I built a CCDV-F practice bank: 3 free questions inside

The Claude Certified Developer, Foundations exam (CCDV-F) is Anthropic's certification for people who build, integrate and ship applications with Claude. Section 5 of the exam guide sets the shape: 53 items in 120 minutes, a mix of multiple-choice and multiple-response, and a scaled passing score of 720 out of 1,000. Section 6 spreads those items across eight domains and 25 weighted skills.

I run an independent practice site for Anthropic's certification exams, and this month I built a 400-question bank for the CCDV-F. Below are three questions from its free sample, each with the full explanation, so you can judge the style before you spend anything. They are written from the published exam guide. They are not copies of exam content, and the site is not affiliated with Anthropic.

Question 1: Model Selection and Optimization (16.8%)

A team wants token-by-token output from the Messages API and starts writing a WebSocket client for the API host, reasoning that streaming needs a persistent two-way socket. Before investing further, the lead checks how the API actually streams a response. What does the streaming documentation describe?

• A. Streaming is delivered on the HTTP response as server-sent events, each carrying a named event type and JSON data.
• B. Streaming is negotiated by upgrading the HTTP connection to a WebSocket, after which events arrive as binary frames.
• C. Streaming is delivered by repeated HTTP polling of a message resource, with each poll returning the text generated so far.
• D. Streaming is available through the SDKs alone, which hold a private channel to the API that raw HTTP clients cannot use.

Show the answer

Correct answer: A. The streaming page says each server-sent event includes a named event type and associated JSON data, that each event uses an SSE event name such as message_stop and includes the matching event type in its data, and that the SDK's stream call keeps the HTTP connection alive with server-sent events; the versioning page records the current format as named server-sent events.

• Why a WebSocket upgrade is wrong: the page describes server-sent events on the HTTP response, and MDN describes WebSockets as a separate API for two-way sessions.
• Why polling is wrong: the page describes events arriving on one kept-alive HTTP connection.
• Why SDK-only is wrong: the page describes the SDK helper as keeping the same HTTP connection alive with server-sent events, the mechanism any HTTP client receives.

Question 2: Agents and Workflows (14.7%)

A lead research agent spawns subagents for every query. For a simple fact lookup it has spawned dozens of subagents and burned a large budget, while complex comparisons sometimes get a single under-resourced subagent. The team asks what Anthropic's write-up of its multi-agent research system did about this. What did it do?

• A. Fixed the subagent count at three for every query, since the write-up found a constant fan-out the one workable way to bound spend.
• B. Removed the subagents and ran every query through a single agent, since the write-up found that parallel agents rarely pay off.
• C. Let the subagents negotiate the fan-out among themselves, since the write-up found that a lead agent cannot judge complexity.
• D. Embedded effort-scaling rules in the prompt: one agent with a few calls for simple facts, many subagents for complex research.

Show the answer

Correct answer: D. Anthropic's write-up of its multi-agent research system says early agents made errors like spawning 50 subagents for simple queries, that agents struggle to judge appropriate effort for different tasks so the team embedded scaling rules in the prompts, with simple fact-finding requiring one agent with 3 to 10 tool calls, direct comparisons 2 to 4 subagents with 10 to 15 calls each, and complex research more than 10 subagents with clearly divided responsibilities, and that these explicit guidelines help the lead agent allocate resources and prevent overinvestment in simple queries.

• Why a fixed count is wrong: the write-up scales the count to query complexity.
• Why dropping subagents is wrong: the write-up reports the multi-agent system outperforming a single agent on its research evaluation.
• Why self-negotiation is wrong: the write-up has the lead agent allocate resources using rules embedded in its prompt.

Question 3: Prompt and Context Engineering (11.0%), multiple response

An architect reviews an agent whose designers load every document that might be relevant into the context at the start of each task, on the grounds that the model has a million-token window. Results get worse as tasks grow longer. Which two statements reflect Anthropic's guidance on context engineering? (Select 2.)

• A. Good context engineering means finding the smallest set of high-signal tokens that makes the desired outcome most likely.
• B. A larger window removes the need to curate, since recall stays constant across the whole window on current models.
• C. Context should be loaded in full at the start of a task, since retrieving data during the task interrupts the model's reasoning.
• D. As the number of tokens in the context grows, the model's ability to recall information from that context accurately drops.
• E. Sub-agents should be avoided for long tasks, since splitting work across context windows loses the shared state.

Show the answer

Correct answers: A, D. Anthropic's engineering article on effective context engineering says that because LLMs are constrained by a finite attention budget, good context engineering means finding the smallest possible set of high-signal tokens that maximize the likelihood of some desired outcome, that needle-in-a-haystack benchmarking has uncovered context rot, where the model's ability to accurately recall information decreases as the number of tokens in the context window increases, so context must be treated as a finite resource with diminishing returns, that just-in-time strategies keep lightweight identifiers and load data dynamically rather than pre-processing everything up front, and that sub-agent architectures let specialized sub-agents handle focused tasks with clean context windows.

• Why constant recall is wrong: the article describes context rot across all models.
• Why loading everything up front is wrong: the article describes just-in-time retrieval as the agentic approach.
• Why avoiding sub-agents is wrong: the article presents sub-agent architectures as a way around context limitations.

Three things worth knowing from the exam guide

Applications and Integration is a third of the exam. At 33.1%, it is the largest of the eight domains, and a good part of it is general software engineering: REST APIs, JSON, asynchronous programming, version control and code review. Working developers start ahead here.

Claude Code's 3.1% is misleading. CLAUDE.md and settings.json also appear under Configuration Management inside Applications and Integration, so skipping Claude Code costs more than the weight suggests.

Multiple-response items need a different habit. Each item states how many responses to select, and the guide does not describe partial credit. Judge every option as true or false for the situation, and pick exactly the number asked for.

Try more

The free Quick Sprint gives you 10 questions from the free sample, with no account needed. The full CCDV-F practice bank has 400 questions across all eight domains, every one explaining the correct answer and why each other option is wrong, plus a full 53-question mock exam in 120 minutes. It's a one-time $69, with a 10-day money-back guarantee.

If you're still deciding between the developer and architect exams, here's how the CCDV-F compares with the CCAR-F.

Which of the eight domains are you least sure about?