Anthropic CCDV-F (Claude Certified Developer - Foundations) Practice Test 2

Question 10

A conversation continues after a published service policy has changed since earlier messages. Which approach best meets this requirement?

  1. Reassert the current task objective concisely
  2. Label external text as untrusted input
  3. Use an external source refresh for changing facts
  4. Segment work into explicit milestones

Explanation

Correct Answer

C:

Refetching authoritative information avoids treating outdated context as eternally current; a crucial qualification is that the model should not invent a change simply because its history is old.

Incorrect Answers

A:

Reintroducing the active goal and constraints guides continuation back to the approved scope; this addresses a separate concern, not the fact that refetching authoritative information avoids treating outdated context as eternally current; repetition cannot override the need to preserve factual state accurately.

B:

Trust-boundary marking prevents source content from being promoted to application authority; here, the controlling capability is that refetching authoritative information avoids treating outdated context as eternally current, but labels should be backed by tool restrictions because prompts can be attacked.

D:

Milestones with acceptance criteria limit aimless elaboration and provide reliable checkpoints; by comparison, the intended solution relies on this: refetching authoritative information avoids treating outdated context as eternally current; a plan without tested completion criteria will still drift.

Question 11

A retailer automates a complicated refund inquiry. One component coordinates the investigation, a narrow worker scans a risky repository separately, and the refund tax must be computed by a repeatable statutory formula. Which THREE approaches cover those distinct duties? Choose THREE.

  1. Use a single supervising agent
  2. Use deterministic code for a known calculation
  3. Use a human approval checkpoint
  4. Delegate retrieval to a research worker
  5. Assign action limits in the agent harness
  6. Use a specialist subagent

Explanation

Correct Answers

A:

A supervisor holds the overall objective and coordinates narrower workers through explicit handoffs; the implementation must still recognize that its instructions and budget must prevent cycles and duplicate work among workers.

B:

Ordinary code computes the numerical result repeatably rather than trusting text generation for arithmetic; a crucial qualification is that a language model may describe the calculation but should not be its financial source of truth.

F:

An isolated worker handles a bounded analysis and returns a concise finding to the parent; the important constraint is that the subagent still needs a scoped tool set and cannot silently inherit unlimited production privileges; the isolated security worker delivers only findings to the refund coordinator rather than its raw logs.

Incorrect Answers

C:

Execution pauses before the consequential action and resumes only after an authorized reviewer agrees; by comparison, the intended solution relies on this: a supervisor holds the overall objective and coordinates narrower workers through explicit handoffs; a natural-language promise to ask permission is weaker than an enforced approval gate.

D:

A research worker gathers references and returns a compact evidence set to the lead; the needed capability is different—a supervisor holds the overall objective and coordinates narrower workers through explicit handoffs—and it should not be allowed to publish the final legal conclusion as an independent authority.

E:

Host-controlled counters and deadlines bound execution regardless of what Claude decides next, whereas the requirement calls for another behavior: a supervisor holds the overall objective and coordinates narrower workers through explicit handoffs; prompt instructions alone cannot enforce a hard external quota.

Question 12

Several teams consume a shared tool API and must understand what changed between releases. What is the most suitable pattern?

  1. Pin reproducible dependencies
  2. Remove deprecated configuration under change control
  3. Write a rollback plan
  4. Maintain a change log with impact scope

Explanation

Correct Answer

D:

Release records state compatibility, affected endpoints, and migration actions, although a model-generated summary alone does not define the actual deployed contract.

Incorrect Answers

A:

Lockfiles and controlled versions reduce unplanned changes between test and production; the missing capability is more specific: release records state compatibility, affected endpoints, and migration actions; automatically taking every new major dependency version can break stable interfaces.

B:

A controlled cleanup traces dependencies, verifies current usage, and deletes unused flags with tests; by comparison, the intended solution relies on this: release records state compatibility, affected endpoints, and migration actions; removing all old switches at once can unintentionally disable required paths.

C:

A rollback procedure identifies safe prior artifacts, migration constraints, and trigger thresholds; the needed capability is different—release records state compatibility, affected endpoints, and migration actions—and reverting an application binary may be insufficient after incompatible schema changes.

Question 13

The app shows model-produced markup on a web page and must guard against script injection. What is the correct engineering decision?

  1. Sanitize generated HTML before rendering
  2. Show uncertainty when evidence is insufficient
  3. Validate generated tool parameters
  4. Display text content while handling tool calls separately

Explanation

Correct Answer

A:

Escaping or trusted sanitization prevents untrusted content from executing in the browser; system prompts cannot provide a browser-level cross-site scripting defense.

