Technology9 min read

CCDV-F Practice Questions for the Claude Messages API

Trace three Claude Messages API scenarios for the Developer Foundations exam. Diagnose a broken stream, return client tool results, and choose when to cache or batch.

OpenExamPrep Editorial TeamNovember 16, 2026

✓Key Facts

  • •Anthropic’s July 2026 Claude Certified Developer – Foundations exam guide lists 53 items for the CCDV-F exam.
  • •The Claude Certified Developer – Foundations v1.0 guide gives candidates 120 minutes to complete the CCDV-F exam.
  • •The CCDV-F v1.0 exam guide sets a passing score of 720 on a scaled range of 100–1,000.
  • •Applications and Integration accounts for approximately 33.1% of scored CCDV-F items in Anthropic’s July 2026 blueprint.
  • •Claude API Mechanics accounts for approximately 6.8% of scored CCDV-F items in Anthropic’s July 2026 blueprint.
  • •Pearson VUE lists December 7, 2026 as the start date for updated Claude Certified Developer – Foundations exam content.
  • •Anthropic Partner Academy lists a $125 USD exam fee for Claude Certified Developer – Foundations before partner discounts.
  • •Anthropic requires a recognized Claude Partner Network organization email address to register for the CCDV-F exam.
  • •Anthropic Partner Academy lists English and Japanese as the two available examination languages for Claude Certified Developer – Foundations.

Start with a trace. In these three original practice questions for Claude Certified Developer – Foundations (CCDV-F), you’ll diagnose a broken Claude response stream, return the results of two client tool calls, and choose how to process live and overnight requests. Before reading each answer, write down the last step that completed, which component owns the next step, and the field or event that proves it. You can replay the examples with local fixtures and a stub tool.

Check which CCDV-F guide applies to your exam date

Anthropic’s July 2026 v1.0 guide lists 53 questions in 120 minutes and a passing score of 720 on a 100–1,000 scale. It includes multiple-choice and multiple-response items, with each item telling you how many answers to select. Applications and Integration accounts for about 33.1% of scored items. Claude API Mechanics, one skill within that domain, accounts for about 6.8% of scored items. The cases here cover a few decisions within that larger scope. Sections 6 and 8 of the guide give you the full task list and Anthropic’s sample questions. Anthropic’s CCDV-F Exam Guide, v1.0

Pearson VUE says updated CCDV-F exam content starts December 7, 2026, with the updated guide scheduled for November 17, 2026. Use v1.0 for an appointment through December 6. If you test on December 7 or later, open the updated guide and check its task statements, weights, and format before setting your practice mix or pacing. The API documentation explains the behavior in these exercises. Your appointment’s guide determines which skills to study for the exam. Pearson VUE’s exam update · Anthropic Partner Academy certification guides

If you haven’t booked, Anthropic lists a $125 USD exam fee before any partner discount. Exam access is through a Claude Partner Network organization and requires a recognized partner work email. The exam is offered in English and Japanese. Anthropic Partner Academy certification FAQ

Question 1. Did Claude finish the turn or did the stream fail?

A repair help desk streams a reply about an equipment fault. The browser shows “Check whether the”. The server has already recorded the response as a successful final answer because it received HTTP 200. Here is an abbreviated server-sent event trace.

HTTP 200
message_start
content_block_start  index=0  type=text
content_block_delta  index=0  text="Check whether the"
error               type=overloaded_error
connection closes

What should the handler change first? Raise the output token cap, treat the visible text as a completed answer, or mark the generation as interrupted and apply the stream failure policy?

Answer. Mark it interrupted. The application received a stream, partial text and then an error. Treat the displayed text as partial and avoid saving it as a completed turn. For this read-only response, the app can offer a retry from the last completed conversation state, subject to its retry policy. Replace or clearly distinguish the partial display if you retry so the user can tell which answer they’re reading. Anthropic documents that an error can arrive inside a stream that began with HTTP 200. Streaming events · Claude API errors

Now replay the same visible text with a different ending. This time you receive content_block_stop, then a message_delta carrying stop_reason: "max_tokens", then message_stop. The stream finished, but the output hit the request’s generation cap. Handle the truncated result or revise the budget for a new request. If the cap cut off a tool call, Anthropic’s guidance is to retry the request with a larger limit to get the full call. Give the two fixtures different outcomes in your test. They look identical in the browser. The events tell you which recovery applies. Anthropic’s guide to stop reasons

Try a third fixture. Stream a tool’s arguments in several input_json_delta fragments, then drop the connection before the call is complete. Assert that your stub tool ran zero times. When you replay a successful turn, assemble fragments by content block index, parse and validate the complete arguments, and dispatch only after the completed turn calls for client tool execution. Use the OpenExamPrep streaming lesson to review the event order. Anthropic’s streaming tool-use example

Question 2. How do you preserve one successful lookup and one failure?

In a fresh request, the equipment assistant returns stop_reason: "tool_use" with two client calls in the same assistant turn. The service looks up whether a replacement fan is available and whether the equipment is under warranty. The stock lookup with ID toolu_stock succeeds. The warranty call with ID toolu_warranty fails because the service is unavailable and the handler has reached its retry limit. A developer proposes sending the text “Fan in stock, warranty unknown” as an ordinary user message.

What should the app send in the next Messages API request? First identify which part of the call history Claude needs in order to continue and how the response should distinguish a failed lookup from a valid answer.

