Question 1
A moderation assistant must apply a policy rubric consistently. The model mislabels positive, negative, and borderline cases; its prompt contains no labeled demonstrations. What should the architect specify in the design?
- Include a minimal pair showing how the exception changes the judgment
- Add counterexamples that show when the rule does not apply
- Adjudicate contradictory samples and keep one coherent set
- Provide carefully labeled positive, negative, and borderline few-shot cases
Explanation
Correct Answer
D:Examples demonstrate how the rubric applies at the decision boundary and can reduce inconsistent interpretation of abstract policy text.
Incorrect Answers
A:Near-identical cases with different required outcomes can clarify the actual boundary introduced by a rule change; the alternative problem occurs when a new exception is introduced to the policy. The present scenario instead requires the system to add representative decision examples.
B:One-sided examples bias the output; contrastive cases teach the difference between related categories; it protects against the case where all examples in the prompt are easy positive cases, rather than the failure described here: the model mislabels positive, negative, and borderline cases; its prompt contains no labeled demonstrations.
C:It would be appropriate when the prompt includes examples with conflicting labels for the same situation; inconsistent demonstrations make the target behavior ill-defined and can be worse than a simpler instruction. Here, the model mislabels positive, negative, and borderline cases; its prompt contains no labeled demonstrations, so a different mechanism is required.
Question 2
An engineering assistant uses tool wrappers to standardize behavior. Two code-search tools return incompatible source-location formats. Which change most directly addresses this condition?
- Normalize both responses to a common file-and-line reference schema
- Check expected test coverage or completion markers in addition to the exit code
- Deny mutation tools while the workflow is in the read-only discovery phase
- Record the original result reference, transformation version, and normalized output
Explanation
Correct Answer
A:Common locations make evidence comparable and simplify provenance preservation across different retrieval implementations.
Incorrect Answers
B:A zero exit status alone may not prove required checks ran; outcome validation should match the acceptance criteria; it protects against the case where a test runner returns a success exit code but logs an incomplete test suite, rather than the failure described here: two code-search tools return incompatible source-location formats.
C:This choice is relevant when a request policy allows read operations but forbids writes during discovery; phase-based permissions ensure exploration cannot silently become execution before an explicit workflow transition. Here, however, two code-search tools return incompatible source-location formats.
D:Traceable normalization allows engineers to determine whether a mismatch arose in the tool, wrapper, or later model reasoning; the alternative problem occurs when a hook modifies returned data, and support needs to debug discrepancies. The present scenario instead requires the system to unify source-location fields for downstream reasoning.
Question 3
A company processes thousands of independent documents through Claude. Each document is self-contained and does not require an immediate interactive response. What is the most defensible next step?
- Record the document revision or digest with each submitted request
- Use a normal request-response call for the interactive decision
- Submit independent requests as a batch and reconcile completed results later
- Record per-request success or error and retry only eligible failures
Explanation
Correct Answer
C:Batch execution suits large offline workloads where delayed completion is acceptable and each request has its own input and output identity.
Incorrect Answers
A:This choice is relevant when the source file changes after the batch is submitted; versioned input identity ensures results are interpreted against the data actually sent, not a later file with different contents. Here, however, each document is self-contained and does not require an immediate interactive response.
B:Deferred batch completion is a poor fit when a customer is waiting for an immediate answer; workload latency requirements control the choice; it protects against the case where a support agent needs a reply within the current live conversation, rather than the failure described here: each document is self-contained and does not require an immediate interactive response.
D:A batch is not necessarily all-or-nothing; item-level bookkeeping avoids rerunning already successful documents; the alternative problem occurs when one batch item fails while the remaining documents complete. The present scenario instead requires the system to use asynchronous batch processing for suitable workloads.
Question 4
A design review automates multi-stage document and code checks. A fixed compliance check requires a predictable list of tests. Which approach satisfies the stated constraint?
- Use a join gate that requires successful results from both checks
- Use budgeted iterative retrieval with stopping conditions
- Stop dependent deployment stages and report the failed security gate
- Use an explicit test checklist and deterministic execution order
Explanation
Correct Answer
D:Stable compliance steps should be reproducible in code, with Claude assisting individual reasoning tasks rather than inventing the pipeline.
Incorrect Answers
A:Approval before every prerequisite completes risks accepting a change that violates an unchecked requirement; the alternative problem occurs when a final reviewer needs both lint and tests to pass. The present scenario instead requires the system to encode a fixed control flow for stable checks.
