Claude Certification Blog

Ten CCDV-F practice questions, worked option by option

These ten CCDV-F practice questions are original scenario items written to the official Exam Guide v1.0, spread across six of the eight Developer domains by weight, each with the key and a stated reason why every option holds or fails.

10 itemsSingle and select-two6 of 8 domains

20 min read

The ten CCDV-F practice questions below are not exam items and were not written from any exam item. Each is a business scenario with a measured symptom and a constraint, then four options, or five for a select-two. Seven are single response and three are select-two, the two formats the guide describes. Answer every item cold, record the misses in the map below, and open the reason for every option afterwards. Three items sit in Applications and Integration because that domain is a third of the paper.

The official CCDV-F certification page and exam guide are the authority on format and weights; the Claude Certified Developer - Foundations page carries the blueprint and six more open items. The reading method these items reward is in the Claude certification sample questions hub, and what official material exists is in the guide to Claude certification practice exams.

Why do these CCDV-F practice questions skip two domains?

Because the blueprint does. Applications and Integration is 33.1% of the exam, about 17 of 53 items; Eval, Testing, and Debugging is 2.6%, a single item, and Claude Code is 3.1%, two. Spread ten items by those weights and the two smallest domains round to zero, so the set covers the six domains that carry 94% of the paper; the free 53-item CCDV-F mock carries the other two at their exam allocation. The omission states how few marks they are worth; it is not advice to skip them.

Eight CCDV-F domains by official weight, with the items each has in this setOfficial weight · items in this setD2 Applications and Integration33.1%3 itemsD5 Model Selection and Optimization16.8%2 itemsD1 Agents and Workflows14.7%2 itemsD6 Prompt and Context Engineering11%1 itemD8 Tools and MCPs10.6%1 itemD7 Security and Safety8.1%1 itemD3 Claude Code3.1%0 itemsD4 Eval, Testing, and Debugging2.6%0 items
Weights from the official CCDV-F Exam Guide v1.0. A pale bar is a domain with no item here.

Applications and Integration: which layer actually changed?

Three items, because the domain is a third of the paper, on three published skills: API mechanics, where prompt caching pays only when the cached prefix is identical from the first token; software engineering foundations, where an idle CPU beside blocked threads is an I/O problem, not a capacity problem; and configuration management, where a floating model alias and an unversioned prompt make an upstream release an unreviewed production change. The distractors share one pattern: a bigger knob where the fault is structural.

Question 1 · Domain 2 · single response · written for this article

Quillstone, a SaaS company, added a cache_control breakpoint at the end of its support assistant's 6,000-token system prompt, expecting most of that input to be served from cache. After a week, the usage fields show cache_read_input_tokens near zero on almost every request while cache_creation_input_tokens is high. The system prompt begins with a line containing the current timestamp and the requesting user's ID, followed by the static instructions. Which change would make the cache effective?

  • A. Raise the cache time-to-live so that entries created for one user survive long enough to be matched by that same user's next request.
  • B. Move the timestamp and user ID after the breakpoint, outside the cached block, so the 6,000 static tokens are an identical prefix on every request.
  • C. Add a second cache_control breakpoint on the most recent user message so that the conversation history is cached even when the system prompt itself is not.
  • D. Pad the system prompt above the minimum cacheable length, because prompts below the threshold are silently processed without caching.
Show the answer and the reason for every option

Answer: B

Caching matches an exact prefix from the first token. Objective O06, Claude API Mechanics.

A. Time-to-live governs how long a matching entry survives. These entries never match because the first line differs on every request, so extending their lifetime leaves the hit rate at zero.

B (correct). Correct. Cache lookups compare the prompt from its first token. With the 6,000 static tokens forming the cached block and the volatile line placed after the breakpoint, every request shares the same cacheable prefix.

C. A breakpoint later in the prompt cannot rescue a prefix that already diverges at the top. Matching proceeds from the start, so the differing first line defeats every downstream breakpoint too.

D. Minimum cacheable length is a real constraint, but a 6,000-token prompt is well above it. The high cache_creation_input_tokens count shows entries are being written; length is not the problem.

