claude must infer which part of a long prompt is the document to analyze versus the instructions for analyzing it versus the background context — and can occasionally misread the structure. XML tagging makes this explicit, reducing errors and improving output organization for complex multi-component prompts.Question 4
Claude adds significant caveats to your request for information about pharmaceutical drug interactions, even though you are a pharmacist with a legitimate professional need. What is the most effective response?
A) Switch to ChatGPT, which will not add these caveats B) Push back harder — tell Claude to stop adding disclaimers C) Rephrase with explicit professional context: "As a licensed pharmacist reviewing patient counseling materials, I need..." D) Accept the caveats as unavoidable — Claude is programmed this way
Answer
**C — Rephrase with explicit professional context: "As a licensed pharmacist reviewing patient counseling materials, I need..."** Claude's caution is calibrated to how requests are framed, not to permanent topic blocks. Adding professional context, stating your purpose, and making the legitimate use case explicit almost always resolves over-caveating on professional queries. Pushing back with pressure tends to be counterproductive — Claude reads pressure as a signal that something unusual is being requested. Switching tools is unnecessary when rephrasing works. Caveats are not a permanent feature; they respond to context.Question 5
What does the Claude Projects feature allow you to do?
A) Share your Claude conversations publicly B) Build custom GPTs similar to ChatGPT's GPT marketplace C) Create persistent project workspaces with shared context and uploaded knowledge files that apply across multiple conversations D) Run automated tasks on a schedule
Answer
**C — Create persistent project workspaces with shared context and uploaded knowledge files that apply across multiple conversations** Claude Projects allows you to create named workspaces where uploaded documents and shared context persist across all conversations within the project. This eliminates the need to re-upload background materials each session and maintains continuity on long-running work. Projects are particularly valuable for consulting engagements, research projects, and any professional work that spans multiple sessions with consistent reference materials.Question 6
You ask Claude for an analysis of a complex strategy document, and the response covers all the details well but does not provide a clear overall conclusion. What is the most direct fix?
A) Switch to Opus model, which provides better synthesis B) Explicitly ask Claude for synthesis: "Given everything you have identified, what is the single most important conclusion for decision-making?" C) Ask Claude to start over with a different approach D) Claude is unable to synthesize — use a separate tool for that step
Answer
**B — Explicitly ask Claude for synthesis: "Given everything you have identified, what is the single most important conclusion for decision-making?"** Claude's tendency to cover all components thoroughly without synthesizing them into a clear overall conclusion is addressed by explicitly asking for that synthesis. This is not a model limitation — Claude can synthesize well when asked to. The issue is that the initial prompt did not request synthesis, so Claude delivered a thorough but unsynthesized analysis. Adding a follow-up or including synthesis in the original prompt resolves this consistently.Question 7
For which of the following tasks does Claude have a clear advantage over ChatGPT?
A) Generating images for a marketing campaign B) Browsing the web for current information on a recent news event C) Analyzing a 150-page regulatory document for key compliance requirements D) Running Python code against an uploaded dataset
Answer
**C — Analyzing a 150-page regulatory document for key compliance requirements** Claude's combination of a very large context window, effective use of long context, careful reading, and willingness to surface inconvenient findings makes it the better choice for long document analysis. ChatGPT generates images (DALL·E) and Claude does not. ChatGPT has web browsing integration that Claude's standard interface lacks. ChatGPT's Code Interpreter runs Python code against files; Claude can discuss and write code but does not natively execute it against uploaded files.Question 8
What is "reduced sycophancy" as a Claude characteristic, and why is it valuable for professional work?
A) Claude refuses to agree with user requests, making it stubborn B) Claude tends to maintain positions under pressure and surface unsolicited critique, making it a more honest thinking partner C) Claude uses less formal language than other models D) Claude is trained to avoid all forms of compliments or positive feedback
Answer
**B — Claude tends to maintain positions under pressure and surface unsolicited critique, making it a more honest thinking partner** Reduced sycophancy means Claude is less likely to tell you what you want to hear at the expense of accuracy and honesty. It will maintain a disagreeing position if you push back, offer critical observations about your work without being asked, and note weaknesses in your plans alongside strengths. This is valuable for strategy, planning, writing, and any work where honest feedback improves outcomes. It does not mean Claude is negative or contrarian — it means its feedback is calibrated to accuracy rather than to user satisfaction.Question 9
You are asking Claude to analyze a document that is over 100,000 words. What step should you take before beginning analysis?
A) Break the document into sections and upload each separately B) Ask Claude to confirm it has received the complete document and identify specific landmarks in it before beginning analysis C) Switch to GPT-4o, which handles very long documents better D) Just proceed — Claude will manage long documents automatically
Answer
**B — Ask Claude to confirm it has received the complete document and identify specific landmarks in it before beginning analysis** Confirming receipt of a very long document — by asking Claude to name the last section or identify a landmark in the middle of the document — catches truncation issues before you invest time in analysis of incomplete content. Proceeding without confirmation risks receiving a thorough analysis of an incomplete document, which can be more misleading than no analysis. Claude handles very long documents better than most alternatives, making switching unnecessary.Question 10
Which of the following best describes the difference between Claude Haiku and Claude Opus?