B:Unknown structure calls for conditional exploration, but it must not continue indefinitely after finding sufficient evidence; it protects against the case where a novel repository may require more or fewer search steps, rather than the failure described here: a fixed compliance check requires a predictable list of tests.
C:This choice is relevant when a failed security check makes downstream deployment work irrelevant; a blocker should halt stages that cannot succeed safely; continuing merely consumes resources and may create risk. Here, however, a fixed compliance check requires a predictable list of tests.
Question 5
An architect is reviewing Claude Code memory usage for a large engineering organization. An instruction is relevant to every repository maintained by one developer. Which mechanism should be selected?
- Keep the billing convention in the billing project or scoped rule
- Reference the maintained build configuration rather than hard-code versions in memory
- Put the shared standard in the repository-level CLAUDE.md
- Store the durable personal convention in the user's Claude Code instruction scope
- Move genuinely team-wide requirements into version-controlled project guidance
Explanation
Correct Answer
D:User guidance follows the developer across projects while remaining separate from team-wide repository policy. Here it serves the explicit need to choose user-level reusable guidance without adding a separate unverified inference.
Incorrect Answers
A:Narrow placement keeps tasks in unrelated services from inheriting incompatible commands or assumptions; it protects against the case where an instruction is mandatory only for the billing service, rather than the failure described here: an instruction is relevant to every repository maintained by one developer.
B:This choice is relevant when a rule changes frequently because its tool version is pinned elsewhere; the build manifest is authoritative for version details; repeating values in instructions creates stale conflicting copies. Here, however, an instruction is relevant to every repository maintained by one developer.
C:It would be appropriate when a convention applies to every service in the same monorepo; a root project instruction communicates cross-cutting conventions without repeating them in every package. Here, an instruction is relevant to every repository maintained by one developer, so a different mechanism is required.
E:Local-only user notes may not travel with the repository; team enforcement needs a shared reviewed artifact; the alternative problem occurs when a teammate edits private local guidance and expects it in code review. The present scenario instead requires the system to choose user-level reusable guidance.
Question 6
A coding assistant reviews generated patch descriptions for consistency. Some changes are wrongly described as refactors rather than bug fixes. What should be implemented at the appropriate boundary?
- Provide few-shot examples that distinguish behavior-preserving and behavior-changing patches
- Add a small labeled example covering coordinated changes across multiple modules
- Use clearly illustrative placeholders and require current identifiers from the input
- Revise labels and re-evaluate examples against the active review criteria
Explanation
Correct Answer
A:Contrastive labels teach the category distinction that ambiguous prose alone may not reliably convey. For the stated case, the check is whether the system can show contrasting labeled change summaries; this approach supplies that control.
Incorrect Answers
B:A representative complex case teaches how to summarize a unit of work that spans files without treating each edit as unrelated; the alternative problem occurs when the team's examples omit multi-file changes. The present scenario instead requires the system to show contrasting labeled change summaries.
C:It would be appropriate when the assistant copies placeholder identifiers from example output; few-shot examples guide format but must not become a source of fabricated ticket IDs or commit hashes. Here, some changes are wrongly described as refactors rather than bug fixes, so a different mechanism is required.
D:This choice is relevant when an example is accurate for old review policy but not the current standard; examples are policy-bearing inputs; inconsistency between versions can undermine otherwise precise instructions. Here, however, some changes are wrongly described as refactors rather than bug fixes.
Question 7
A code-inspection MCP server reports results and operational errors to Claude Code. The source file exists but cannot be read due to repository permissions. Which option best meets the operational need?
- Return a successful empty hit collection with the applied search scope
- Return an unsupported-encoding error and safe remediation choices
- Reject incomplete tool results before marking the inspection complete
- Report access denial with path scope but without privileged file contents
Explanation
Correct Answer
D:A permission-denied file is not equivalent to missing code; the operator may need an approved access adjustment. For the stated case, the check is whether the system can identify a read authorization problem; this approach supplies that control.
Incorrect Answers
A:Zero results is useful evidence about a bounded search; presenting it as a tool failure would trigger unnecessary recovery; it protects against the case where a search query is valid and returns no matches in the target directory, rather than the failure described here: the source file exists but cannot be read due to repository permissions.
