Home / Claude Certified Developer Foundations

Anthropic · CCDV-F

Claude Certified Developer Foundations Practice Exam (CCDV-F)

Free Claude Certified Developer Foundations practice exam questions from all 8 official domains, with every answer option explained.

Start practicing free Take a mock examSee plans

Claude Certified Developer – Foundations (CCDV-F)

Questions
53

Multiple-choice and multiple-response.

Time
120 min
To pass
720 / 1000

Scored on a 100–1000 scale.

Exam Price
$125

What is on the CCDV-F exam?

Open a domain for its objectives and their practice questions.

Try free CCDV-F practice questions

Pick an answer to see why each option is right or wrong. A preview of 4 of the 16 free questions, which cover every domain.

Domain 1: Agents and Workflows

1.1 Agent Architectureeasy

A research system with a lead agent and four parallel subagents gives good results, but its token spend is too high. Which model assignment is the recommended cost lever?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

This is the most expensive configuration; reserving the capable model for the steps that need it is the point of model selection as a lever.

Each subagent spends its own tokens, so the multiplied cost sits mostly in the workers; keeping the stronger model on the lead, which plans and synthesizes, while the workers run cheaper models cuts spend where it multiplies.

Smaller models are a valid lever where possible, which makes this tempting, but downgrading the lead too gives up the capable model on the planning and synthesis step that needs it.

This inverts the guidance: it puts the expensive model where the cost is multiplied by the number of workers and weakens the step that plans and synthesizes.

1.2 Agent Construction with Claudeeasy

An agent's refund tool is described only as "Issues refunds." Which rewritten description follows the guidance for a tool that changes data?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

This widens the trigger instead of restricting it, inviting calls on any payment mention without checking the order first.

A restrictive description tells Claude what must be true before the call and what the tool is never for, which narrows when a data-changing tool gets selected.

Pointing to an external document gives the model nothing to act on; the conditions and exclusions must be in the description itself.

This adds detail about how the tool works but still gives no precondition and no limit on when it should not be called.

Domain 2: Applications and Integration

2.2 Systems Life Cycleeasy

A regulated team wants to start building before confirming that the chosen platform can process data in the required region. Why is that risky?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Evals belong to test; the risk here is the unmet residency requirement.

The platform determines where data is processed.

Gates are where regulated engagements keep control; moving to build before residency is satisfied invites late rejection.

Version pinning is unaffected; the residency failure is the risk.

2.1 Understanding Requirementseasy

Months after launch, an auditor asks a team why their legal-research assistant runs on its current platform rather than an alternative. The engineers had picked it early because they had used it before. Which artifact, prepared during scoping, would have let them defend the choice?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

The requirements record lists the functional behaviors, infrastructure constraints and the regulation behind each. Because it feeds the deployment decision, it shows the platform followed from requirements rather than from familiarity.

An eval suite is a real scoping output and shows the assistant behaves correctly, which makes it tempting. But it checks behavior, not placement; it says nothing about why one platform meets the constraints and another doesn't.

This documents familiarity, which is exactly the reason the auditor's question exposes. A defensible choice has to follow from requirements, not experience.

A benchmark looks like hard evidence, but a laptop measurement doesn't reflect the users' region and covers only one constraint. It can't explain the platform choice as a whole.

The other 12 free questions, by domain:

Domain 3: Claude Code · 2 questions
3.1 Claude Code Operationeasy

A developer creates .claude/rules/frontend/components.md with no YAML frontmatter, expecting it to load only when Claude works on frontend files. What actually happens?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Scoping comes from a paths glob in the frontmatter, not the folder; without paths the rule loads every session.

Rules files without paths do load, at launch.

Subfolders in .claude/rules are only for organization; they do not scope anything.

Rules files are not invoked by name; that is Skill behavior.

3.1 Claude Code Operationeasy

A developer wants to allow a few extra commands for their own use in one project without affecting teammates or committing the change. Where should the setting go?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

The project file is committed, so the change would apply to everyone who clones the repo.

User settings apply to every project on the machine, not just this one.

Local project settings are personal overrides for one project and are git-ignored automatically.

Enterprise managed settings are set by admins for the whole organization.

Domain 4: Eval, Testing, and Debugging · 2 questions
4.1 Debugging and Error Handlingeasy

A model-graded judge returns scores clustered around 6 out of 10 for almost every answer. What change most directly addresses this?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

A wider scale does not change the judge's tendency toward the middle.

Without structured reasoning first, judges drift toward a safe middle score; calibration checks agreement with humans.

More cases do not fix a judge that cannot discriminate.

Exact match cannot grade open-ended quality.

4.1 Debugging and Error Handlingeasy

A dependency returns a status code the team has never seen and cannot yet classify. According to sound failure-handling practice, how should the code treat it until it is understood?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Silently swallowing errors leads to confident, wrong downstream results.

When unsure, treat an error as terminal: a wrongly terminal error gets fixed, while a wrongly retriable one hammers the service and hides the problem.

A refusal is a specific stop_reason, not a catch-all for unknown codes.

Guessing retriable risks repeated identical failures that hide the real cause.

