Anthropic CCDV-F Exam Dumps

Claude Certified Developer-Foundations

813 Reviews

Exam Code	CCDV-F
Exam Name	Claude Certified Developer-Foundations
Questions	95 Questions & Answers
Update Date	September 04, 2026
Price	Was : $106.2 Today : $59 Was : $124.2 Today : $69 Was : $142.2 Today : $79

Add To Cart

Demo Questions

Question # 1

A long-running coding agent degrades over an extended session: it repeats work it already completed and loses track of the original objective. Inspection shows the context is dominated by verbose tool output from earlier steps. Which two techniques directly address this? 

A. Prune or summarize stale tool results, retaining only the conclusions needed for subsequent steps.
B. Raise temperature so the agent explores new approaches instead of repeating old ones.
C. Maintain a compact, persistently re-stated record of the objective and completed steps, so goal state does not depend on surviving raw history. 
D. Disable tool use once the context reaches a threshold, forcing the agent to reason from memory.
E. Increase max_tokens so the agent can produce longer responses that restate everything it has learned. 

Show Answer

Question # 2

A platform team supports twenty internal Claude applications. They need central control over a small number of non-negotiable settings, team-level control over shared project conventions, and individual control over personal preferences — with predictable resolution when these conflict. Which two principles should govern the design? 

A. Use a defined precedence hierarchy in which organization-level policy overrides project settings, which in turn override       individual settings, so conflicts resolve deterministically.
B. Keep the non-negotiable settings in an enforcement layer that individual configuration cannot override, rather than             relying on convention or documentation.
C. Merge all settings into a single flat file that every team edits, so all configuration is visible in one place.
D. Resolve conflicts by whichever configuration file was modified most recently.
E. Allow individual settings to override organization policy, since developers understand their own context best.

Show Answer

Question # 3

An agent built on a custom loop works correctly for short tasks but behaves erratically on longer ones: it occasionally re-executes tools it already ran, sometimes loses the result of an earlier call, and intermittently produces malformed requests after a tool error.Which underlying implementation defect best explains all three symptoms together?

A. The loop is not faithfully maintaining conversation state — assistant turns containing tool_use blocks and the corresponding tool_result turns are not being appended consistently and in order, so the model reasons over an incomplete or malformed history.
B. The model tier is too low for long tasks and should be upgraded.
C. The temperature is set too high, causing the model to forget prior steps.
D. The context window is being exceeded, which silently drops the oldest tool results.

Show Answer

Question # 4

A team wants Claude to help modernize a large legacy codebase with sparse test coverage. The proposed plan is to have an agent rewrite whole modules in one pass and rely on manual review to catch problems. Which sequencing produces materially better outcomes?

A. Establish characterization tests that pin current behavior first, then refactor incrementally behind that safety net,                 verifying after each step.
B. Rewrite the modules first and write tests afterward against the new implementation, since the new code is cleaner to           test.
C. Rewrite everything at once but keep the legacy code in a parallel branch for comparison during review.
D. Skip tests entirely and rely on staged production rollout to surface regressions.

Show Answer

Question # 5

A developer needs Claude to normalize free-text job titles into a fixed internal taxonomy. Instructions alone yield inconsistent formatting and occasional invented categories. The developer decides to add examples to the prompt. Which approach to the examples is most effective?

A. Include a large number of examples drawn exclusively from the most common, unambiguous cases.
B. Include a smaller set of examples that spans the boundary and edge cases, shows the exact output format, and demonstrates the correct behavior when no taxonomy match exists
C. Include examples only of incorrect outputs, labeled as things to avoid. 
D. Include no examples and instead repeat the instruction three times to reinforce it.

Show Answer

Question # 6

An application must reliably extract twelve structured fields from unstructured contracts. Some fields are frequently absent from a given contract. The current implementation uses a single tool call with all twelve fields marked required, and the team observes that the model fabricates plausible values for absent fields rather than omitting them. What is the best remedy?

