CCDV-F practice questions

Question 11· Multiple choice

Applications and Integration

A workflow drafts a refund note and posts it to a ticketing system. Sometimes the post times out after the ticket has already been created, the workflow retries, and the customer ends up with two refund notes on one ticket.

What removes the duplicate?

Send a key with the post that the ticketing system takes as the note's identity.
Stop retrying the post and let a person handle the timed out cases.
Wait longer for the post, so the timeout stops firing on slow days.
Read the ticket for an existing note before the post is attempted.

Question 12· Multiple choice

Agents and Workflows

A team adds a second model call that reviews the first one's output and asks for a revision. On translation, where a glossary fixes how certain terms must be rendered, the loop improves the result and settles after two or three passes. On a request to make a blog post livelier, the loop revises without end and nobody can say whether the fourth version beats the second.

What separates the two cases?

The translation task produces shorter output, so fewer passes are needed.
The translation task has a standard the reviewing call can apply.
The blog post task needs the stronger model on the reviewing call.
The blog post task should generate its revisions in parallel.

Question 13· Multiple choice

Applications and Integration

A pipeline runs in two steps. The first reads an inbound customer email and produces a short description of what the customer wants. The second call places that description into its own system prompt, as the instructions for what to do about it.

What risk does this design create?

The second call costs more, since a system prompt is billed twice.
Text that began in the email now arrives as instructions in step two.
The two calls cannot share a conversation, so context is lost.
The description may be truncated, so step two sees less than it should.

Question 14· Multiple choice

Model Selection and Optimization

A summarisation prompt states the output format three separate ways, warns against six specific mistakes and carries a worked example of each. All of it was added over months to hold an older, weaker model to the format. The route has since moved to a much stronger model, with the prompt carried across untouched. Output has been correct throughout.

What does the migration leave the team owing?

Nothing, since the output is correct and the prompt is what makes it correct.
A rewrite of the prompt from scratch against the stronger model's documentation.
A pass that removes the scaffolding piece by piece, scoring the format each time.
A rule that the prompt is frozen, so the format cannot drift after a migration.

Question 15· Multiple choice

Applications and Integration

A team finished moving its summarisation calls from an old internal wrapper service to a new client library six weeks ago. The new path carries most of the traffic. Two callers were never migrated, so the wrapper is still deployed, still holds a credential, and nobody has looked at its logs since the move.

Callers reaching the API through the new client library and through the wrapper service it replaced

What should close the migration out?

Migrate the two callers, then remove the wrapper and revoke its credential.
Add the wrapper to the dashboards so its traffic is monitored again.
Freeze the wrapper and leave it for callers that have not moved.
Point the wrapper at the new library so both paths share one client.

Question 16· Multiple choice

Agents and Workflows

A migration agent runs for hours over hundreds of turns. As the context fills, the harness drops the oldest turns to make room for new ones. Later in the run the agent redoes a table it converted two hours earlier and contradicts a naming decision it made near the start.

What should the harness do instead of dropping the oldest turns?

Replace them with a written summary of what they settled.
Drop turns from the middle and keep the start and the end.
Move the run to a model with a larger context window.
Start a fresh conversation with the original brief when the window fills.

Question 17· Multiple choice

Applications and Integration

A pricing team needs 9,000 documents scored every weekday and the scores loaded before the market opens at 09:30. Nobody reads a single score before then. A developer proposes the batch path, submitting at 04:00, on the grounds that no caller is waiting.

What is the risk in that plan, and what covers it?

The batch path bounds completion rather than promising a time, so the plan needs a cutoff and a fallback.
There is no risk, since the batch path is designed for work like this.
The documents may exceed the batch size limit, so they need splitting.
The batch path costs less, so a nightly run of this size needs approval.

Question 18· Multiple choice

Model Selection and Optimization

A service written in a language with no official SDK calls the Messages API over plain HTTP. The body and the authentication header are right, and every request is rejected with a message naming a header the request does not carry. The same call from the Python SDK succeeds.

Which header is missing, and why did the SDK not need it?

The API version header, which the SDKs send on every call.
The content type header, which the SDKs set from the body.
The user agent header, which the SDKs build from their version.
The accept header, which the SDKs set for the streaming case.

Question 19· Multiple choice

Applications and Integration

A compliance reviewer asks the team to be able to show, months later, exactly what the model said to a particular customer. The team's plan is to set the sampling temperature to zero and store the request, so the reply can be produced again whenever it is asked for.

What is wrong with that plan?

Zero is not an accepted value, so those requests would be refused.
A repeat can differ, so the reply itself is the thing to store.
Requests may not be stored, so there would be nothing to replay.
Temperature is set per account, so one request cannot choose it.

Question 20· Select all that apply

Tools and MCPs

A workspace has twelve custom Skills installed, several of them carrying large reference files. The team is working out what those Skills cost on a request that triggers none of them.

Select the two statements that are true of an untriggered Skill.

Select all that apply.

Its instructions have not been read, so its body costs nothing on this request.
Its name and description are already in the context, at roughly a hundred tokens.
Its bundled reference files were loaded when the workspace started.
It is invisible to the model until a person names it in a message.
Its instructions were summarised into the system prompt when it was installed.