Incorrect Answers

B:

The product communicates the gap and offers verification rather than inventing a confident conclusion, whereas the requirement calls for another behavior: escaping or trusted sanitization prevents untrusted content from executing in the browser; more assertive response wording would make an unsupported answer misleading.

C:

Server-side validation constrains types, ranges, ownership, and supported operations; however, the critical mechanism here is that escaping or trusted sanitization prevents untrusted content from executing in the browser, and a well-described function schema does not replace authorization checks.

D:

The UI renders text but dispatches structured tool blocks through the application control path; the requirement instead turns on this behavior: escaping or trusted sanitization prevents untrusted content from executing in the browser; showing a raw tool call as a final answer leaves the workflow incomplete.

Question 14

A team uses Claude Code both in a nightly pipeline and at developers' terminals. The pipeline cannot wait for prompts, but human engineers need to investigate and approve file changes interactively. Which TWO operating modes fit? Choose TWO.

  1. Use non-interactive print mode for CI automation
  2. Use a fresh isolated CI workspace
  3. Use an interactive session for exploratory editing
  4. Use CLAUDE.md for repository instructions
  5. Use structured or streaming output for pipeline consumers

Explanation

Correct Answers

A:

Headless command-line invocation can run a prompt and return consumable output under controlled settings; a crucial qualification is that CI still needs noninteractive credentials and appropriately limited permissions.

C:

An interactive coding session supports iterative commands and review as context evolves; automating every command without supervision is unnecessary for a human-led investigation.

Incorrect Answers

B:

Per-job isolation avoids file collisions and cross-run contamination; by comparison, the intended solution relies on this: headless command-line invocation can run a prompt and return consumable output under controlled settings; shared mutable directories can corrupt outputs even when prompts are correct.

D:

A project instruction file provides contextual guidance when Claude Code works in that repository, whereas the requirement calls for another behavior: headless command-line invocation can run a prompt and return consumable output under controlled settings; it is explanatory context rather than a security boundary enforcing tool privileges.

E:

Machine-friendly output formats allow programmatic consumers to handle individual events; however, the critical mechanism here is that headless command-line invocation can run a prompt and return consumable output under controlled settings, and human-oriented terminal decoration is not a stable protocol contract.

Question 15

A document analysis fails because the upstream service is temporarily unavailable but the input remains valid. Which design would meet this acceptance criterion?

  1. Show progress for asynchronous jobs
  2. Expose a safe retry action after transient faults
  3. Clearly distinguish a draft from a committed action
  4. Invalidate session privileges after role changes

Explanation

Correct Answer

B:

The UI can offer a bounded retry with clear status rather than silently losing the task; notably, the backend must still prevent duplicated side effects when retrying.

Incorrect Answers

A:

Progress state and job identifiers give a reliable view of work in flight; the requirement instead turns on this behavior: the UI can offer a bounded retry with clear status rather than silently losing the task; a fake static percentage would misrepresent actual completion.

C:

Labels and separate commit controls prevent users from assuming a draft has been submitted; in contrast, the UI can offer a bounded retry with clear status rather than silently losing the task, while model-generated wording should not make a simulated operation appear executed.

D:

Authorization checks must reflect current permissions rather than those cached at session creation; this addresses a separate concern, not the fact that the UI can offer a bounded retry with clear status rather than silently losing the task; prior approval of one request does not authorize later actions forever.

Question 16

The workflow must enforce approvals, but a bounded middle stage requires flexible information gathering. What would best address the stated failure mode?

  1. Prefer a workflow for regulated approvals
  2. Prefer a static lookup and rule table
  3. Prefer a hybrid workflow-agent design
  4. Run an explicit fixed-stage workflow

Explanation

Correct Answer

C:

Code preserves mandatory outer checkpoints while an agent explores only within an approved segment; its operational boundary is that failure to isolate the exploratory stage could let it bypass the required controls.

Incorrect Answers

A:

Explicit branches and immutable records make the process stable under regulatory review; the requirement instead turns on this behavior: code preserves mandatory outer checkpoints while an agent explores only within an approved segment; free-form agent planning is unsuitable for mandatory stage-order guarantees.

B:

A rule engine evaluates the defined cases predictably and yields transparent reasons; this cannot substitute for the behavior that code preserves mandatory outer checkpoints while an agent explores only within an approved segment, because a language model would introduce avoidable nondeterminism into a deterministic policy.

D:

Application code controls the predetermined stage transitions and makes them reproducible in tests; this addresses a separate concern, not the fact that code preserves mandatory outer checkpoints while an agent explores only within an approved segment; it cannot decide to introduce an unplanned research step after unexpected evidence appears.

Question 17

A compliance workflow must preserve a legally binding threshold across context reductions. Which solution should the engineer adopt?