A. Model absence explicitly in the schema — make genuinely optional fields optional or nullable, provide an explicit "not         present" representation, and instruct the model to use it rather than infer.
B. Keep all fields required and add a post-processing step that discards values the team judges implausible.
C. Split the extraction into twelve separate model calls, one per field, all with required schemas.
D. Increase temperature so the model is less likely to commit to a fabricated value.

Show Answer

Question # 7

A regulated financial institution is deploying an agent that processes documents containing customer financial data. Policy requires that this data never leave infrastructure the institution controls, apart from the model inference call itself, which is covered by an existing approved agreement.Which deployment consideration follows most directly?

A. The agent harness, tool implementations, and any persisted agent state should run on institution-controlled infrastructure, with only the inference call crossing the boundary.
B. The agent must run entirely on the model provider's hosted infrastructure to inherit its compliance posture.
C. The requirement cannot be satisfied, since agents inherently require third-party state storage.
D. Encrypting the API key in transit satisfies the data residency requirement.

Show Answer

Question # 8

A long-running chat assistant appends every turn to the message history without bound. Users report that after roughly an hour of conversation the assistant begins returning errors instead of responses. What is the most likely cause?

A. The accumulated conversation has exceeded the model's context window, so the request is rejected.
B. The API expires conversations after a fixed wall-clock duration regardless of length.
C. The assistant has exhausted its max_tokens budget, which is consumed cumulatively across a session.
D. Long sessions trigger automatic safety filtering that blocks further responses.

Show Answer

Question # 9

A business sponsor states the requirement as: "The assistant must be right at least 95% of the time." The engineering lead pushes back before accepting it. What is the most important clarification to obtain?

A. What counts as "right" for this task, who adjudicates it, and on what distribution of inputs the 95% is measured.
B. Whether the 95% should be measured in production or in staging.
C. Whether the sponsor would accept 94% if the cost were lower.
D. Which model tier the sponsor prefers, since accuracy is a property of the model.

Show Answer

Question # 10

A prompt supplies a contract document, a set of processing instructions, and three worked examples in one undifferentiated block of prose. Claude intermittently treats sentences from the contract as instructions to follow. Which prompt change most directly addresses this? 

A. Increase max_tokens so the model has room to restate the instructions before answering.
Delimit each part of the prompt with clear structural tags, such as <document>, <instructions>, and<examples>.
C. Move the entire prompt into the system parameter so all of it carries system-level authority. 
D. Lower the temperature to 0 so the model stops improvising against the instructions. 

Show Answer

Question # 11

A team is designing a Claude-powered assistant that must serve three surfaces: an interactive web chat, an asynchronous email responder, and a batch enrichment job. All three share the same domain knowledge and business rules, but they differ in latency tolerance, output format, and available tools. Which two design decisions best serve this situation? 

A. Factor the shared domain knowledge and business rules into a single versioned prompt component reused across all         three surfaces, with surface-specific instructions layered on top.
B. Select model tier, streaming behavior, and tool surface independently per surface, since their latency and capability             requirements differ materially.
C. Use one identical request configuration across all three surfaces, so behavior is guaranteed to be consistent.
D. Duplicate the domain knowledge into three independent prompts so each team can iterate without coordination.
E. Route all three surfaces through the batch API, since a single ingestion path simplifies operations.

Show Answer

Question # 12

A Claude-powered feature has been in production for a month. The team monitors HTTP error rates and average latency only. Which two additional signals are most valuable for operating this system? 

A. Token consumption per request and aggregate spend, broken down by feature and by prompt version.
B. Output quality sampled continuously against a rubric or eval set, so silent degradation is detectable.
C. The number of characters in each system prompt, tracked over time.
D. CPU utilization of the machine that assembles the prompt string.
E. The alphabetical distribution of user query first letters.

Show Answer

Question # 13

A team must select a model tier for a new agentic workflow. Which three considerations should carry the most weight?