B:An encoding-specific response directs the workflow to conversion or review instead of claiming the file contains no data; the alternative problem occurs when the file parser crashes on an unsupported encoding. The present scenario instead requires the system to identify a read authorization problem.
C:This choice is relevant when a tool reports success but omits a required file hash; structural success is insufficient when required provenance fields are missing; validation must check the actual promised result schema. Here, however, the source file exists but cannot be read due to repository permissions.
Question 8
An internal knowledge system is being exposed through the Model Context Protocol. Claude must actively run a database query with parameters. What is the most appropriate design choice?
- Publish a parameterized MCP prompt for the triage workflow
- Expose the query as an MCP tool with a defined input schema
- Expose the document as an MCP resource with suitable identity and content type
- Handle MCP content blocks according to their declared type instead of flattening everything to text
Explanation
Correct Answer
B:Tools are callable actions with arguments; a parameterized database operation is not just a static document resource.
Incorrect Answers
A:It would be appropriate when a team wants a reusable suggested workflow template for incident triage; prompts package reusable instructions and messages, whereas a tool would imply executable service behavior. Here, Claude must actively run a database query with parameters, so a different mechanism is required.
C:Resources are suited to retrievable data; they do not by themselves encode an arbitrary side-effecting invocation; it protects against the case where the assistant needs to read the contents of a stable policy document, rather than the failure described here: Claude must actively run a database query with parameters.
D:Typed content lets a client interpret data faithfully and prevents binary or structured outputs from being silently mangled; the alternative problem occurs when a connection returns JSON records and images through multiple content parts. The present scenario instead requires the system to represent an invocable operation appropriately.
Question 9
An enterprise design team compares agent organization patterns for different requests. A small deterministic classification needs no external calls. Which response is most reliable?
- Expose a narrowly defined reusable translation worker behind a clear input/output contract
- Delegate independent evidence collection then combine results in a coordinator
- Use sequential stages with explicit prerequisite checks
- Use one constrained model call instead of a multi-agent chain
Explanation
Correct Answer
D:Simple classification has no meaningful subtask parallelism; multiple agents add latency and reconciliation cost without improving the required outcome.
Incorrect Answers
A:A well-defined specialist reduces duplicated orchestration logic while keeping translation behavior consistent across callers; the alternative problem occurs when the same specialized translation step repeats across many workflows. The present scenario instead requires the system to avoid unjustified orchestration complexity.
B:This topology exploits concurrency without letting each worker independently publish a different final answer; it protects against the case where an investigation has three independent evidence streams and one final answer, rather than the failure described here: a small deterministic classification needs no external calls.
C:It would be appropriate when a legal review cannot begin until a contract-clause extractor finishes; the second stage depends on verified clauses, so parallel dispatch would let it act on absent or provisional data. Here, a small deterministic classification needs no external calls, so a different mechanism is required.
Question 10
A Claude Code team audits the quality of its persistent project instructions. CLAUDE.md contains vague advice to 'write good code' but omits repository commands. What should the team do to resolve this particular gap?
- Enforce the gate in CI or permission policy, while keeping guidance aligned
- Document the actual lint, test, and review procedures with relevant paths
- Summarize service boundaries and point to authoritative files for details
- Resolve the conflict and express one authoritative formatting rule
- Replace the secret with a safe placeholder and managed-secret guidance
Explanation
Correct Answer
B:Specific commands and expected outcomes give Claude verifiable steps; broad style slogans offer little operational guidance.
Incorrect Answers
A:This choice is relevant when one developer asks Claude Code to ignore a required security gate for speed; prompt instructions are not an authoritative security boundary; a required check should fail independently of the assistant's compliance. Here, however, CLAUDE.md contains vague advice to 'write good code' but omits repository commands.
C:A navigational map improves codebase discovery while leaving large or frequently changing implementation text in its source files; the alternative problem occurs when the team wants Claude to know repository architecture without pasting full source files. The present scenario instead requires the system to replace generic slogans with actionable conventions.
D:Conflicting directives can produce nondeterministic edits; a clear standard prevents wasted patch iterations; it protects against the case where two instruction files issue contradictory formatting requirements, rather than the failure described here: CLAUDE.md contains vague advice to 'write good code' but omits repository commands.
E:It would be appropriate when a temporary secret appears in a copied setup example; persistent instructions may be read by agents and committed to version control; secrets do not belong in them. Here, CLAUDE.md contains vague advice to 'write good code' but omits repository commands, so a different mechanism is required.