CAF PT1 Q01: Which company develops the Claude family of large language models?

CAF PT1 Q02 Answer C: Haiku

CAF PT1 Q03: A solutions architect must process a 300-page contract in a single request. Which…

CAF PT1 Q04: Roughly how many tokens correspond to 100 English words in typical text?

CAF PT1 Q05: Which input types can current Claude models natively accept in the Messages API?…

CAF PT1 Q06: What does the term 'multimodal' mean when applied to Claude?

CAF PT1 Q07: Which statement best describes a 'hallucination' in the context of Claude?

CAF PT1 Q08: A team needs Claude to answer questions about events from last week. What is the…

CAF PT1 Q09 Answer A: Task complexity and required reasoning depth

CAF PT1 Q10: What is 'extended thinking' in newer Claude models?

CAF PT1 Q11 Answer B: x-api-key

CAF PT1 Q12 Answer B: /v1/messages

CAF PT1 Q13 Answer B: max_tokens

CAF PT1 Q14: In the Messages API, which roles are valid inside the messages array?

CAF PT1 Q15: What does a stop_reason of 'max_tokens' indicate in a Messages API response?

CAF PT1 Q16: Which HTTP status codes from the Anthropic API should trigger retry logic with…

CAF PT1 Q17 Answer B: Users see output tokens as they are generated, dramatically improving…

CAF PT1 Q18 Answer B: In a server-side secret manager, with the backend proxying model calls

CAF PT1 Q19 Answer A: Truncate or summarize older conversation history before sending

CAF PT1 Q20: What is the purpose of the anthropic-version HTTP header?

CAF PT1 Q21: What is the primary function of a system prompt?

CAF PT1 Q23: What is few-shot prompting?

CAF PT1 Q22 Answer B: Tags clearly delimit different parts of the prompt, and Claude was trained...

CAF PT1 Q24 Answer B: Think through the problem step by step before giving your final answer

CAF PT1 Q25 Answer B: If the answer is not in the document, say you cannot find it rather than...

CAF PT1 Q26: What is 'prefilling' the assistant response in the Messages API?

CAF PT1 Q27: Which techniques help ensure Claude outputs valid JSON for downstream systems?...

CAF PT1 Q28: What is role prompting?

CAF PT1 Q29 Answer A: Documents near the top, with the specific question or instructions after them

CAF PT1 Q30 Answer B: Build a small evaluation set of representative inputs and iterate on the...

CAF PT1 Q31: What does 'tool use' (function calling) enable Claude to do?

CAF PT1 Q32 Answer B: A JSON Schema describing the tool's expected parameters

CAF PT1 Q33 Answer B: A user message containing a tool_result block referencing the tool_use id

CAF PT1 Q34: Which tool_choice setting forces Claude to call one specific named tool?

CAF PT1 Q35: What is the Model Context Protocol (MCP)?

CAF PT1 Q36: Which primitives can an MCP server expose to clients? (Select ALL that apply)

CAF PT1 Q37 Answer B: An agent runs a loop: it plans, calls tools, observes results, and iterates...

CAF PT1 Q38: In an agent loop, why is it important to set a maximum iteration count?

CAF PT1 Q39 Answer B: Return an error message in the tool_result (marking it as an error) so...

CAF PT1 Q40: Which use cases are natural fits for Claude tool use? (Select ALL that apply)

CAF PT1 Q41: What is Constitutional AI, the training approach associated with Claude?

CAF PT1 Q42: What is prompt injection?

CAF PT1 Q43 Answer A: Clearly delimit untrusted page content with tags and tell Claude to treat it..

CAF PT1 Q44 Answer B: Ensuring the deployment meets regulatory requirements (e.g., HIPAA)...

CAF PT1 Q45: By default, does Anthropic train its models on business customers' API data?

CAF PT1 Q46: What is the purpose of red teaming an LLM application before launch?

CAF PT1 Q47: Which controls form a sensible output-safety layer for a customer-facing Claude...

CAF PT1 Q48 Answer B: Because LLMs can err or be manipulated, and human review bounds the impact...

CAF PT1 Q49: What is a jailbreak attempt against Claude?

CAF PT1 Q50 Answer B: Enforce authorization in the tool layer so it only returns data the...

