# Free CCDV-F practice questions

> 10 free CCDV-F practice questions with the correct answer and an explanation.

Page: https://devopscodelab.com/exams/claude-developer-foundations/free-practice-questions/
Practice exams: https://devopscodelab.com/exams/claude-developer-foundations/

### Question 1 of 10 (choose 2)

A developer is reviewing the stop_reason values a Messages API integration must handle. Which TWO values indicate that the application should take an action before Claude's turn can be considered complete? (Choose two.)

- A. pause_turn
- B. stop_sequence
- C. content_filter
- D. end_turn
- E. tool_use

**Answer:** A, E

**Explanation:** tool_use means Claude is waiting for the client to run tools and return results, and pause_turn means a long server-tool turn was paused and should be sent back so Claude can continue. end_turn is a natural finish and stop_sequence means a custom stop string was hit, so neither requires continuation. content_filter is not a Claude stop_reason; safety declines appear as refusal.

### Question 2 of 10

A team stores conversation history between turns as plain strings extracted from each response's text blocks. After enabling tools and thinking, requests start failing or Claude loses context about tool calls. What is the fix?

- A. Store and resend the full assistant content arrays, including tool_use and thinking blocks
- B. Store only the final text but add a summary of the tool calls to the next system prompt
- C. Store only the user turns and let Claude regenerate its earlier answers on each new request
- D. Store the full responses but strip the thinking blocks before each new request to save tokens

**Answer:** A

**Explanation:** When tools or thinking are used, the assistant content array must be appended as returned, because tool_result blocks must match prior tool_use blocks and thinking blocks must be passed back unchanged during tool use. Text-only history breaks those links. Removing thinking blocks during a tool loop can cause errors, and dropping assistant turns loses the conversation.

### Question 3 of 10

A nightly job must classify 400,000 support tickets. Results are needed by the next morning, and cost matters more than latency. Which design is MOST cost-effective?

- A. Stream every ticket individually through the Messages API so tokens arrive more quickly
- B. Submit the tickets through the Message Batches API and collect the results by custom_id
- C. Send the tickets concurrently through the Messages API using many parallel worker threads
- D. Put all tickets in one Messages API request and ask Claude to return a list of their labels

**Answer:** B

**Explanation:** The Message Batches API processes asynchronous requests at roughly half the standard price, usually within 24 hours, which fits an overnight deadline. Parallel synchronous calls pay full price and press against rate limits. Streaming lowers perceived latency but not cost, and a single giant prompt would exceed output limits and make per-ticket errors hard to isolate.

### Question 4 of 10

A summarization route on Claude Opus 5.5 returns good results, but latency and token spend are higher than needed for this simple task. Which change trades quality for speed MOST directly without switching models?

- A. Add a stop sequence after the first paragraph of output
- B. Set output_config.effort to low for this route
- C. Raise max_tokens so that the model can finish sooner
- D. Set thinking to disabled on every request of this route

**Answer:** B

**Explanation:** The effort setting controls how much the model thinks and how many tokens it spends overall, so low effort is the intended lever for routine work. On Claude Opus 5.5 thinking cannot be switched off, and a disabled thinking setting returns an error. A higher max_tokens only raises the ceiling, and a stop sequence would truncate the summary rather than make the model more efficient.

### Question 5 of 10

Before sending a large prompt, an application must check whether it will fit the budget for a route and estimate its input cost. Which approach gives an accurate number for the target model?

- A. Call the token counting endpoint with the same model, system prompt, tools and messages
- B. Use an open-source tokenizer from another model family and add a ten percent margin
- C. Send the real request with max_tokens set to 1 and then read input usage from the response
- D. Divide the character count by four, which matches Claude's tokenizer for all text

**Answer:** A

**Explanation:** The count_tokens endpoint returns the input token count for the exact model and payload, including tools and images, at no charge. A characters-divided-by-four rule is only a rough estimate, and other tokenizers do not match Claude's, especially since tokenizers differ between model generations. A real request with max_tokens 1 still bills the full input.