Answer. Keep the assistant’s tool-use turn and add one user message holding both tool results. Include the conversation history for the standard stateless Messages API call, including the assistant turn as returned. Follow that assistant turn immediately with the user tool results. Each block’s tool_use_id answers the matching call. The outage gets is_error: true and a short explanation. The next turn in the request looks like this. Messages API conversation history · Anthropic’s tool-result format

{
  "role": "user",
  "content": [
    {
      "type": "tool_result",
      "tool_use_id": "toolu_stock",
      "content": "Replacement fan is in stock."
    },
    {
      "type": "tool_result",
      "tool_use_id": "toolu_warranty",
      "is_error": true,
      "content": "Warranty service unavailable. Coverage could not be determined."
    }
  ]
}

That gives Claude enough information to report what the service actually found. A warranty service that successfully reports an expired policy would return a normal result. Marking the outage as a successful “no coverage” response would turn missing evidence into a business decision. Dropping the failed result leaves a tool call without its matching result. A text summary alone loses the link between each call ID and its outcome. Anthropic’s guidance on tool execution errors

Try a retry after both tools ran. Make the follow-up API request fail at the transport layer. Keep the completed tool outcomes in the application’s state and test a retry that resends the same assistant turn and its results. Count tool invocations to verify that the retry of the model request didn’t rerun the handlers. This test is especially relevant when a tool has a side effect, such as opening a service ticket. In that case, test duplicate protection at the system that performs the write as well. Review how the application assembles the request in the OpenExamPrep Messages API fundamentals lesson.

Question 3. Which work should be streamed, cached or batched?

The help desk now has two workloads. Technicians need to see answers as they are generated during a repair. Many requests use the same long equipment manual followed by a different question. A separate job classifies thousands of independent historical tickets, and its downstream report can wait for the job to finish. The team proposes sending both workloads to the Message Batches API to simplify the integration.

Which constraint should decide the API path for each workload? For the live route, also name the measurement that would tell you the shared manual was reused.

Answer. Stream the live Messages API response, and consider prompt caching for the stable manual. Streaming lets the technician see output during the request. For an explicit cache breakpoint, put it at the end of the shared prefix, before the varying question. Keep the prefix identical across calls and use a cacheable length for the chosen model. Within the cache lifetime, a positive usage.cache_read_input_tokens on the next call shows input was read from cache. In a controlled test with one breakpoint at the end of the manual, compare that count with the original cache_creation_input_tokens to verify reuse of the full prefix. Wall-clock latency alone can vary for other reasons. Anthropic’s streaming guide · Anthropic’s prompt caching guide

Use Message Batches for the historical job if its deadline can tolerate asynchronous completion. Give each request a unique custom_id and associate results with those IDs. The results can arrive out of input order. Check the outcome for each request even when the batch itself has finished, so one failed classification can go to the appropriate recovery path. Batching addresses the delivery schedule of independent requests. Caching addresses shared input. If the offline job also repeats eligible context, the batch API supports caching on a best-effort basis. Anthropic’s batch processing guide

Try two fixture changes. Move a different timestamp into the shared prefix on each request and explain why those prefixes no longer match. Then reverse two batch result lines and verify your code still attaches each result to the right ticket. Replace one success with an error and check that only that ticket is marked for follow-up. Those tests expose assumptions that a single successful run can hide.

Turn each missed scenario into one test

For each case, write the fix you considered, the evidence for or against it, and a follow-up test that changes one condition. For example, increasing the output limit only fits the first case’s completed max_tokens branch. Your test should hold the displayed text constant and change the events that ended the turn. If you were drawn to the wrong option, record the evidence you missed before you read the explanation.

Evidence you overlookedWhat to inspect on the next questionWhat to practice
Partial output followed by an errorWhich stream events actually arrived?Compare interruption with a normal end at the output cap.
Two tool IDs with different outcomesDoes the follow-up carry both results and preserve their meaning?Return one success and one explicit error in the same user turn.
Interactive deadline and repeated inputWhat must arrive during the request, and which part of the request stays stable?Pick a delivery path first, then verify cache reuse separately.

Then open the free CCDV-F practice questions. Pick 12 unfamiliar questions across more than one topic. For the July v1.0 format, give yourself about 27 minutes to match the average pace of 53 questions in 120 minutes. Include your reading time and flag any answers you guessed. Review the flagged items and the misses together. Use the practice percentage to decide what to revisit. The certification uses a scaled score, so the raw percentage from a practice set cannot tell you whether you met the 720 standard. Anthropic’s score explanation in section 9

Before the next set, change the condition that made you hesitate. If you missed the stream question, answer a case where the stream finishes but the output is unusable. If you missed the workload question, decide what you would change if the overnight report suddenly required a result during the request. Once you can explain how the new evidence changes your action, move on to a different skill.

For candidates still choosing which Claude exam to take, see the Claude certification comparison.

Test Your Knowledge
Question 1 of 3

A streamed response closes normally. Near the end, the message delta reports max_tokens and the final tool block has incomplete arguments. What should the application do?

A

Run the tool with any fields that were received.

B

Send a tool result claiming success for the incomplete call.

C

Retry the model request with a larger output limit to get a complete tool call.

D

Record the result as an overloaded stream based on the incomplete arguments.

CCDV-FClaude Certified DeveloperMessages APIClaude APIClaude tool use

Related Articles

Stay Updated

Get free exam tips and study guides delivered to your inbox.

Free exam tips & study guides. Unsubscribe anytime.