CAF PT1 Q51 Answer A: The Anthropic API (Claude Developer Platform)

CAF PT1 Q52: What is Retrieval-Augmented Generation (RAG)?

CAF PT1 Q53 Answer A: A vector database

CAF PT1 Q54: What does prompt caching optimize in Claude API workloads?

CAF PT1 Q55: Which workload is the BEST fit for the Message Batches API?

CAF PT1 Q56 Answer A: The context window cannot hold anywhere near that volume, and retrieval...

CAF PT1 Q57 Answer A: Token usage and cost per request

CAF PT1 Q58 Answer B: Retry with exponential backoff and, if needed, degrade gracefully to a...

CAF PT1 Q59 Answer B: To keep traffic within their AWS environment and reuse existing AWS IAM...

CAF PT1 Q60 Answer A: Streaming tokens to the UI as they generate

CAF PT2 Q01: What does the temperature parameter control in Claude's text generation?

CAF PT2 Q02 Answer A: temperature: 0

CAF PT2 Q03: What is the purpose of the stop_sequences parameter?

CAF PT2 Q04: Which tasks are good fits for Claude's vision capability? (Select ALL that apply)

CAF PT2 Q05 Answer B: Claude cannot generate images; it can, however, produce SVG markup or code...

CAF PT2 Q06 Answer B: Pinned versions keep behavior stable; silent model upgrades can change...

CAF PT2 Q07: Does Claude's context window budget include the generated output tokens?

CAF PT2 Q08: Is Claude suitable for non-English workloads?

CAF PT2 Q09: How does output token pricing compare to input token pricing for Claude models?

CAF PT2 Q10 Answer B: Documenting the model's capabilities, limitations, evaluations, and safety...

CAF PT2 Q11 Answer A: Python

CAF PT2 Q12 Answer B: content_block_delta

CAF PT2 Q13 Answer B: Use the API's token counting endpoint to get an exact count for the request

CAF PT2 Q14 Answer A: Requests per minute (RPM)

CAF PT2 Q15: An API call returns 401 authentication_error. What is the most likely cause?

CAF PT2 Q16: How does the Messages API handle conversation state across turns?

CAF PT2 Q17 Answer A: Base64-encoded image data in an image content block

CAF PT2 Q18: What is Anthropic's guidance regarding adjusting temperature and top_p together?

CAF PT2 Q19 Answer A: Use streaming so the connection stays active while tokens are delivered...

CAF PT2 Q20: Where in the Messages API response can you find how many tokens were consumed?

CAF PT2 Q21: Which prompt is likely to produce the MOST reliable results?

CAF PT2 Q22: Anthropic suggests a simple test for prompt clarity. What is it?

CAF PT2 Q23 Answer A: Examples should be diverse and representative, covering edge cases and all...

CAF PT2 Q24: Which phrasing style tends to work better in Claude prompts?

CAF PT2 Q25: What is prompt chaining?

CAF PT2 Q26: Why might you ask Claude to put its final answer inside [answer] tags?

CAF PT2 Q27 Answer A: Instruct Claude to respond directly without preamble or filler phrases

CAF PT2 Q28: Which practices improve the faithfulness of citations when Claude answers from...

CAF PT2 Q29 Answer B: Ask a targeted clarifying question when intent is ambiguous and the cost of...

CAF PT2 Q30 Answer B: It separates trusted instructions from untrusted input, aiding both...

CAF PT2 Q31 Answer B: A detailed, precise description explaining what the tool does, when to use...

CAF PT2 Q32: Can Claude request multiple tool calls in a single response turn?

CAF PT2 Q33: What does stop_reason 'tool_use' signify?

CAF PT2 Q34 Answer A: stdio - communication over standard input/output for locally spawned servers

CAF PT2 Q35: In MCP architecture, what is the role of the client?

CAF PT2 Q36: An agent working through a very long multi-step task is approaching its context...

CAF PT2 Q37: What is the orchestrator-workers (subagent) pattern in agent design?

CAF PT2 Q38 Answer B: No - use a conventional deterministic workflow; reserve agentic LLM loops...

CAF PT2 Q39 Answer A: Use descriptive parameter names with types, constraints, and descriptions in..