A. Measured accuracy of each candidate tier on a representative eval set built from the actual task, rather than on general benchmark rankings.
B. The interaction between tier and total cost in an agentic loop, where a stronger model may complete a task in fewer turns and therefore cost less overall than a weaker one that iterates more.
C. Latency requirements of the surface the workflow serves, including whether extended thinking is compatible with those requirements.
D. Which tier the team used on a previous unrelated project, for organizational consistency. 
E. Which tier has the largest number of published third-party benchmark wins.

Show Answer

Question # 14

A team is promoting a prototype agent to production. It currently runs an unbounded loop, executes any tool the model requests, logs nothing beyond the final answer, holds a single broadly privileged credential, and has no evaluation suite. Which three changes should be prioritized first? 

A. Introduce enforced bounds — an iteration ceiling, a per-task budget, and a defined escalation path when either is reached.
B. Introduce authorization at the point of execution, with least-privilege credentials and approval gates on irreversible actions.
C. Introduce execution tracing and structured telemetry — prompts, tool calls, tool results, token usage, stop reasons, and errors — so production behavior is diagnosable.
D. Introduce a requirement that the agent explains its reasoning in every user-facing response, as the primary accountability mechanism.
E. Introduce a second model that reviews the first model's output and approves it, as a replacement for human approval on irreversible actions.

Show Answer

Question # 15

A team is hardening the client layer of a high-volume Claude integration. Which three client-side behaviors most improve resilience in production?

A. Distinguishing retryable failures (rate limiting, transient server errors) from non-retryable ones (authentication, request validation) and applying retry only to the former.
B. Setting explicit request timeouts, with the value chosen to accommodate long generations rather than defaulting to a short web-service timeout. 
C. Emitting structured telemetry per request — model identifier, prompt version, token usage, latency, stop reason, and error class — to make production behavior diagnosable. 
D. Retrying every failed request indefinitely until it succeeds, to guarantee no request is lost.
E. Disabling streaming in all environments, since streamed responses cannot be logged. 

Show Answer

Question # 16

An agent's create_shipment tool times out at the network layer, so the client never receives a response. The retry logic reissues the call, and customers begin receiving duplicate shipments. What is the correct remedy?

A. Make the operation idempotent — have the caller supply an idempotency key that the shipment service uses to                   deduplicate, so a retried call returns the original result rather than creating a second shipment.
B. Remove retry logic from all tools, accepting failure on any transient network error.
C. Increase the client timeout so that slow responses are always eventually received.
D. Ask the model to check whether a shipment already exists before each retry.

Show Answer

Question # 17

A team is designing bounds for a long-running autonomous agent. Which three bounds are most important to define explicitly before deployment? 

A. A maximum number of iterations or tool calls per task, after which the agent halts and escalates rather than continuing indefinitely.
B. A token or monetary budget per task, enforced by the harness rather than requested of the model.
C. An explicit set of actions that require human approval, enforced at the point of execution.
D. A maximum length for the agent's individual responses, which is the primary safeguard against runaway behavior.
E. A rule that the agent must apologize when it makes a mistake, to maintain user trust.

Show Answer

Question # 18

A team is building retrieval over an internal document corpus to ground Claude's answers. Which two design decisions most directly affect answer quality? 

A. The chunking strategy — how documents are segmented, and whether chunks preserve enough surrounding context         to be independently meaningful.
B. How retrieved passages are presented to the model, including clear delimitation, source attribution, and instructions on       how to handle conflicting or missing evidence.
C. Whether the vector database is deployed in the same cloud region as the application.
D. Whether the embedding vectors are stored as 32-bit or 64-bit floating point values.
E. Whether the retrieval step uses the same model tier as the generation step.

Show Answer

Question # 19

A team's monthly spend has tripled with no increase in request volume. Investigation shows the agent now performs many more tool-calling turns per task than it did previously, and each turn resends the full accumulated conversation including all prior tool results. Which analysis is most accurate?