Question 2 · Domain 2 · single response · written for this article

Saltmarsh Bank runs a Python service in which each web request calls the Claude API synchronously with the standard client and waits about four seconds for the reply. With 300 concurrent users, the 32 worker threads are all blocked on network waits, p95 latency has reached 11 seconds and requests queue at the load balancer. CPU on the service sits below 10 percent. The bank's platform team will not add instances or memory to the service. Which change addresses the cause?

  • A. Increase the worker thread pool from 32 to 512 so that many more requests can wait on the API concurrently without queuing at the balancer.
  • B. Cache completed responses keyed by the user's message text so repeated questions are served without an API call at all.
  • C. Switch to the asynchronous client and await each API call, so a single worker serves many in-flight requests while they wait on the network.
  • D. Move to a Haiku-class model so each call returns in roughly one second and the existing threads free up about four times faster.
Show the answer and the reason for every option

Answer: C

Network waits call for async, not more threads. Objective O07, Software Engineering Foundations.

A. More threads raise the ceiling but keep capacity tied to thread count, each one idle during a four-second wait. Every thread costs memory the instances do not have, and the queue returns at the next traffic increase.

B. Caching by message text only helps when users send identical messages, which is uncommon in a conversational service. It reduces a few calls without changing how the remaining calls block threads.

C (correct). Correct. The threads are blocked on I/O, not work. Awaiting the async client frees the worker during the network wait, so a single process can hold hundreds of concurrent calls.

D. A faster model shortens each wait, so threads turn over sooner, but the synchronous design still blocks one thread per call. The ceiling rises somewhat while the cause remains.

Question 3 · Domain 2 · select two · written for this article

Fennel Robotics, a manufacturer, runs a production service that turns inspection notes into structured quality reports. On a Tuesday with no deployment, the reports began arriving with a different section order and downstream parsing failed for six hours. The code names the model by a floating alias, which follows each new release automatically, and the prompt is a string edited in place in the source file with no record of which version was tested. Which TWO changes prevent a recurrence? (Select TWO.)

  • A. Set temperature to zero so the model's output becomes deterministic and the section order can no longer change between runs.
  • B. Pin the request to a specific dated model version and move to newer releases only after the evaluation suite passes against them.
  • C. Wrap the parsing step in retry logic with exponential backoff so transient formatting differences are absorbed before they reach downstream systems.
  • D. Version the prompt in source control alongside the model version it was validated with, so any change to either is reviewed and traceable.
  • E. Move the prompt into a CLAUDE.md file so that it is governed by the configuration hierarchy rather than by application code.
Show the answer and the reason for every option

Answer: B and D

Pin the model, version the prompt. Objective O09, Configuration Management.

A. Temperature zero narrows sampling variation within one model. It cannot prevent a new release from ordering sections differently, because the change came from the model, not from sampling.

B (correct). Correct. A dated version pin makes the model a controlled dependency. New releases are adopted when the team chooses and after the eval passes, never on a Tuesday by surprise.

C. Retries recover from transient faults. A structural change present in every response fails identically on each attempt, so backoff only delays the failure while adding load.

D (correct). Correct. Recording the prompt and its validated model version together in source control gives every change a review and a history, which is what the incident response lacked.

E. CLAUDE.md is read by Claude Code sessions, not by a production API service. Relocating the prompt there leaves the service's behaviour ungoverned and the alias problem untouched.

Model Selection and Optimization: is the extra spend buying anything?

Two items on the second-heaviest domain, both turning on a measurement the scenario already contains: an evaluation that shows no meaningful gap from extended thinking on a bounded classification, and an invoice that doubled behind a service that logs nothing. The defensible option acts on the evidence in the stem, not on a general belief about reasoning or cheaper tiers. Watch the mechanism error too: the thinking budget must fit inside the output limit.

Question 4 · Domain 5 · single response · written for this article

Oxbow Outfitters, a retailer, classifies inbound customer emails into one of six routing queues. The team enabled extended thinking with a 10,000-token budget on every call because an engineer expected better accuracy. Median latency is now 6.2 seconds against a 1.5-second target for the routing step. On a 500-email labelled set, accuracy is 96.4 percent with thinking enabled and 96.2 percent with it disabled. Which adjustment is most defensible?

