Claude Architect Foundations Practice Test

Answer choices

1. A. In a message with role system inside the messages array
2. B. In the anthropic-version header
3. C. In the stop_sequence field
4. D. In a top-level system parameter

Wrong-answer review

• A. In a message with role system inside the messages array: This distractor describes the idea that In a message with role system inside the messages array. In "An architect wants a durable instruction to apply across a Messages API interaction, but does not want to fake that instruction as a user turn. According to Anthropic's Messages docs, where should that instruction go?", that misses the required action because the correct response is "In a top-level system parameter". On the job, mixing up that distractor with "In a top-level system parameter" can lead to the wrong inferred coverage: claude api architecture and messages api action or troubleshooting path.
• B. In the anthropic-version header: This distractor describes the idea that In the anthropic-version header. In "An architect wants a durable instruction to apply across a Messages API interaction, but does not want to fake that instruction as a user turn. According to Anthropic's Messages docs, where should that instruction go?", that misses the required action because the correct response is "In a top-level system parameter". On the job, mixing up that distractor with "In a top-level system parameter" can lead to the wrong inferred coverage: claude api architecture and messages api action or troubleshooting path.
• C. In the stop_sequence field: This distractor describes the idea that In the stop_sequence field. In "An architect wants a durable instruction to apply across a Messages API interaction, but does not want to fake that instruction as a user turn. According to Anthropic's Messages docs, where should that instruction go?", that misses the required action because the correct response is "In a top-level system parameter". On the job, mixing up that distractor with "In a top-level system parameter" can lead to the wrong inferred coverage: claude api architecture and messages api action or troubleshooting path.

Answer choices

1. A. Always choose the fastest model because latency is the only production metric that matters
2. B. Keep the cheapest model and compensate only with longer prompts
3. C. Avoid testing and assume all Claude models can reach the same quality level
4. D. Favor capability even if latency or cost are higher when the workload genuinely requires it

Wrong-answer review

• A. Always choose the fastest model because latency is the only production metric that matters: This distractor describes the idea that Always choose the fastest model because latency is the only production metric that matters. In "A workflow truly depends on the strongest reasoning available, and lower-capability models miss critical steps even after testing. Which tradeoff should the architect accept according to Anthropic's model-selection guidance?", that misses the required action because the correct response is "Favor capability even if latency or cost are higher when the workload genuinely requires it". On the job, mixing up that distractor with "Favor capability even if latency or cost are higher when the workload genuinely requires it" can lead to the wrong inferred coverage: claude fundamentals and model selection action or troubleshooting path.
• B. Keep the cheapest model and compensate only with longer prompts: This distractor describes the idea that Keep the cheapest model and compensate only with longer prompts. In "A workflow truly depends on the strongest reasoning available, and lower-capability models miss critical steps even after testing. Which tradeoff should the architect accept according to Anthropic's model-selection guidance?", that misses the required action because the correct response is "Favor capability even if latency or cost are higher when the workload genuinely requires it". On the job, mixing up that distractor with "Favor capability even if latency or cost are higher when the workload genuinely requires it" can lead to the wrong inferred coverage: claude fundamentals and model selection action or troubleshooting path.
• C. Avoid testing and assume all Claude models can reach the same quality level: This distractor describes the idea that Avoid testing and assume all Claude models can reach the same quality level. In "A workflow truly depends on the strongest reasoning available, and lower-capability models miss critical steps even after testing. Which tradeoff should the architect accept according to Anthropic's model-selection guidance?", that misses the required action because the correct response is "Favor capability even if latency or cost are higher when the workload genuinely requires it". On the job, mixing up that distractor with "Favor capability even if latency or cost are higher when the workload genuinely requires it" can lead to the wrong inferred coverage: claude fundamentals and model selection action or troubleshooting path.

Answer choices

1. A. XML tags to separate the prompt sections
2. B. A longer unstructured paragraph
3. C. A different workspace for each section
4. D. A stop_sequence for each section

Wrong-answer review

• B. A longer unstructured paragraph: This distractor describes the idea that A longer unstructured paragraph. In "When a Claude prompt contains instructions, examples, and background material that must not blur together, what does Anthropic recommend using?", that misses the required action because the correct response is "XML tags to separate the prompt sections". On the job, mixing up that distractor with "XML tags to separate the prompt sections" can lead to the wrong inferred coverage: prompting, system prompts, multi-turn design, and context management action or troubleshooting path.
• C. A different workspace for each section: This distractor describes the idea that A different workspace for each section. In "When a Claude prompt contains instructions, examples, and background material that must not blur together, what does Anthropic recommend using?", that misses the required action because the correct response is "XML tags to separate the prompt sections". On the job, mixing up that distractor with "XML tags to separate the prompt sections" can lead to the wrong inferred coverage: prompting, system prompts, multi-turn design, and context management action or troubleshooting path.
• D. A stop_sequence for each section: A stop_sequence can be used to signal the end of a response but does not provide a method for structuring or separating different parts of the prompt itself.