Domain 5: Model Selection and Optimization · 2 questions
5.1 LLM Fundamentalseasy

A team is choosing a model for a new production workload and has no evaluation data yet. What starting point does the recommended practice suggest?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Defaulting to the most capable model is a common and expensive mistake.

Reasoning mode is a separate per-call setting, not the basis for choosing a tier.

Stepping down to Haiku should follow an eval showing the quality drop is acceptable.

Sonnet is the balanced default; tier changes should follow eval evidence.

5.2 Technical Fundamentalseasy

A developer is porting a Python service that uses AsyncAnthropic to TypeScript and looks for the matching async client class in the TypeScript SDK. What should they use?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

The TypeScript SDK is Promise-based rather than blocking, so dropping to raw REST gains nothing; both reach the same API anyway.

Carrying the Python pattern over is natural, but AsyncAnthropic is the Python SDK's async client. The TypeScript SDK has no separate async class.

Batches solve bulk offline work at lower cost, with long turnaround. A port of an async client needs requests that still return in real time.

In TypeScript the regular Anthropic client already returns Promises, so non-blocking concurrency comes from awaiting its calls; there is no extra class to find.

Domain 6: Prompt and Context Engineering · 2 questions
6.1 Context Engineeringeasy

An accessibility service receives a single product image, writes alt text for it, returns the text and ends. No request ever refers back to an earlier one. Which memory scope fits this agent?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

External storage is the right scope when state must survive across sessions or users. Here no request reuses earlier state, so it only adds retrieval latency and engineering for data that is never read back.

Keeping everything in context feels safe, but cost then scales with history length. Unrelated past jobs add tokens to every call without helping the current image.

Summarized memory helps a long-running conversational agent that has outgrown its token budget. These jobs are not a conversation, so a summary of unrelated images is pure overhead.

Stateless memory suits task agents that finish and close, and independent pipelines. Each alt-text job is self-contained, so there is no prior context worth paying to keep.

6.2 Prompt Engineeringeasy

A support assistant stays on topic for the first few turns, then gradually starts answering unrelated questions in a more casual tone. Which missing piece does this pattern point to?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

A schema controls output shape, not what the assistant is willing to discuss.

Higher temperature increases variation; it does not hold scope steady.

Scope and tone drift that worsens deeper into a conversation points to an underspecified system prompt.

Few-shot examples fix invented structure, not drifting scope and tone.

Domain 7: Security and Safety · 2 questions
7.2 Guardrails and Safe Deploymenteasy

A company with an existing zero data retention (ZDR) agreement wants to adopt Anthropic's newest, most capable model for a regulated workload. What should the team do during scoping?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Data retention is a platform policy, not something a prompt can control.

Consumer surfaces are the wrong route for a regulated workload.

Eligibility is per model and per platform; a new model may not be covered.

ZDR eligibility varies by model and platform and is not guaranteed by an existing agreement; newer models may not yet be confirmed.

7.1 AI Application Securityeasy

A team wraps fetched web content in <untrusted> tags and tells the model to ignore any instructions inside them. They consider prompt injection solved. What is the flaw in this reasoning?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Model-level defenses are probabilistic; least privilege and enforced controls bound the damage when an injection lands.

No agent reading untrusted content is fully immune; delimiters alone are not a security control.

There are no reserved tag names; the issue is that any delimiter is a soft boundary.

Injections commonly arrive through fetched pages, documents and tool results.

Domain 8: Tools and MCPs · 2 questions
8.1 Tool Implementationeasy

Claude returns two tool_use blocks in one assistant turn. How must the application respond?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

Results must be in the immediately following user turn.

Claude does not run tools; your application executes them.

Every tool_use needs a matching tool_result in the next user turn; results are matched by ID.

A missing tool_result causes a validation error.

8.2 MCP Server Developmenteasy

A developer connects several MCP servers in Claude Code and is worried that every tool definition will be loaded into context as soon as a session starts. What does Claude Code do by default?

Included in Sample: sample every domain.See plans

Choose an answer to check it. Afterward, select any option to review its explanation.

By default Claude Code defers MCP tool definitions and uses a search step to load only the tools a given task needs, which keeps unused definitions out of context.

Server order has no bearing on loading. Tools from any connected server can be found by the search step when a task needs them.

Allow rules decide whether a tool may run without a prompt. They do not decide which definitions are loaded into context.

Upfront loading is an opt-in mode, and even then it applies only while the definitions fit within about 10% of the context window.

Practice all 16 free questionsGet 150+ questions

CCDV-F question bank coverage

DomainExam weightFree sampleMaster (full bank)
1. Agents and Workflows15%222
2. Applications and Integration33%251
3. Claude Code3%210
4. Eval, Testing, and Debugging3%210
5. Model Selection and Optimization17%225
6. Prompt and Context Engineering11%215
7. Security and Safety8%211
8. Tools and MCPs11%215
Total100%16159

Looking for a Claude Certified Developer – Foundations practice exam? Start with the free questions above, then take the timed mock exam to rehearse the CCDV-F format under exam conditions.

How these CCDV-F questions are made

The domains, weights and objectives come from Anthropic's official CCDV-F exam guide (v1.0). Questions are original and written only from the vendor's official prep material, and each one cites the passage it is based on. Every question is then answered independently, without seeing its key, and every answer option's explanation is checked for accuracy before it is published.