  • A. Keep extended thinking enabled but switch to a Haiku-class model, whose faster generation brings the thinking phase inside the latency target.
  • B. Keep thinking enabled and lower max_tokens sharply, so the response is forced to finish inside the 1.5-second target.
  • C. Keep the 10,000-token budget, because reasoning before answering always improves accuracy and the latency target should be renegotiated with the business.
  • D. Disable extended thinking for this call, since the evaluation shows no meaningful accuracy difference on a bounded six-way classification.
Show the answer and the reason for every option

Answer: D

Thinking is a per-call dial, spent where it pays. Objective O12, LLM Fundamentals.

A. A faster tier generates thinking tokens more quickly but still generates them. The 10,000-token budget remains the dominant cost, and the evaluation shows thinking is not earning it here.

B. The thinking budget must fit within max_tokens. Cutting max_tokens sharply while keeping the budget risks a truncated or empty answer, and it does not make the deliberation itself shorter.

C. Thinking improves outcomes on hard, multi-step reasoning, not on every task. The measured 0.2-point difference contradicts the claim, and renegotiating the target pays for a benefit that does not exist.

D (correct). Correct. The evaluation is the evidence: 96.4 versus 96.2 percent on 500 emails is not a meaningful gap. Turning thinking off removes the latency cost for a task that never needed it.

Question 5 · Domain 5 · single response · written for this article

Halyard Freight, a logistics firm, received a Claude API invoice 2.4 times higher than the prior month. The service handles shipment summaries for 60 customer accounts, logs nothing about individual requests, and reports only the monthly total from the billing console. Leadership wants to know within a week which accounts and which request types drive the increase before approving any optimisation work. Which step should the team take first?

  • A. Record the input, output and cache token counts from each response's usage field, tagged with the account and request type, and aggregate them daily.
  • B. Estimate each account's consumption by dividing the total invoice across the 60 accounts in proportion to the shipment volume each one generated for the month.
  • C. Set a global max_tokens cap on every request so output spending is bounded while the team investigates where the increase originated.
  • D. Move all 60 accounts to a Haiku-class model immediately, cutting the per-token rate so the next invoice falls regardless of the cause.
Show the answer and the reason for every option

Answer: A

Measure usage per account before changing anything. Objective O15, Cost and Token Management.

A (correct). Correct. The usage field on each response is the ground truth for tokens consumed. Tagging it by account and request type is the cost model leadership needs, and it can exist within days.

B. Allocating the invoice by shipment volume assumes every shipment costs the same tokens. Long documents, retries and cache misses break that assumption, so the estimate would misattribute the increase.

C. A max_tokens cap bounds output spend going forward but says nothing about which account caused the increase, and it may cut off summaries for legitimate long shipments.

D. A blanket tier change before diagnosis trades quality across all 60 accounts to treat a cause that may sit in one. It also removes the chance to learn what happened.

Agents and Workflows: what must happen on every run?

One single and one select-two. The first is the compliance case the Agent SDK's hooks exist for: a step the model usually performs is not a step that always happens, and a regulator who accepts no probabilistic control has ruled out every option that is still an instruction. The second is a context problem disguised as a runtime problem; isolated subagent contexts and an external store for findings fix both symptoms at once.

Question 6 · Domain 1 · single response · written for this article

Wrenfield Health, a hospital group, uses a coding agent, built with the Claude Agent SDK, to edit interface-mapping files. Compliance requires that every file write be checked by a schema validator before it is accepted. The team currently tells the agent in its system prompt to run the validator after each edit, and audit logs show the step was skipped in six percent of runs last month. Compliance will not accept any probabilistic control. Which implementation meets the requirement?

  • A. Add a reviewing subagent that inspects every edit after the main agent finishes and runs the validator before the change is committed.
  • B. Move the instruction into CLAUDE.md and also repeat it at the top and bottom of the system prompt to give it more salience with the model.
  • C. Register a hook on the file-write event that runs the validator in code and rejects the write on failure, independent of the model.
  • D. Upgrade the agent to a top-tier model, on the basis that stronger models follow multi-step procedural instructions with far fewer omissions than lighter ones.
Show the answer and the reason for every option

Answer: C

A required step belongs in a hook, not a prompt. Objective O02, Agent Construction with Claude.

A. A second agent is another probabilistic component. It will catch most skipped validations and miss some, so the audit finding changes from six percent to a smaller number rather than to zero.

B. Repeating an instruction and placing it in CLAUDE.md increases the chance the model follows it. It remains a request, and compliance has stated that a request is not an acceptable control.