CAF PT2 Q40: What is Claude's 'computer use' capability?

CAF PT2 Q41 Answer B: Apply data minimization - send only fields necessary for the task and redact..

CAF PT2 Q42: What is the function of Anthropic's Usage Policy for API customers?

CAF PT2 Q43: What distinguishes INDIRECT prompt injection from direct prompt injection?

CAF PT2 Q44 Answer A: Adversarial tests with jailbreak and injection attempts

CAF PT2 Q45 Answer B: Include appropriate disclaimers, qualify limitations, and route users to...

CAF PT2 Q46 Answer A: Evaluate outputs across demographic slices to detect disparate patterns

CAF PT2 Q47 Answer A: The principle of least privilege - tools should carry only the minimum...

CAF PT2 Q48 Answer B: It enables incident investigation, compliance evidence, abuse detection, and..

CAF PT2 Q49 Answer B: Zero data retention arrangements or deployment channels that meet the...

CAF PT2 Q50 Answer A: Disclose clearly that users are interacting with an AI system

CAF PT2 Q51 Answer B: Anthropic does not offer a first-party embeddings API; teams typically use a..

CAF PT2 Q52: When chunking documents for a RAG system, which approach is generally recommended?

CAF PT2 Q53: What is the default lifetime of a prompt cache entry in the Anthropic API?

CAF PT2 Q54: For prompt caching to produce cache hits, what must be true about the requests?

CAF PT2 Q55 Answer A: Use prompt caching for repeated large prefixes

CAF PT2 Q56: What is a 'model router' in an LLM application architecture?

CAF PT2 Q57 Answer B: Hybrid retrieval covers both conceptual similarity and exact-term matches...

CAF PT2 Q58 Answer B: Implement retries, timeouts, circuit breakers, and a fallback path (e.g...

CAF PT2 Q59 Answer B: Prompts are versioned artifacts; changes run through automated evals in...

CAF PT2 Q60 Answer A: Centralized API key management and authentication

CAF PT3 Q01 Answer B: Haiku-class for the bulk tagging, a stronger model (Sonnet or Opus class)...

CAF PT3 Q02: When sending images to Claude, why does image size matter?

CAF PT3 Q03: What does the top_k sampling parameter do?

CAF PT3 Q04: What is the difference between the context window and max_tokens?

CAF PT3 Q05 Answer B: Give Claude a calculator or code-execution tool and have it delegate...

CAF PT3 Q06 Answer B: Recall over long contexts is generally strong but should be tested...

CAF PT3 Q07: How is an image in a prompt converted for billing purposes?

CAF PT3 Q08 Answer A: Report today's stock price for a company

CAF PT3 Q09 Answer B: Plan a migration: run your eval suite against the successor model, fix...

CAF PT3 Q10 Answer A: Better complex-task performance in exchange for higher token usage and latency

CAF PT3 Q11 Answer A: Messages must start with a user turn, and roles should alternate between...

CAF PT3 Q12: What is the general workflow of the Message Batches API?

CAF PT3 Q13: Why should each request in a Message Batch include a meaningful custom_id?

CAF PT3 Q14: Which practices make a robust streaming implementation? (Select ALL that apply)

CAF PT3 Q15 Answer A: The model string is misspelled, or references a model unavailable to your...

CAF PT3 Q16: How do Anthropic API rate limits typically evolve for a growing customer?

CAF PT3 Q17: What retry behavior do Anthropic's official SDKs provide out of the box?

CAF PT3 Q18 Answer B: Generous but bounded timeouts sized to expected generation length, with...

CAF PT3 Q19 Answer A: Use separate API keys per environment (dev, staging, prod) and per major...

CAF PT3 Q20: When is it appropriate to include the anthropic-beta header in requests?

CAF PT3 Q21 Answer A: Specify the expected tone and audience explicitly (formal, for corporate...

CAF PT3 Q22 Answer A: Exhaust prompt engineering first - clearer instructions, examples, chaining...

CAF PT3 Q23 Answer B: Organized sections with clear headers or tags: role, rules, tools guidance...

CAF PT3 Q24 Answer A: Define every output field with its type and allowed values

CAF PT3 Q25 Answer A: Only include dates and figures that appear verbatim in the source; if a date..

