CCDV-F Claude API Mechanics Practice Test 3 — 15 Questions

Question 1

A developer is building a document-summarization service on the Anthropic Messages API. The service must always return a summary followed by exactly one structured JSON block containing metadata (e.g., word count, sentiment). To guarantee the model never emits any text after the JSON block, the developer adds the closing brace of the JSON structure to the 'stop_sequences' array. What is the effect of this configuration on the API response?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

The response's 'stop_reason' is 'stop_sequence' and the text field contains the JSON block with the stop string trimmed off.. A stop sequence halts generation as soon as the matching string is produced, and the matched string itself is not included in the returned text. The response's stop_reason becomes 'stop_sequence', letting the application know a configured terminator fired rather than the model finishing naturally. This is why developers who place a closing JSON brace in stop_sequences see the brace omitted from the output.

Answer

The API returns an error because stop_sequences cannot contain characters that might appear inside the model's generated JSON.

Answer

The response's 'stop_reason' is 'max_tokens' because stop sequences only apply to the final assistant turn, not to content inside a single message.

Answer

The API silently strips the entire JSON block from the response, returning only the summary text to the caller.

Question 2

A developer is streaming a response from the Anthropic Messages API using server-sent events and wants to assemble the final text on the client. They observe that each event delivers a small fragment of the answer. Which event type carries the incremental text delta that must be concatenated to reconstruct the full completion?

Accepted Answer

content_block_delta. Streaming responses arrive as a sequence of typed server-sent events, and the actual generated text is delivered piecewise inside content_block_delta events. A client must concatenate those deltas in order and treat message_start, content_block_stop, message_delta, and message_stop as structural markers rather than content.

Answer

content_block_stop

Answer

message_start

Answer

ping

Question 3

A developer is building a Python application that streams responses from the Anthropic Messages API. They notice that when they set stream=True, the response object is an iterator of server-sent events. They want to extract only the incremental text chunks as they arrive, without waiting for the full message. Which event type should they filter for to get the text deltas?

Accepted Answer

content_block_delta. When streaming with the Messages API, the server sends a series of events. The content_block_delta event is specifically designed to carry incremental text deltas as Claude generates the response. By listening for this event type, the developer can process text as it arrives, enabling real-time user experiences. Other events like message_start, message_delta, and ping serve different purposes such as initialization, final metadata, or keep-alive.

Answer

message_delta

Answer

message_start

Answer

ping

Question 4

A developer sends a Messages API request that includes a system prompt and several prior turns, and Claude replies with stop_reason set to "max_tokens" even though the conversation is short. They want Claude to finish its thought rather than truncate. Which change most directly addresses the truncation?

Accepted Answer

Increase the max_tokens value in the request so Claude has more room to complete its response.. The stop_reason value max_tokens is a direct signal that the output token limit was reached before the model finished. The remedy is to raise max_tokens within the model's context constraints; stop sequences, temperature, and prompt placement do not affect how many output tokens are permitted.

Answer

Move the instructions from the system prompt into the first user turn to free output tokens.

Answer

Lower the temperature to 0 so the response becomes deterministic and naturally shorter.

Answer

Add a stop_sequence that matches the end of the expected answer so Claude knows when to stop.

Question 5

A developer is using the Anthropic Messages API and wants to ensure that Claude's response is deterministic and reproducible for a given prompt. They set the temperature parameter to 0. However, they observe that repeated calls with the same input sometimes yield slightly different outputs. Which factor is the most likely cause of this non-determinism?

Accepted Answer

The model's internal computations involve non-deterministic operations such as floating-point arithmetic or parallel processing.. Even with temperature set to 0, which makes the model choose the most probable next token, the underlying probability distribution can vary slightly between calls due to non-deterministic operations in the model's execution on distributed hardware. This includes floating-point arithmetic order and parallel processing. The API does not provide a seed parameter to enforce determinism. Therefore, the most likely cause is the model's internal non-determinism, not misconfiguration of temperature or top_p.

Answer

The API automatically injects a random seed into each request unless one is explicitly provided.

Answer

The top_p parameter is set to a value less than 1, introducing randomness even with temperature 0.

Answer

The temperature parameter is ignored when set to 0, and the default temperature is used instead.

Question 6

A developer is using the Anthropic Messages API and wants to limit the maximum number of tokens that Claude can generate in its response. Which parameter should they set in the request body?