C (correct). Correct. A hook runs deterministic code on the tool event, every time, regardless of what the model decides. Blocking the write on validation failure satisfies a compliance requirement that rejects probabilistic controls.

D. A more capable model omits steps less often, which improves the number without meeting the requirement. Compliance asked for a guarantee, and no tier turns an instruction into an enforcement point.

Question 7 · Domain 1 · select two · written for this article

Marlow and Tice, a law firm, built a due-diligence agent that reads about 300 documents for each acquisition in a single agentic session. By roughly the 120th document the context window is full, findings from early documents are no longer reflected in the final report, and a run takes four hours. The partners need every document's findings in the report and want the run under an hour. Which TWO patterns address both problems? (Select TWO.)

  • A. Fan the documents out to subagents that each process a batch in their own context and return a condensed findings summary to an orchestrator.
  • B. Raise max_tokens on every call so the agent can hold more of the earlier documents' findings in its working output.
  • C. Migrate the agent from the Claude Agent SDK to a graph-based framework, on the basis that such frameworks manage long-document context on the agent's behalf.
  • D. Instruct the agent to keep its intermediate reasoning brief so that more documents fit into the session before the window fills.
  • E. Have the agent write each document's findings to an external store as it goes, then assemble the report from that store rather than from the window.
Show the answer and the reason for every option

Answer: A and E

Isolate context per subagent and persist findings. Objective O03, Agent Patterns and Frameworks.

A (correct). Correct. Each subagent works in an isolated context, so no single window carries all 300 documents, and independent batches can run in parallel, which is what brings the runtime down.

B. max_tokens limits how long a single response may be. The context window stays the same size and earlier findings are not preserved, so the loss at document 120 continues unchanged.

C. A framework changes how the loop is expressed, not how many tokens a window holds. Without a fan-out or memory pattern, a graph-based agent hits the same limit at the same point.

D. Terser reasoning postpones the overflow by some documents, but the 300-document total still exceeds one window, and shorter reasoning does nothing to shorten a four-hour sequential run.

E (correct). Correct. Persisting findings outside the window as they are produced makes the report independent of context retention. Early documents' findings survive because they were written down, not remembered.

Security and Safety: what reduces the exposure before the audit?

One select-two item on a domain worth 8.1% of the exam. The exposure is not the model call but the copy of every prompt and completion that hundreds of staff can read for two years. The published skill is data leakage prevention and PII handling, and the two options that meet it act on the data and on the audience. The distractors act elsewhere: on the model's output, on the storage medium, on a credential nobody has exposed. Encryption at rest is a real control aimed at a threat the stem never describes.

Question 8 · Domain 7 · select two · written for this article

Northwater Mutual, an insurer, runs a claims assistant whose prompts include policyholder names, dates of birth and medical notes. Every prompt and completion is logged in full to an analytics dashboard that 300 staff across the company can open, and the logs are retained for two years. A regulator's data-protection audit is scheduled in 60 days. Which TWO changes most directly reduce the exposure? (Select TWO.)

  • A. Insert a system-prompt rule forbidding the model from repeating a policyholder's personal details in any completion it produces.
  • B. Encrypt the log storage at rest while leaving the dashboard's access list and the two-year retention period unchanged.
  • C. Redact or tokenise personal identifiers before prompts and completions are written to the log, so the dashboard never holds raw policyholder data.
  • D. Restrict the dashboard to the roles that need it for their work and record who accesses which logs, replacing company-wide open access.
  • E. Rotate the Claude API key and move it into a secrets manager, so the credential that sends policyholder data to the API cannot itself be misused.
Show the answer and the reason for every option

Answer: C and D

Minimise stored PII and restrict who can read it. Objective O19, AI Application Security.

A. An output instruction does not touch the prompts, which already carry names, dates of birth and medical notes into the logs. The dashboard exposure is unchanged.

B. Encryption at rest defends against someone obtaining the disks. Authorised dashboard users still read plaintext, and 300 of them can, so the audit finding remains.

C (correct). Correct. Removing or tokenising identifiers before the log write means the dashboard holds no raw policyholder data. It reduces the exposure at its source rather than around its edges.

D (correct). Correct. Company-wide access to medical notes is an authorization failure. Role-based restriction with access logging limits who can read the data and shows the regulator that access is controlled.