### Question 6 of 10

A customer-service system receives refund requests, technical questions and sales inquiries, and each type needs a very different prompt and tool set. Which workflow pattern is MOST appropriate?

- A. Evaluator-optimizer: have one model draft and another critique each reply
- B. Voting: run the same prompt several times and keep the majority answer
- C. Prompt chaining: send each request through all three prompts in sequence
- D. Routing: classify each request first, then pass it to a specialised prompt

**Answer:** D

**Explanation:** Routing separates distinct categories so each downstream prompt and tool set can be optimised without hurting the others. Running every request through all three prompts wastes calls and mixes concerns. Evaluator-optimizer improves quality of one output through iteration but does not separate categories, and voting increases confidence on one task rather than directing traffic.

### Question 7 of 10 (choose 2)

A developer implements the memory tool so an agent can keep notes across conversations. Which TWO statements about this tool are correct? (Choose two.)

- A. The memory tool only works inside the Message Batches API for asynchronous jobs
- B. The memory tool replaces the need to send any conversation history in requests
- C. Claude can create, read and update files in a memory directory between sessions
- D. The application implements the storage backend that executes the memory commands
- E. Anthropic stores the memory files permanently in the organisation's workspace

**Answer:** C, D

**Explanation:** The memory tool is client-side: Claude issues file-style commands against a memory directory, and the application decides where and how they are stored, allowing knowledge to persist across sessions. Anthropic does not host the memory store for this tool. Conversation history still has to be sent for the current session, and the tool is used in normal Messages API requests.

### Question 8 of 10

A financial analysis prompt produces conclusions that skip important reasoning steps. The team wants Claude to reason carefully on hard cases without paying for long reasoning on easy ones. What is the recommended configuration on current models?

- A. Use adaptive thinking, which lets the model decide how much to think per request
- B. Ask for the reasoning inside the final answer text on every request instead
- C. Disable thinking and raise max_tokens so the answer can contain more steps
- D. Send a fixed thinking budget of 32,000 tokens on every request to the model

**Answer:** A

**Explanation:** Adaptive thinking lets the model scale its reasoning to the difficulty of each request, and effort tunes the overall depth. Fixed thinking budgets are deprecated or rejected on current models. Forcing visible reasoning into every answer inflates output on easy cases, and disabling thinking removes the internal reasoning the team wants.

### Question 9 of 10

An MCP server exposes a tool that permanently deletes records. Which MCP tool annotation helps clients warn users before the call?

- A. openWorldHint set to true
- B. readOnlyHint set to true
- C. idempotentHint set to true
- D. destructiveHint set to true

**Answer:** D

**Explanation:** destructiveHint signals that a tool may perform destructive updates, so hosts can ask for confirmation. readOnlyHint true would wrongly claim the tool has no side effects. openWorldHint describes whether a tool reaches external entities, and idempotentHint says repeated calls have no extra effect; neither signals deletion risk.

### Question 10 of 10

A retrieval-augmented assistant pulls web pages into context. Which practice BEST reduces the risk that instructions hidden in those pages change the assistant's behaviour?

- A. Raise the effort level so the assistant reasons longer before following any instruction
- B. Retrieve pages only from search results that rank on the very first page of the search engine
- C. Mark retrieved content as untrusted data in tags and restrict which tools it can trigger
- D. Put the retrieved pages before the system prompt so they are read with lower priority

**Answer:** C

**Explanation:** Separating untrusted content clearly and limiting the actions available while processing it reduces both the chance and the impact of indirect prompt injection. Search ranking says nothing about safety. More reasoning does not reliably stop injected instructions, and retrieved content cannot be placed ahead of the system prompt, which is a separate field.

Ready for a timed run? Take the free 15-question CCDV-F mini exam: https://devopscodelab.com/app/exams/claude-developer-foundations/mini/

DevOpsCode Lab is not affiliated with Anthropic. CCDV-F and related names are trademarks of their owners. We write every question ourselves; these are not real exam questions.