CAF PT3 Q26: Why might you instruct Claude to express uncertainty when it is not confident?

CAF PT3 Q27: A global product serves users in many languages. Which prompt-level instruction...

CAF PT3 Q28 Answer A: Exact-match or rule-based checks against expected outputs for structured tasks

CAF PT3 Q29: What is 'LLM-as-judge' evaluation?

CAF PT3 Q30 Answer B: A higher temperature (e.g., around 0.8-1.0) to increase diversity of ideas

CAF PT3 Q31 Answer B: Claude emits a tool_use call like get_weather(location: Mumbai), your app...

CAF PT3 Q32 Answer B: Claude calls tool A, receives its result, then in a subsequent turn calls...

CAF PT3 Q33: What characterizes the ReAct-style pattern many Claude agents follow?

CAF PT3 Q34 Answer B: Build 8 MCP servers once and every MCP-compatible app can use them - turning..

CAF PT3 Q35 Answer B: Trustworthiness of the server: it will receive tool calls and supply content..

CAF PT3 Q36 Answer A: Goal achieved - a defined success check ends the loop

CAF PT3 Q37 Answer B: Design the tool to paginate, summarize, or cap results (e.g., top-N rows...

CAF PT3 Q38 Answer A: Strengthen the tool description and system prompt to require tool use for...

CAF PT3 Q39: What is Claude Code?

CAF PT3 Q40 Answer A: Require human approval for destructive or irreversible operations

CAF PT3 Q41 Answer A: Such attempts will occur routinely, so defenses must not depend solely on...

CAF PT3 Q42: Where can input moderation fit in a Claude application pipeline?

CAF PT3 Q43 Answer A: Redact or tokenize PII in logs, restrict log access, and set retention...

CAF PT3 Q44 Answer A: SOC 2 attestation of the provider's security controls

CAF PT3 Q45 Answer A: Verification of critical facts against primary sources before the decision...

CAF PT3 Q46: What is automation bias, and why does it matter in AI-assisted workflows?

CAF PT3 Q47 Answer A: Consequential actions like forwarding email require explicit user...

CAF PT3 Q48 Answer B: Staged rollout: internal dogfood, then a small user percentage with...

CAF PT3 Q49 Answer A: Detection: monitoring and alerting on quality and safety signals

CAF PT3 Q50 Answer B: Decline to reproduce copyrighted text at length, offering a summary or...

CAF PT3 Q51 Answer A: RAG over the help-center corpus for grounded answers, confidence-based...

CAF PT3 Q52 Answer A: Storing complete responses for repeated identical (or semantically...

CAF PT3 Q53 Answer A: Stream responses and render tokens immediately

CAF PT3 Q54: What is a 'golden dataset' in LLM application evaluation?

CAF PT3 Q55: How can a team compare two prompt variants live in production responsibly?

CAF PT3 Q56 Answer A: Spend alerts and budget thresholds on API usage

CAF PT3 Q57 Answer A: Retrieval must filter by tenant at the query level (or use per-tenant...

CAF PT3 Q58 Answer A: A queue between ingestion and processing, with workers consuming at a rate...

CAF PT3 Q59 Answer A: RAG serves updated content as soon as the index refreshes, while fine-tuning..

CAF PT3 Q60 Answer A: So quality issues can be attributed to specific prompt or model changes...

CAF PT4 Q01 Answer A: Input tokens

CAF PT4 Q02: When using extended thinking, what does the thinking budget control?

CAF PT4 Q03: Which image formats does the Claude API accept?

CAF PT4 Q04 Answer B: The request fails with a validation error, so the application must manage...

CAF PT4 Q05: An engineering team uses Claude to generate code for a payment service. Which...

CAF PT4 Q06 Answer A: Low confidence indications, failed validation of the small model's output...

CAF PT4 Q07 Answer A: Tokenizers are typically optimized around English-heavy corpora, so many...

CAF PT4 Q08 Answer B: Running your own task-specific evaluation suite against candidate models and..

CAF PT4 Q09 Answer A: A longer input prompt

CAF PT4 Q10: What does the date-like suffix in a Claude model string represent?