Question 180
An unstable page may keep sending an assistant into repeated searches with little progress. What change is best aligned with the goal?
- Require explicit approval for irreversible actions
- Set a maximum tool iteration count
- Gate a write tool behind a transaction boundary
- Configure an allowlist of available tools
Explanation
Correct Answer
B:A harness enforces a fixed ceiling on action cycles and reports incomplete execution safely—with this limitation: model self-assessment is too probabilistic to serve as the only loop brake.
Incorrect Answers
A:The side-effect layer blocks execution until an authorization token or confirmation arrives; however, the critical mechanism here is that a harness enforces a fixed ceiling on action cycles and reports incomplete execution safely, and labeling a step dangerous in the prompt does not implement authorization.
C:The application coordinates the write with transactional or compensating logic outside the model; this cannot substitute for the behavior that a harness enforces a fixed ceiling on action cycles and reports incomplete execution safely, because language-model intent is not a database rollback guarantee.
D:Its registered and authorized tool set excludes mutating functions before Claude can request them; in contrast, a harness enforces a fixed ceiling on action cycles and reports incomplete execution safely, while an instruction not to delete data is insufficient if destructive tools remain callable.
Question 181
A shipping API adds a breaking field change, and the release manager needs a recoverable production deployment plus clear guidance for downstream consumers. Which TWO release-management artifacts directly address those concerns? Choose TWO.
- Write a rollback plan
- Maintain a change log with impact scope
- Require pull-request review
- Write a design review before implementation
- Pin reproducible dependencies
Explanation
Correct Answers
A:A rollback procedure identifies safe prior artifacts, migration constraints, and trigger thresholds; the implementation must still recognize that reverting an application binary may be insufficient after incompatible schema changes.
B:Release records state compatibility, affected endpoints, and migration actions; a crucial qualification is that a model-generated summary alone does not define the actual deployed contract.
Incorrect Answers
C:Peer review examines diff, tests, and ownership before a controlled merge; by comparison, the intended solution relies on this: a rollback procedure identifies safe prior artifacts, migration constraints, and trigger thresholds; direct unreviewed edits reduce accountability and increase regression risk.
D:Architecture review assesses interfaces, failure modes, and control ownership before code changes; the needed capability is different—a rollback procedure identifies safe prior artifacts, migration constraints, and trigger thresholds—and implementing immediately can lock in unsafe assumptions and costly redesign.
E:Lockfiles and controlled versions reduce unplanned changes between test and production, whereas the requirement calls for another behavior: a rollback procedure identifies safe prior artifacts, migration constraints, and trigger thresholds; automatically taking every new major dependency version can break stable interfaces.
Question 182
A team needs to test whether Claude can accurately extract fields from messy supplier emails before building a UI. What is the best fit for the described situation?
- Prototype with a representative evaluation set
- Identify external integration contracts
- Use a threat model before tool rollout
- Implement a narrow vertical slice
Explanation
Correct Answer
A:A focused spike with real representative examples exposes feasibility and failure modes early; in production, a polished front end does not demonstrate model accuracy on difficult inputs.
Incorrect Answers
B:Interface requirements document endpoints, authentication, payloads, limits, and failure handling; the needed capability is different—a focused spike with real representative examples exposes feasibility and failure modes early—and a prose-level prompt is not a substitute for a validated service contract.
C:Mapping assets, adversaries, trust boundaries, and side effects exposes what must be protected, whereas the requirement calls for another behavior: a focused spike with real representative examples exposes feasibility and failure modes early; security should not be postponed until the tool can already send messages.
D:A small end-to-end slice tests integration boundaries and production assumptions together; however, the critical mechanism here is that a focused spike with real representative examples exposes feasibility and failure modes early, and building many disconnected components first delays discovery of interface mismatches.
Question 183
An executed lookup must be supplied back for Claude to incorporate into its next answer. Which of the following would work as specified?
- Send tool_result in a following user message
- Place reusable material before per-user content
- Register a client tool with input_schema
- Execute the returned tool_use in application code
Explanation
Correct Answer
A:The reply associates content with the tool_use identifier in the continued message history; the implementation must still recognize that a standalone text string is not the protocol-level tool result envelope.