  1. Retain exact quoted requirements when necessary
  2. Keep pointers to source documents
  3. Start a new clean window after task boundary
  4. Use targeted tool requests instead of bulk dumps

Explanation

Correct Answer

A:

Exact critical terms should remain intact as source data rather than be loosely paraphrased, although a small numerical wording change can reverse a regulatory decision.

Incorrect Answers

B:

References and identifiers enable later retrieval without copying the entire source into every prompt; however, the critical mechanism here is that exact critical terms should remain intact as source data rather than be loosely paraphrased, and pointers alone do not make an inaccessible document available at inference time.

C:

Separating independent sessions avoids accidental carryover of assumptions and private context; the requirement instead turns on this behavior: exact critical terms should remain intact as source data rather than be loosely paraphrased; the new session must still receive any relevant persistent project policy.

D:

Narrow queries reduce context consumption and sensitive data exposure simultaneously; this cannot substitute for the behavior that exact critical terms should remain intact as source data rather than be loosely paraphrased, because returned results still need authorization before the query executes.

Question 18

Two subagents report inconsistent current limits for the same API capability. Which development approach should be selected?

  1. Detect and resolve conflicting worker findings
  2. Use a reducer with a fixed schema
  3. Observe a tool result before choosing the next action
  4. Propagate unresolved uncertainties

Explanation

Correct Answer

A:

The coordinator compares sources, dates, and evidence before emitting a final claim—with this limitation: voting by worker count is not a substitute for checking authoritative documentation.

Incorrect Answers

B:

A merger validates and combines structured outputs rather than concatenating arbitrary prose; this cannot substitute for the behavior that the coordinator compares sources, dates, and evidence before emitting a final claim, because different worker vocabularies can break aggregation unless normalized first.

C:

Alternating action and observation permits evidence-driven choices rather than a fully predetermined plan; in contrast, the coordinator compares sources, dates, and evidence before emitting a final claim, while every loop must check for stop conditions and tool failures.

D:

The aggregate output carries gaps and confidence limitations from each contributor; this addresses a separate concern, not the fact that the coordinator compares sources, dates, and evidence before emitting a final claim; a polished synthesis should not erase unsupported assumptions.

Question 19

A deployment modifies safety-related prompts for a large population and impact is uncertain. Which action most directly closes the gap?

  1. Run adversarial prompt-injection evaluations
  2. Perform a least-privilege permissions review
  3. Use a policy-based refusal and safe redirect
  4. Canary a risky policy change

Explanation

Correct Answer

D:

A limited rollout measures refusal, error, and unauthorized action attempts before broad exposure; immediate all-user changes make regressions harder to contain.

Incorrect Answers

A:

An abuse test suite probes trust boundaries and confirms tools remain protected; here, the controlling capability is that a limited rollout measures refusal, error, and unauthorized action attempts before broad exposure, but passing a few normal demonstrations does not measure adversarial resilience.

B:

Reviewing actual credentials and granted scopes shrinks the impact of a compromised prompt or tool; the missing capability is more specific: a limited rollout measures refusal, error, and unauthorized action attempts before broad exposure; the prompt's intended task does not limit the external credential automatically.

C:

A policy layer handles the request appropriately and communicates a safe boundary; by comparison, the intended solution relies on this: a limited rollout measures refusal, error, and unauthorized action attempts before broad exposure; policy adherence should be evaluated; it is not guaranteed by a single warning prompt.

Question 20

A regulated advice assistant lacks enough verified evidence to make a material recommendation. What should the team do at this decision point?

  1. Provide an escalation channel for uncertain outcomes
  2. Show progress for asynchronous jobs
  3. Clearly distinguish a draft from a committed action
  4. Use CLAUDE.md for repository guidance in Claude Code

Explanation

Correct Answer

A:

Escalation transfers context to an authorized person instead of fabricating a final decision, although human review must be a real operational path, not only a sentence in a prompt.

Incorrect Answers

B:

Progress state and job identifiers give a reliable view of work in flight; however, the critical mechanism here is that escalation transfers context to an authorized person instead of fabricating a final decision, and a fake static percentage would misrepresent actual completion.

C:

Labels and separate commit controls prevent users from assuming a draft has been submitted; the requirement instead turns on this behavior: escalation transfers context to an authorized person instead of fabricating a final decision; model-generated wording should not make a simulated operation appear executed.

D:

Repository instruction files provide persistent project context to the Claude Code environment; this cannot substitute for the behavior that escalation transfers context to an authorized person instead of fabricating a final decision, because they are not automatically loaded into unrelated Messages API clients.

Claude Certified Developer - Foundations (CCDV-F) Practice Tests