A. Cost in an agentic loop grows super-linearly with turn count, because every additional turn re-sends the entire accumulated context as input; reducing turns and pruning accumulated tool output attack the dominant term.
B. Cost grows linearly with turn count, so the tripling must be caused by a price change rather than the additional turns.
C. Only output tokens are billed in a tool-use loop, so the accumulated context is not a cost factor.
D. Prior tool results are cached automatically after the first turn, so context accumulation has no cost impact.

Show Answer

Question # 20

A team is choosing between three candidate designs for a document-processing system: a deterministic workflow, a single agent with a tool-use loop, and an orchestrator with parallel subagents. Which three factors should most influence the decision?

A. Whether the sequence of steps is known in advance or must be discovered per input.
B. Whether subtasks are independent and generate large intermediate context that would otherwise crowd a single window.
C. The degree of determinism, auditability, and per-request cost predictability the use case requires.
D. Whether the engineering team finds the agentic design more interesting to build and maintain.
E. Whether the chosen model tier supports a larger context window than competing tiers.

Show Answer

Question # 21

An incident review finds that nobody can determine which prompt text and which model produced a customer-visible bad answer three weeks ago. Which two configuration practices would have made this determinable? 

A. Record the prompt version identifier and the exact model identifier on every request log entry.
B. Store prompts as versioned artifacts with immutable identifiers, rather than editing them in place.
C. Log the full response text only, since the prompt can always be reconstructed from the current codebase.
D. Rely on the model to state which version of its instructions it used when asked.
E. Increase log retention to twelve months without adding any new fields.

Show Answer

Question # 22

A team is analyzing why an agent that performed well in testing degrades on very long production sessions, even though every request stays within the context window and no errors are returned. Which explanation is most technically accurate?

A. As context grows, relevant information competes with a large volume of lower-value content; attention is finite, so retrieval of the pertinent detail becomes less reliable even when the token count is technically admissible.
B. The model discards the oldest tokens automatically once the context is half full, so early instructions are silently removed.
C. Tokens beyond a fixed count are processed at reduced numeric precision, degrading their influence.
D. Long contexts trigger an automatic switch to a lower-capability model tier to control cost.

Show Answer

Question # 23

A high-throughput service calls Claude concurrently from many workers. Under peak load the team observes a rising rate of 429 responses, and their naive retry logic causes throughput to oscillate — long stalls followed by bursts of successful requests, then more stalls. Which combination of measures best stabilizes throughput?

A. Apply client-side concurrency limiting and request admission control so the fleet does not exceed its allocated throughput, combined with exponential backoff plus jitter and respect for retry-after on the requests that are still throttled. 
B. Increase the number of workers so more requests are in flight, ensuring capacity is always fully utilized. 
C. Remove backoff entirely and retry immediately, so throttled requests recover the moment capacity frees up.
D. Set a very long fixed sleep after any 429, applied uniformly across all workers.

Show Answer

Question # 24

A team asks Claude to perform a large-scale refactor across roughly 400 files. The agent produces a single enormous change set that reviewers cannot meaningfully assess. Which approach produces a better outcome?

A. Decompose the refactor into independently reviewable, independently testable increments — for example by module         or by mechanical transformation type — with tests run at each step.
B. Merge the change set unreviewed and rely on the test suite alone to catch defects.
C. Ask the agent to compress the diff by removing whitespace changes so it appears smaller.
D. Assign one reviewer per hundred files and merge once all four approve in parallel.

Show Answer

Question # 25

An engineering team is designing an agent to triage production incidents. The agent must gather evidence from logs, metrics, and deployment history; form a hypothesis; and either propose a remediation or escalate. Investigation paths vary widely by incident, most incidents resolve in a few steps, and a minority require deep multi-branch exploration.Which architecture best balances these characteristics?

A. A fixed three-stage workflow that always queries all three data sources in parallel and then summarizes.
B. An agent with a bounded tool-use loop, an explicit step and cost ceiling, subagents for deep exploration of a single hypothesis, and a defined escalation path when the ceiling is reached.
C. A fully autonomous agent with no step limit, allowed to investigate until it reaches a conclusion.
D. A prompt chain of ten predetermined analysis steps applied identically to every incident.

Show Answer