200
Scenario-based questions
8
Domains at the guide's weights
53
Question mock exam, 120 minutes
10-day
Money-back guarantee
Practice Questions by Exam Domain
The bank follows the domain weights in Section 6 of the exam guide (CCDV-F).
Applications and Integration
Model Selection and Optimization
Agents and Workflows
Prompt and Context Engineering
Tools and MCPs
Security and Safety
Claude Code
Eval, Testing, and Debugging
Practice Modes
Timed at the exam's own pace: 120 minutes for 53 items.
- Quick Sprint: 10 questions in 23 minutes; free without an account, drawn from a 28-question sample of the bank
- Focused Session: 20 questions in 46 minutes; the first five need no account and come from the 28-question sample
- Deep Practice: 30 questions from the full 200-question bank across all eight domains, in 68 minutes
- Domain Drill: up to 20 questions from a single exam domain, or the whole domain where it is smaller, timed at the exam's pace
- Full Mock Exam: 53 questions in 120 minutes, drawn fresh each attempt at the guide's domain weights: 17 Applications, 9 Model Selection, 8 Agents, 6 Prompt and Context, 6 Tools, 4 Security, 2 Claude Code, 1 Eval
Sample Questions
Three questions from the free sample, with their full explanations.
Question 1: Agents and Workflows
A product team describes its new document pipeline as an agent: a fixed sequence in which one model call extracts fields, code validates them, a second call drafts a letter, and a third call checks tone, with every branch decided in code. An architect says the label matters for how the system is tested and governed. How does Anthropic's guidance classify this system?
- A. As an agent, since any system with more than one model call directs its own process by the guidance's definition.
- B. As a workflow, since its LLM calls and tools are orchestrated through predefined code paths rather than by the model.
- C. As an agent, since the tone check lets a model judge the output, which is the defining mark of agency in the guidance.
- D. As neither, since the guidance reserves both terms for systems that call external tools during a task.
Show answer and explanation
Correct answer: B
Anthropic's guidance on building effective agents draws an architectural distinction between workflows and agents: workflows are systems where LLMs and tools are orchestrated through predefined code paths, while agents are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks; it calls both kinds agentic systems. Why more than one call is wrong: the distinction turns on who directs the process, not on the number of calls. Why the tone check is wrong: a model judging an output inside paths fixed in code is still orchestrated by code. Why neither is wrong: the guidance applies the two terms to exactly these systems and does not require external tools.
Question 2: Applications and Integration
A triage agent calls a route_ticket tool whose priority field is a free-text string. The application switches on that value and keeps falling into its default branch, because Claude sends variants such as High, high priority and urgent. The developer wants values the application can rely on. Which schema change fits best?
- A. Declare priority as an enum of the allowed values in input_schema, so the schema states what the application accepts.
- B. List the allowed values in the tool's description and keep the field a free string, since the description constrains the input.
- C. Change the field to an integer priority code that the application maps to labels, since a number cannot be misspelled.
- D. Add a system prompt instruction requiring the value in uppercase, so that the application needs a single string comparison.
Show answer and explanation
Correct answer: A
The define tools page's example tool expects an input object with a required location string and an optional unit string that must be either celsius or fahrenheit, which is the schema itself stating the allowed values, and the prompting best practices page says that for classification tasks you should use either tools with an enum field containing your valid labels or structured outputs. Why a description alone is wrong: the define tools page says the description explains what the tool does, when to use it and how it behaves, while the input_schema is what defines the expected parameters, so a free string stays free. Why an integer code is wrong: an unconstrained integer is as open as an unconstrained string, and the page's mechanism for a fixed set of values is the enum. Why an uppercase instruction is wrong: it asks the model to follow a convention instead of letting the schema define the accepted values, so the application still cannot rely on what arrives.
Question 3: Security and Safety
A CRM assistant builds each tool result by concatenating a customer record's free-text notes field into a free-form string with a label in front of it. A record edited by an outside party contained a closing tag followed by a directive addressed to the assistant, and the agent treated the rest of the notes as instructions. Which change does the guidance recommend for how the untrusted content is packaged?
- A. HTML-escape the notes text before insertion, since escaped angle brackets stop the model from reading the text as markup.
- B. Truncate the notes to their first paragraph, since injected directives sit at the end of a field where they are read last.
- C. Wrap the notes text as a string inside a JSON object, since JSON escaping gives delimiters that an attacker cannot break out of.
- D. Prefix the notes text with a warning sentence in capital letters, since the model gives a capitalized warning precedence over the content.
Show answer and explanation
Correct answer: C
The mitigate jailbreaks page says to JSON-encode untrusted content: where possible, wrap third-party strings in a JSON object rather than concatenating them into free-form text, because JSON escaping provides unambiguous delimiters between the untrusted payload and the surrounding structure, so an attacker cannot close a quote or tag to break out into an instruction context. Why HTML-escaping is wrong: escaping changes how brackets render, but the model still reads the words, and the page's mechanism is a delimited JSON payload. Why truncation is wrong: nothing fixes where an injected directive sits in a field, and truncation discards the content the assistant needs. Why a capitalized warning is wrong: the page's remedy is structural encoding, and a warning is more free-form text in the same string.
Preparing for both exams?
Get the CCAR-F bank (400 questions) and the CCDV-F bank (200 questions) together, with both mock exams, in one dashboard.
Save $19 on buying them separately
Frequently Asked Questions
How many questions are in the CCDV-F practice bank?
200 scenario-based questions across the eight domains in Section 6 of the exam guide (CCDV-F), weighted like the guide, including multiple-response items that state how many answers to select.
How is it different from the CCAR-F practice bank?
They prepare for different exams. CCAR-F follows the five Architect domains; CCDV-F follows the eight Developer domains, from the API and SDKs to agents, tools, security and model selection. The CCDV-F questions were written separately for the developer guide.
What do the explanations cover?
Every CCDV-F question explains why the correct answer is right and why each other option is wrong, with the documentation it is based on named in the explanation.
Is this the official exam?
No. We are an independent exam-prep site, not affiliated with Anthropic. Registration for the exam itself happens through Anthropic Partner Academy.
What if it isn't for me?
There is a 10-day money-back guarantee: request a full refund within 10 days of purchase.
Start Practicing Today
Jump into 200 scenario-based questions and see where you stand across all eight CCDV-F domains.
Unlock all 200 CCDV-F questions: $6910-day money-back guarantee · or try 10 free questions