Answer choices

1. A. Into the max_tokens field
2. B. Into the top-level system parameter
3. C. Into a role system message in the messages array
4. D. Into the request-id response header

Wrong-answer review

• A. Into the max_tokens field: This distractor describes the idea that Into the max_tokens field. In "During migration from Text Completions to the Messages API, where should the durable system instruction move?", that misses the required action because the correct response is "Into the top-level system parameter". On the job, mixing up that distractor with "Into the top-level system parameter" can lead to the wrong inferred coverage: claude api architecture and messages api action or troubleshooting path.
• C. Into a role system message in the messages array: This distractor describes the idea that Into a role system message in the messages array. In "During migration from Text Completions to the Messages API, where should the durable system instruction move?", that misses the required action because the correct response is "Into the top-level system parameter". On the job, mixing up that distractor with "Into the top-level system parameter" can lead to the wrong inferred coverage: claude api architecture and messages api action or troubleshooting path.
• D. Into the request-id response header: This distractor describes the idea that Into the request-id response header. In "During migration from Text Completions to the Messages API, where should the durable system instruction move?", that misses the required action because the correct response is "Into the top-level system parameter". On the job, mixing up that distractor with "Into the top-level system parameter" can lead to the wrong inferred coverage: claude api architecture and messages api action or troubleshooting path.

Answer choices

1. A. Managed Agents are only for Claude Code and cannot be used with Claude Platform APIs
2. B. The Messages API gives fine-grained control for custom agent loops, while Managed Agents provide a pre-built managed agent harness
3. C. The Messages API can only be used for single-turn prompts, while Managed Agents are the only way to do multi-turn work
4. D. Managed Agents remove the need for authentication, while the Messages API requires it

Wrong-answer review

• A. Managed Agents are only for Claude Code and cannot be used with Claude Platform APIs: Managed Agents are not limited to Claude Code; they can be used with other Claude Platform APIs. The source does not restrict Managed Agents to Claude Code alone.
• C. The Messages API can only be used for single-turn prompts, while Managed Agents are the only way to do multi-turn work: This distractor describes the idea that The Messages API can only be used for single-turn prompts, while Managed Agents are the only way to do multi-turn work. In "What is the key architectural difference between the Messages API and Claude Managed Agents in Anthropic's platform positioning?", that misses the required action because the correct response is "The Messages API gives fine-grained control for custom agent loops, while Managed Agents provide a pre-built managed agent harness". On the job, mixing up that distractor with "The Messages API gives fine-grained control for custom agent loops, while Managed Agents provide a pre-built managed agent harness" can lead to the wrong inferred coverage: claude fundamentals and model selection action or troubleshooting path.
• D. Managed Agents remove the need for authentication, while the Messages API requires it: This explanation is inaccurate as it confuses authentication requirements with the nature of the APIs. Both the Messages API and Managed Agents require authentication, but they serve different purposes and have different features.

Answer choices

1. A. Judge prompt quality only by whether the response is longer
2. B. Make controllable prompt changes and measure the impact rather than changing everything at once
3. C. Add random extra examples until the answer looks good
4. D. Switch models first and skip prompt testing entirely

Wrong-answer review

• A. Judge prompt quality only by whether the response is longer: This distractor describes the idea that Judge prompt quality only by whether the response is longer. In "A team changes several parts of a Claude prompt at once and gets better results, but cannot tell why. According to Anthropic's prompt engineering overview, what is the better improvement pattern?", that misses the required action because the correct response is "Make controllable prompt changes and measure the impact rather than changing everything at once". On the job, mixing up that distractor with "Make controllable prompt changes and measure the impact rather than changing everything at once" can lead to the wrong inferred coverage: prompting, system prompts, multi-turn design, and context management action or troubleshooting path.
• C. Add random extra examples until the answer looks good: Adding random extra examples does not necessarily improve the prompt's effectiveness; it may introduce irrelevant information.
• D. Switch models first and skip prompt testing entirely: This distractor describes the idea that Switch models first and skip prompt testing entirely. In "A team changes several parts of a Claude prompt at once and gets better results, but cannot tell why. According to Anthropic's prompt engineering overview, what is the better improvement pattern?", that misses the required action because the correct response is "Make controllable prompt changes and measure the impact rather than changing everything at once". On the job, mixing up that distractor with "Make controllable prompt changes and measure the impact rather than changing everything at once" can lead to the wrong inferred coverage: prompting, system prompts, multi-turn design, and context management action or troubleshooting path.