E. Key hygiene matters, but nothing in the scenario suggests the key is exposed, and a misused key would not read the logs. Rotating it removes no personal data from the dashboard and narrows no one's access.

Tools, MCPs and output handling: where do data, identity and shape get decided?

Two single-response items from domains that together carry 21.6% of the paper. The MCP item is a transport decision: where the data lives, where credentials may live and whose identity each query must carry settle the choice between a local process and a remote server; the closest miss authenticates the server when the constraint was the person. The output item is the structured-output contract in miniature: a schema through a forced tool call, validation before insert and a retry path.

Question 9 · Domain 8 · single response · written for this article

Ashgrove University wants 200 registry staff to query its student-records engine from Claude Desktop through an MCP server. The engine runs only inside the university data centre, staff laptops hold no database credentials, and the security office requires that every query be authenticated against the staff member's own identity and logged centrally. Which deployment fits these constraints?

  • A. Run a stdio MCP server on each laptop that connects to the engine over the campus VPN using a shared service credential.
  • B. Host a remote MCP server in the data centre over streamable HTTP with per-user authentication, and point each Claude Desktop at it.
  • C. Run one remote MCP server in the data centre under a single service account, and have each query carry the staff member's name as a parameter for the central log.
  • D. Export a nightly snapshot of the student records to a local file and have a stdio MCP server on each laptop read from it.
Show the answer and the reason for every option

Answer: B

Match the transport to where data and identity live. Objective O24, MCP Server Development.

A. A stdio server runs locally and would need the database credential on each laptop, which the constraints forbid. A shared credential also makes every query look like the same user.

B (correct). Correct. A remote server keeps credentials and logging inside the data centre, authenticates each staff member individually, and serves 200 clients through one maintained endpoint. Every stated constraint is met.

C. Central and free of laptop credentials, but a name carried as a parameter is asserted by the client, not authenticated. Any caller could log queries under a colleague's name, so the identity requirement is not met.

D. A nightly export places student records on 200 laptops, serves data up to a day old, and loses per-query identity and central logging. It fails the security requirements outright.

Question 10 · Domain 6 · single response · written for this article

Harrow Vale Transit, a public agency, uses Claude to extract incident reports into JSON for a database. About three percent of responses either open with a sentence before the JSON or contain a severity value outside the allowed set, and the nightly ingest job crashes on the first bad record, delaying every report behind it. The agency needs no report lost and no operator attention overnight. Which approach is most defensible?

  • A. Force a tool call whose input schema defines the fields and the severity enum, validate each response against it, and retry failures.
  • B. Strip any text before the first opening brace and after the last closing brace before parsing, so an introductory sentence can no longer break the ingest job.
  • C. Strengthen the prompt with an explicit instruction to return only JSON with no preamble, list the permitted severity values, and re-run the nightly job.
  • D. Wrap the database insert in exception handling so a malformed record is logged and skipped and the remaining reports still load overnight.
Show the answer and the reason for every option

Answer: A

Constrain the shape, validate it, never trust it blind. Objective O18, Output Handling.

A (correct). Correct. The tool schema constrains shape and the enum at generation time, validation catches residual errors, and retrying repairs them. No record is dropped and no operator is needed.

B. Trimming to the braces fixes the preamble case but leaves a severity value outside the allowed set untouched, so the ingest job still crashes on those records.

C. Stronger wording raises compliance but remains probabilistic. The three percent failure rate is what happens when formatting is requested rather than enforced by a schema.

D. Skipping malformed records keeps the job alive by discarding reports, which the agency has said is unacceptable. It converts a crash into silent data loss.

How to score these CCDV-F questions: the miss map

Score exactly: a select-two item counts only when both letters match. Then write each miss into its domain row rather than adding up a total, because ten items cannot estimate a scaled score. What they can do is point. One miss in a domain means reread the skill named in that row in the official guide. Two or more means the domain link, which opens the free runner's untimed review filtered to that domain, before you sit the full timed paper.