Accepted Answer

max_tokens. The max_tokens parameter is specifically designed to cap the number of tokens generated in the response. It is a required field in the Messages API request. Other parameters like stop_sequences control stopping based on content, not token count. Parameters such as max_length or token_limit are not part of the API. Therefore, setting max_tokens is the correct approach to limit response length.

Answer

token_limit

Answer

max_length

Answer

stop_sequences

Question 7

A developer is building a document Q&A service on the Anthropic Messages API. Their integration currently sends the entire 240,000-token knowledge base on every turn of a long conversation, and they are hitting the model's context window limit. They want to keep the full conversation history and the knowledge base available without exceeding the window. Which approach best addresses the problem?

Accepted Answer

Summarize or truncate older conversation turns and retrieve only the most relevant knowledge-base chunks to include in each request.. The context window is a fixed ceiling on combined input and output tokens per request. To keep a long conversation and a large corpus usable, the developer must shrink what is actually sent: condense prior turns and inject only the retrieved passages relevant to the current question. Caching and streaming change cost or delivery mechanics, not the window itself, and max_tokens governs output only.

Answer

Enable prompt caching on the static knowledge base so repeated prefixes are served from cache and no longer count toward the context window.

Answer

Increase the max_tokens parameter so Claude can internally compress the knowledge base before answering.

Answer

Switch the request to stream: true so the server transmits the knowledge base in smaller chunks that bypass the context limit.

Question 8

A developer is using the Anthropic Messages API and wants to implement a retry mechanism for handling rate limit errors. They receive an HTTP 429 response with a 'retry-after' header. Which approach is the most appropriate for handling this error?

Accepted Answer

Wait for the number of seconds specified in the 'retry-after' header before retrying the request.. When the API returns a 429 status with a 'retry-after' header, the best practice is to pause and wait for the specified number of seconds before retrying. This respects the server's rate limiting policy and avoids making unnecessary requests that could lead to further throttling. Immediate retries or switching models do not address the root cause. Failing without retry is also not ideal for transient rate limits.

Answer

Immediately retry the request in a tight loop until it succeeds.

Answer

Switch to a different model that has higher rate limits and retry immediately.

Answer

Log the error and fail the request without retrying, as rate limits are permanent.

Question 9

A developer's agent calls the Messages API with several custom tools defined. During testing, the response repeatedly returns stop_reason "tool_use" with a tool_use block, but the application crashes because it expects a text block. The developer wants to handle this correctly so the agent can continue. What should the application do when it receives stop_reason "tool_use"?

Accepted Answer

Parse the tool_use block, execute the tool, then send a new request containing the assistant's tool_use turn and a user turn with the corresponding tool_result block.. A tool_use stop reason means the model has deliberately halted to request external execution. The application must run the named tool with the provided input, then continue the conversation by appending the assistant's tool_use turn followed by a user turn carrying the matching tool_result block. Only after receiving that result will the model resume and produce its final answer.

Answer

Set tool_choice to "none" and resend the request so Claude answers directly without invoking tools.

Answer

Resend the original request with the same parameters, because a tool_use stop reason indicates a transient server condition.

Answer

Ignore the tool_use block and read the accompanying text field, since Claude always includes a natural-language summary alongside tool calls.

Question 10

A developer is streaming a long Claude response using the Messages API with stream: true. Their client code reads Server-Sent Events and appends text to the UI. Mid-stream, the connection drops and the client reconnects by re-issuing the same request from scratch. Users complain that the answer restarts from the beginning. What is the most accurate explanation of what is happening?

Accepted Answer

Streaming responses are stateless; a dropped connection cannot be resumed, so re-issuing the request starts a brand new generation and the partial text already shown is discarded.. Streaming responses from the Messages API are not resumable: each HTTP request produces an independent generation delivered as Server-Sent Events. When the transport fails, the client cannot ask the server to continue from the last token it saw. The correct mitigation is to persist partial output locally, avoid blindly restarting the UI, and design retries that tolerate duplicate or truncated content.

Answer

The client must call the count_tokens endpoint before reconnecting so the server can splice the new stream onto the previous tokens.

Answer

The Messages API supports resuming a stream from the last received event, so the client should have sent the last event ID back on reconnect.

Answer

Setting a higher max_tokens on the reconnect lets the API detect the earlier partial output and continue from where it stopped.

Question 11

A developer is preparing a Messages API request and wants to give Claude a persistent persona: "You are a concise legal assistant. Never give advice outside contract review." Where should this instruction be placed so it applies to the whole conversation?