Incorrect Answers
B:Organizing requests into a shared prefix and variable suffix improves the chance of cache hits; the missing capability is more specific: the reply associates content with the tool_use identifier in the continued message history; if the variable field appears before a checkpoint, the hashed prefix changes per request.
C:Tool definitions describe names, parameters, and types so Claude can propose structured calls; by comparison, the intended solution relies on this: the reply associates content with the tool_use identifier in the continued message history; defining a schema does not grant access to an external system automatically.
D:The integrator validates and runs the operation locally, then returns the result; the needed capability is different—the reply associates content with the tool_use identifier in the continued message history—and the Claude API cannot directly call arbitrary internal customer functions by itself.
Question 184
A document-processing service caches a long shared instruction prefix, but its hit rate drops after a schema revision and tenants make requests infrequently. Which TWO diagnostic checks explain whether cache reuse is effective? Choose TWO.
- Compute cost from input and output token usage
- Move unchanged long instructions into a shared prefix
- Compare cache TTL with request arrival pattern
- Read cache diagnostics when hit rate falls
- Remove unnecessary cache breakpoints
Explanation
Correct Answers
C:Lifetime selection must account for actual reuse frequency and write overhead; a longer cache is not automatically economical for rarely repeated prompts.
D:Cache usage metrics and prefix comparisons reveal the changed content causing misses, although increasing the context window will not restore a mismatched cache prefix.
Incorrect Answers
A:Per-model input and output token rates combine with measured usage for accurate costing; the missing capability is more specific: lifetime selection must account for actual reuse frequency and write overhead; number of messages alone is not an adequate unit of resource consumption.
B:Stable prefix design allows eligible prompt caching to amortize repeated input processing; by comparison, the intended solution relies on this: lifetime selection must account for actual reuse frequency and write overhead; putting request-specific content before a cache checkpoint defeats reuse.
E:Simpler eligible breakpoint design reduces confusion and aligns with documented limits; however, the critical mechanism here is that lifetime selection must account for actual reuse frequency and write overhead, and arbitrary markers cannot create savings if prefixes keep changing.
Question 185
A product owner wants answers that cite the exact policy clause used for each compliance conclusion. Which response is most defensible under these conditions?
- Specify source attribution for factual answers
- Document data-handling and retention constraints
- Define the authority of each tool action
- Write testable functional acceptance criteria
Explanation
Correct Answer
D:Requirements should specify observable inputs, citation fields, and acceptance tests for the behavior, subject to this requirement: a statement that the system should be helpful is not independently verifiable.
Incorrect Answers
A:Product requirements describe links or identifiers for supporting evidence and how claims map to sources; the requirement instead turns on this behavior: requirements should specify observable inputs, citation fields, and acceptance tests for the behavior; requiring a confident tone does not establish provenance.
B:Privacy requirements identify permitted collection, transmission, storage, retention, and deletion controls; this cannot substitute for the behavior that requirements should specify observable inputs, citation fields, and acceptance tests for the behavior, because developer convenience cannot override legal or organizational handling obligations.
C:Access requirements tie each permitted side effect to identity, scope, and approval evidence; in contrast, requirements should specify observable inputs, citation fields, and acceptance tests for the behavior, while a model's confidence in a request is not proof of caller authority.
Question 186
A prompt asks Claude to record a shipment adjustment and execution needs typed parameters. How should this requirement be met?
- Treat retrieved source text as evidence
- Use tool schemas for action interfaces
- Use application-controlled system instructions
- Specify a strict response schema
Explanation
Correct Answer
B:Tool definitions and server validation specify callable operations more reliably than prose parsing; its operational boundary is that a natural-language command is not an adequate machine-checked action contract.
Incorrect Answers
A:External content is data to quote or analyze only within the authorized task; this cannot substitute for the behavior that tool definitions and server validation specify callable operations more reliably than prose parsing, because a source document has no standing to instruct tool use or access policy.
C:The system-level field communicates application guidance separately from task content; this addresses a separate concern, not the fact that tool definitions and server validation specify callable operations more reliably than prose parsing; backend authorization still must be enforced in code and tools.
D:Schema constraints and subsequent validation improve consistent downstream parsing; here, the controlling capability is that tool definitions and server validation specify callable operations more reliably than prose parsing, but they cannot guarantee that the values are factually correct.
Question 187
A deployed classifier changes its decision instructions and previous results must remain reproducible. Which action most directly closes the gap?