A) Haiku generates only short responses; Opus generates longer ones B) Haiku is Anthropic's safest model; Opus has fewer content restrictions C) Haiku is optimized for speed and cost on simpler tasks; Opus is Anthropic's most capable model for complex, judgment-intensive tasks D) They are the same model with different names for different markets
Answer
**C — Haiku is optimized for speed and cost on simpler tasks; Opus is Anthropic's most capable model for complex, judgment-intensive tasks** Anthropic's model family is tiered by capability and cost. Haiku is designed for high-volume, lower-complexity tasks where latency and cost matter (classification, simple Q&A, data extraction). Sonnet is the balance model for most professional use. Opus provides maximum reasoning and instruction-following capability for tasks where quality matters more than speed or cost. The differences are about capability and optimization, not content restrictions or output length.Question 11
What is the recommended approach when working on a very long writing project with Claude — for example, a 10,000-word strategy report?
A) Send one large prompt requesting the complete report in a single response B) Break the project into phases (outline, section-by-section drafting, integration review) with separate requests for each C) Use the largest context window model and write everything in one session D) Use ChatGPT for long projects since it has better length handling
Answer
**B — Break the project into phases (outline, section-by-section drafting, integration review) with separate requests for each** Section-by-section drafting produces better individual sections, allows course correction between sections, prevents context degradation that can affect very long single-session outputs, and gives you natural checkpoints to validate the document is developing as planned. The outline phase is particularly valuable because it catches structural problems cheaply. An integration review at the end ensures the sections cohere. This disciplined approach consistently outperforms single-prompt requests for long documents on any AI model.Question 12
Claude expresses that it is "uncertain" about a specific statistic in its response. What should you do with that claim?
A) Discard the entire response since Claude acknowledged uncertainty B) Use Claude's expressed uncertainty as a signal to prioritize verification of that specific claim C) Ask Claude again — it will be more confident the second time D) Accept it anyway since Claude's uncertainty expressions are just boilerplate
Answer
**B — Use Claude's expressed uncertainty as a signal to prioritize verification of that specific claim** Claude's confidence calibration is a genuine feature, not boilerplate. When Claude explicitly flags uncertainty, it is providing actionable guidance about where to focus verification effort. This is more useful than models that present everything with equal confidence. It does not mean you should discard the uncertain claim entirely — it means you should verify it before relying on it. Claude's confident claims still warrant verification for high-stakes use, but uncertainty flags are your highest-priority verification targets.Question 13
You are building an application on the Claude API and want Claude to express explicit uncertainty when it does not know something, rather than generating confident-sounding answers. Where and how should you configure this behavior?
A) You cannot configure this — Claude handles uncertainty automatically B) In the user prompt for each individual request C) In the system prompt, with explicit instructions about how to handle uncertainty D) In the API request parameters using the "uncertainty_mode" flag
Answer
**C — In the system prompt, with explicit instructions about how to handle uncertainty** System prompt configuration is the appropriate place for persistent behavioral rules. Include explicit instructions like: "When you encounter a question where you are uncertain, say so explicitly. Use phrases like 'I'm not confident about this specific point — I'd recommend verifying' rather than generating a confident-sounding answer to fill gaps." There is no API parameter for uncertainty mode. User-prompt-level instructions are inconsistent and require repetition every call. Claude does handle uncertainty somewhat automatically, but explicit system prompt instructions make the behavior more consistent and specific to your application's needs.Question 14
Which statement about Claude's image capabilities is accurate?
A) Claude can generate images using DALL·E integration B) Claude can analyze images you share with it but cannot generate images C) Claude can generate and analyze images through its Projects feature D) Claude has no image capabilities whatsoever
Answer
**B — Claude can analyze images you share with it but cannot generate images** Claude has vision capability — it can receive and analyze images, describe what it sees, respond to questions about image content, and interpret diagrams and charts. However, Claude does not generate images. There is no equivalent of DALL·E in Claude's feature set. For image generation needs, users must use a different tool (ChatGPT/DALL·E, Midjourney, Adobe Firefly, etc.).Question 15
You ask Claude to give you feedback on your business proposal, and it responds with thorough critical analysis identifying three significant weaknesses. You are convinced your proposal is actually strong and the critique reflects misunderstanding. What is the appropriate next step?
A) Dismiss Claude's critique — it doesn't understand your business B) Accept all of Claude's critique since it has reduced sycophancy and is therefore always right C) Respond with the specific context that would address Claude's misunderstandings and ask whether its critique still holds given that context D) Switch to ChatGPT, which will give you more validating feedback