Answer choices

1. A. message_stop
2. B. workspace_end
3. C. message_start
4. D. content_block_delta

Wrong-answer review

• B. workspace_end: This distractor describes the idea that workspace_end. In "Which event marks the end of a Claude streaming response in Anthropic's documented event flow?", that misses the required action because the correct response is "message_stop". On the job, mixing up that distractor with "message_stop" can lead to the wrong inferred coverage: claude api architecture and messages api action or troubleshooting path.
• C. message_start: This distractor describes the idea that message_start. In "Which event marks the end of a Claude streaming response in Anthropic's documented event flow?", that misses the required action because the correct response is "message_stop". On the job, mixing up that distractor with "message_stop" can lead to the wrong inferred coverage: claude api architecture and messages api action or troubleshooting path.
• D. content_block_delta: This distractor describes the idea that content_block_delta. In "Which event marks the end of a Claude streaming response in Anthropic's documented event flow?", that misses the required action because the correct response is "message_stop". On the job, mixing up that distractor with "message_stop" can lead to the wrong inferred coverage: claude api architecture and messages api action or troubleshooting path.

Answer choices

1. A. Structured outputs
2. B. Extended thinking
3. C. Multimodal input
4. D. Tool use

Wrong-answer review

• A. Structured outputs: Structured outputs help format responses in a predictable way, which can be beneficial for clarity and organization, but they do not enhance the depth of reasoning required for difficult tasks.
• C. Multimodal input: Multimodal input allows Claude to process and respond based on various types of data, but it does not directly address the need for deeper reasoning in difficult tasks.
• D. Tool use: Tool use enables Claude to access external resources or perform specific actions, which can be useful for complex tasks, but it is not specifically designed for extended thinking.

Answer choices

1. A. Only in the last user message
2. B. Only in a stop_sequence field
3. C. In the system prompt
4. D. Repeated in every assistant response

Wrong-answer review

• A. Only in the last user message: This distractor describes the idea that Only in the last user message. In "A team wants Claude to act as a cautious finance reviewer across an entire interaction, not just for one turn. According to Anthropic's system-prompt guidance, where should that durable behavior instruction live?", that misses the required action because the correct response is "In the system prompt". On the job, mixing up that distractor with "In the system prompt" can lead to the wrong inferred coverage: prompting, system prompts, multi-turn design, and context management action or troubleshooting path.
• B. Only in a stop_sequence field: Placing the instruction in a stop_sequence field is incorrect because stop sequences are used to signal when Claude should stop generating text, not for providing instructions or behaviors.
• D. Repeated in every assistant response: This distractor describes the idea that Repeated in every assistant response. In "A team wants Claude to act as a cautious finance reviewer across an entire interaction, not just for one turn. According to Anthropic's system-prompt guidance, where should that durable behavior instruction live?", that misses the required action because the correct response is "In the system prompt". On the job, mixing up that distractor with "In the system prompt" can lead to the wrong inferred coverage: prompting, system prompts, multi-turn design, and context management action or troubleshooting path.

Answer choices

1. A. Treat the answer as a fully completed natural-language response and return it immediately
2. B. Discard the response because tool_use indicates an API authentication failure
3. C. Retry the same request without tools because tool_use means the model is confused
4. D. Execute or handle the requested tool interaction before continuing the loop

Wrong-answer review

• A. Treat the answer as a fully completed natural-language response and return it immediately: Incorrect. Treating the answer as a fully completed natural-language response when a tool_use stop reason is received would not allow for the execution of the requested tool interaction.
• B. Discard the response because tool_use indicates an API authentication failure: Incorrect. Discarding the response because tool_use indicates an API authentication failure is incorrect. A tool_use stop reason does not imply an API authentication issue; it simply means Claude is calling a tool.
• C. Retry the same request without tools because tool_use means the model is confused: Incorrect. Retrying the same request without tools when a tool_use stop reason is received would not address the fact that Claude is expecting to handle a tool interaction.