- Assign semantic versions to prompts
- Retain a tested rollback version
- Deploy prompts with feature flags
- Test a prompt change on a frozen dataset
Explanation
Correct Answer
A:Versioned templates link each inference to the rules actually active during its execution; its operational boundary is that keeping only the newest text prevents reliable regression analysis.
Incorrect Answers
B:Keeping the previously tested artifact enables a measured rollback when safe; this addresses a separate concern, not the fact that versioned templates link each inference to the rules actually active during its execution; a rollback should consider state migrations and outstanding jobs.
C:Configurable release controls can switch a bounded population between approved prompt versions; here, the controlling capability is that versioned templates link each inference to the rules actually active during its execution, but editing production text in place without history prevents controlled reversal.
D:A regression suite compares outcomes with expected labels before promotion; the missing capability is more specific: versioned templates link each inference to the rules actually active during its execution; a valid JSON response alone cannot establish behavior equivalence.
Question 188
A screenshot contains account numbers unrelated to the visual classification requested. Which solution should the engineer adopt?
- Correct a malformed request before resubmission
- Validate MIME type and size before upload
- Redact unnecessary sensitive content before submission
- Pass an image as a typed image content block
Explanation
Correct Answer
C:Preprocessing removes irrelevant secrets and personal fields prior to sending the image content, although asking Claude to ignore a secret still transmits that data to the service.
Incorrect Answers
A:Client-side schema validation fixes the payload and prevents fruitless retries; by comparison, the intended solution relies on this: preprocessing removes irrelevant secrets and personal fields prior to sending the image content; a network backoff cannot make an invalid request body well formed.
B:Server-side checks bound resource use and ensure supported input formats before model submission; the needed capability is different—preprocessing removes irrelevant secrets and personal fields prior to sending the image content—and client-provided extensions and names are not reliable content validation alone.
D:Image blocks identify the media source while text instructions explain the intended analysis; however, the critical mechanism here is that preprocessing removes irrelevant secrets and personal fields prior to sending the image content, and putting a local path in ordinary text does not upload the image bytes to the API.
Question 189
A later agent step must act on the same shipment that an earlier search identified. How should the implementation handle this need?
- Use idempotency keys for retriable writes
- Preserve tool error codes separately from text
- Limit untrusted text in tool outputs
- Include stable identifiers in tool results
Explanation
Correct Answer
D:IDs let downstream operations refer unambiguously to retrieved business objects; its operational boundary is that human-readable names alone may collide across tenants or records.
Incorrect Answers
A:Stable transaction identifiers allow duplicate attempts to be detected before replaying the business action; however, the critical mechanism here is that IDs let downstream operations refer unambiguously to retrieved business objects, and replaying without deduplication can create multiple charges.
B:Typed errors guide safe retries, clarification, or escalation without inventing data; the requirement instead turns on this behavior: IDs let downstream operations refer unambiguously to retrieved business objects; free-form error prose can be misread or stripped by downstream clients.
C:Treating tool output as low-trust evidence and trimming unnecessary text reduces injection surfaces; this cannot substitute for the behavior that IDs let downstream operations refer unambiguously to retrieved business objects, because a tool result is not an application-level system command.
Question 190
A web worker handles dozens of independent Claude calls without tying up one thread per waiting network request. Which solution matches the actual constraint?
- Use a work queue for long offline jobs
- Use async I/O for many concurrent inference requests
- Use POST for a Messages generation request
- Cancel timed-out tasks cleanly
Explanation
Correct Answer
B:Asynchronous clients yield during I/O and resume when responses arrive; the implementation must still recognize that an event loop does not make a single model inference computationally faster.
Incorrect Answers
A:A persistent queue decouples submission from workers and supports retry state; this addresses a separate concern, not the fact that asynchronous clients yield during I/O and resume when responses arrive; keeping a web request open indefinitely is fragile under client disconnects.
C:The endpoint creates a new inference result from a JSON request and is invoked as an HTTP POST; the missing capability is more specific: asynchronous clients yield during I/O and resume when responses arrive; treating it as a cacheable read-only GET would not express the required request body.
D:Cancellation and timeout handling release resources and prevent orphaned work where supported; by comparison, the intended solution relies on this: asynchronous clients yield during I/O and resume when responses arrive; merely hiding the client spinner does not stop backend consumption.