DomainItemsYour missesObjective to rereadPractise the domain
D1 Agents and Workflows2___ of 2O02 Agent Construction with Claude; O03 Agent Patterns and FrameworksDomain 1 in the free mock
D2 Applications and Integration3___ of 3O06 Claude API Mechanics; O07 Software Engineering Foundations; O09 Configuration ManagementDomain 2 in the free mock
D5 Model Selection and Optimization2___ of 2O12 LLM Fundamentals; O15 Cost and Token ManagementDomain 5 in the free mock
D6 Prompt and Context Engineering1___ of 1O18 Output HandlingDomain 6 in the free mock
D7 Security and Safety1___ of 1O19 AI Application SecurityDomain 7 in the free mock
D8 Tools and MCPs1___ of 1O24 MCP Server DevelopmentDomain 8 in the free mock
Four steps from answering the ten items cold to acting on the misses by domain1 · Answer all ten coldCommit to a letter before any reveal2 · Count misses by domainOne row per domain in the table3 · One miss: reread the skillThe objective named in that row4 · Two or more: run the filterThe domain link, untimed, then timed
A Cred Farmer routine for a ten-item set. It produces a reading list, not a readiness verdict.

The wrong option is usually a bigger knob

Across these ten, the most tempting distractor makes an existing number larger: a longer cache lifetime, more threads, a higher tier, a stricter instruction. The key changes what breaks. Before you pick, ask whether the option alters the mechanism or only its size.

The set pairs with the two-week CCDV-F study plan, which schedules the domains in weight order, and the CCDV-F cheat sheet, whose decision rules these items exercise. The alternative, a memorised answer key, is covered in what Claude certification exam dumps actually cost you. The CCDV-F blueprint sets the weights used here. Checked against the official CCDV-F Exam Guide v1.0 (July 2026) on 17 September 2026.

Key takeaways

  • Ten original items, six domains, by weight. Applications and Integration gets three because it is a third of the exam; the two smallest domains live in the full mock.
  • Answer cold, then reveal. Reading the reasons before committing turns a diagnostic into a reading exercise.
  • Score by domain, not by total. Ten items cannot estimate the scaled score; a miss pattern in one domain can name the skill to reread.
  • Distrust the bigger knob. Longer, more, higher and stricter are the shape of most distractors here; the key changes the mechanism.
  • Read the constraints as a checklist. Every distractor in the MCP and security items fails at least one stated constraint; the key fails none.

Now sit the full 53 under a clock

The free CCDV-F mock is 53 timed items across all eight domains with a reason for every option, no account needed. Signed in, the account adds 5 aligned timed forms, 265 items, plus drills by objective. Nothing reports a scaled score or a pass probability.

Start the free CCDV-F mock

Questions

Frequently asked

The follow-up questions people search next.

Are these real CCDV-F exam questions?

No. All ten were written fresh for this page from the published CCDV-F Exam Guide v1.0 and its domain skills, and none is drawn from, or paraphrased from, any exam form or any gated Cred Farmer form. Anthropic keeps its live items confidential; a page offering CCDV-F exam questions that claims otherwise is selling a breach.

How many questions are on the CCDV-F exam?

The official guide sets 53 items in 120 minutes across eight domains, mixing single-answer and multiple-response items, with every multiple-response stem stating how many options to select. The pass mark is a scaled 720 on a 100 to 1,000 scale; the fee is $125 per attempt before partner-tier discounts.

Is there a free CCDV-F mock exam?

Yes. The Cred Farmer runner at /practice/ccdv-f sits 53 original items in 120 minutes with no account, allocated across the eight domains by official weight, and explains every option after you submit. It reports a raw count by domain, never a scaled score, a pass verdict or a percentile.

Where can I find CCDV-F sample questions?

Three places, all free: the illustrative samples inside the official exam guide, the six open items on the Cred Farmer certification page, and the ten worked items here. All three follow the same eight-domain blueprint; only the guide is official. The retired official practice exam has not been replaced.

How should I score myself on these CCDV-F questions?

Count an item right only when your letters match the key exactly, so a select-two item with one correct pick is a miss. Write the misses into the map by domain and act on the row, not the total. Ten items say nothing about the scaled score, and no practice percentage converts to it.

Keep reading

Related posts

Not affiliated with, or endorsed by, Anthropic or Pearson VUE. Details are summarised from publicly published program information and can change. Always confirm against the official exam guide before booking.

We use cookies and privacy-friendly analytics to understand usage and improve Cred Farmer. Essential features work either way. See our Cookie Policy.