Accepted Answer

In the system parameter, as the top-level system prompt for the request.. The system parameter is purpose-built for top-level instructions that define Claude's role, tone, and boundaries for the whole request. Placing the legal-assistant persona there gives it consistent influence over all turns and separates standing guidance from conversational content. Other fields either carry no model-visible text or serve unrelated control functions.

Answer

As the first user message in the messages array, prefixed with "System:".

Answer

Inside a metadata object attached to the request body.

Answer

In the stop_sequences parameter so Claude stops when it deviates from the persona.

Question 12

A team is building a retrieval-augmented assistant that sends a large document set inside the system prompt on every turn. To reduce cost, they enable prompt caching. They notice cache_read_input_tokens is high on most turns but intermittently drops to zero even though the system prompt text has not changed. Which explanation best fits this behavior?

Accepted Answer

The cache has a short time-to-live, so if enough time passes between requests the cached prefix expires and the next request pays full input cost before re-populating the cache.. Prompt caching entries expire after a time-to-live if they are not referenced. Bursty or idle traffic patterns cause the prefix to be evicted between requests, so the next call misses the cache and reports zero cache reads before re-creating the entry. Keeping steady traffic, or reducing gaps between calls, stabilizes cache_read_input_tokens and preserves the intended cost savings.

Answer

The API randomly invalidates cache entries per account to distribute load, so intermittent cache misses are expected and cannot be controlled.

Answer

Prompt caching only applies to user messages, so any content placed in the system prompt bypasses the cache and is billed at full input rate.

Answer

Cache reads only register when max_tokens is large enough to cover the entire cached prefix, so smaller responses skip the cache.

Question 13

A developer wants Claude to return a fixed set of user-profile fields extracted from free-text bios. They want the response to be valid JSON they can parse without regex. The model occasionally adds a friendly sentence before the JSON. Which Messages API feature most reliably constrains the output?

Accepted Answer

Define a tool whose input_schema describes the profile fields, then set tool_choice to force Claude to call it.. Structured extraction is best enforced by defining a tool whose input_schema specifies the exact fields and types, then using tool_choice to require that the model call it. Claude then returns a tool_use block whose input is JSON matching the schema, eliminating stray prose. Prompt wording, temperature, stop sequences, and token limits influence style or length but cannot enforce a schema.

Answer

Increase max_tokens so Claude has more room to format the JSON correctly.

Answer

Set stop_sequences to "}" so generation halts exactly at the end of the JSON object.

Answer

Add the word "JSON" to the prompt and set temperature to 0.

Question 14

A developer is using the Messages API and wants Claude to reply in strict JSON matching a schema their downstream service expects. They have already written a clear instruction in the system prompt describing the schema. Which additional API feature most directly enforces that the output conforms to the schema?

Accepted Answer

Defining a tool whose input_schema describes the desired fields, then instructing Claude to call that tool for its response.. Tool definitions carry a JSON Schema in input_schema, and when Claude invokes a tool, the API returns arguments that conform to that schema. Routing the model's answer through a tool call therefore gives the strongest structural guarantee in the Messages API. Prompt instructions help guide behavior but do not enforce shape, so pairing them with a tool schema is the reliable pattern for strict JSON.

Answer

Setting temperature to 0 to eliminate all variation in the response format.

Answer

Increasing max_tokens so the model has room to complete the full JSON object without truncation.

Answer

Adding a stop_sequence equal to the closing brace so the API halts generation exactly at the end of the JSON.

Question 15

A developer is diagnosing a production Messages API workload where some requests fail with an overloaded_error and others return stop_reason "max_tokens". They want to handle both conditions correctly. Which TWO actions are appropriate? (Choose two.)

Accepted Answer

Implement exponential backoff with jitter and retry requests that return overloaded_error.. The two conditions require different handling. Transient overloaded_error should be retried with exponential backoff and jitter because capacity pressure is temporary. A max_tokens stop reason indicates the output was truncated by the configured ceiling, so the developer must raise max_tokens or reduce input to leave room for a complete answer. Retrying truncation unchanged, caching around an unprocessed request, or shrinking output for server load all fail to address the actual cause.

Answer

Convert overloaded_error responses into successful responses by caching the last valid completion for the same prompt.

Answer

Retry every request that returns stop_reason "max_tokens" with the identical parameters until it completes.

Answer

Lower max_tokens to 1 whenever overloaded_error occurs so the server has less work to do.