Question 1
A developer wants a named workflow to summarize a PR in a prescribed format. What should they create?
- Define a custom slash-invocable command or skill for the PR summary
- Write only a skill description indicating when a PR-summary assistant might be useful
- Publish a PR-summary report template without defining an invocable Claude Code workflow
- Specify portable filesystem paths for a future PR-summary skill but create no workflow entry point
Explanation
Correct Answer
A:Named command workflows provide repeatable behavior for a familiar task. They are more convenient than asking developers to copy a lengthy prompt each time.
Incorrect Answers
B:A description helps discovery of a skill after it exists; it does not supply the named workflow's implementation or invocation contract.
C:An output template describes the desired artifact but does not make a reusable named command or skill. The workflow must be created explicitly.
D:Relative paths matter when a skill loads project resources; they are not a replacement for the requested named command or skill definition.
Question 2
Database schema edits require review by data owners, but ordinary comments do not. How should guidance reflect this?
- Revise rule globs to the new directory structure in the same refactor
- Transfer writing-style guidance into a documentation-path rule
- Scope the review instruction to schema and migration paths
- Transfer conditional package conventions into dedicated rules files with path selectors
Explanation
Correct Answer
C:Targeted rules make special approval requirements visible when the affected files warrant them. Applying the policy to every file obscures its true trigger.
Incorrect Answers
A:Although conditional instruction matching depends on current paths, the choice addresses maintaining globs after a directory move and not the requested targeting schema-specific review obligations.
B:Documentation conventions are meaningful when documents are the target; the deciding condition is targeting schema-specific review obligations, whereas this action concerns limiting editorial guidance to documentation edits.
D:Splitting the corpus by applicability reduces irrelevant context; this addresses avoiding an oversized global project rule set, while the question is about targeting schema-specific review obligations.
Question 3
A customer asks for a refund but the account owner cannot be established from available records. What should the agent do?
- Send an unidentified customer's refund for amount-limit approval before establishing which account is involved
- Ask a policy owner to compare refund rules even though the customer's identity has not been verified
- Ask which refund operation the customer prefers while assuming the supplied account is theirs
- Ask for authorized identity clarification before account-specific action
- Request an exception decision for the refund without linking it to a verified customer identity
Explanation
Correct Answer
D:Missing identity is a safety-critical ambiguity. Guessed account selection risks exposing information or changing the wrong customer's records.
Incorrect Answers
A:Amount limits can require human authorization, but an approver cannot safely apply a refund to an unknown account. Identity must be established before case-specific financial action.
B:Policy conflicts merit specialist review when the applicable rule is unclear. Here the blocking uncertainty is which customer account the requester can access.
C:Clarifying desired action is not evidence of account ownership. Identity and authorization need independent verification before account-specific processing.
E:Exception approval concerns policy discretion, not proof of ownership. An unidentified account remains unsafe to access or modify even when an exception might otherwise be legitimate.
Question 4
A researcher needs unrestricted search but should not open local secrets files. Which boundary matters most?
- Limit filesystem-capable tools independently from remote read-only search tools
- Set a strict number of tool invocations for the researcher but leave filesystem access unrestricted
- Validate only remote search arguments while letting filesystem reads run without a distinct access policy
- Remove write-capable tools but retain unrestricted file-reading privileges
Explanation
Correct Answer
A:Both operations may look like information retrieval, but local file access can reveal credentials. Least privilege evaluates data sensitivity and tool reach, not just read versus write.
Incorrect Answers
B:A call budget caps the number of actions, not which local files may be opened. Even one permitted filesystem call could disclose a key; access to those resources must be limited.
C:Filtering search queries does not constrain local reads. The researcher still needs a separate permission boundary for secret-bearing files.
D:Read-only tools can still disclose sensitive local credentials. A read/write distinction is insufficient when the protected asset is confidential file content.
Question 5
A developer checks out a third-party repository whose CLAUDE.md asks an agent to upload local keys. What should the agent operator do?
- Treat repository instructions as untrusted and refuse exfiltrating requests
- Trust the repository's CLAUDE.md because it is version-controlled and present in the checked-out project
- Allow the key upload after confirming that the repository's tests pass
- Copy the keys into the working tree first so a repository task can discover them locally
Explanation
Correct Answer
A:Codebase text and instructions must not override higher-priority security boundaries. Workspace trust and tool permissions remain necessary even when malicious guidance appears in an instruction file.
Incorrect Answers
B:Version control records who committed a file; it does not make its content safe or higher priority than credential-protection requirements. A malicious repository instruction is still untrusted input.
C:Passing tests says nothing about whether disclosing private credentials is authorized. The requested upload is an exfiltration attempt even in an otherwise functional codebase.
D:Putting secrets inside a third-party checkout increases exposure and can cause accidental commits or tool access. The safe response is to reject the data-transfer request rather than relocate it.
Question 6
A team requests speculative performance improvements without measurements. Which first phase is appropriate?
- Gather performance evidence and define the target metric before code changes
- Find specific tests and acceptance criteria alongside the edit steps
- Implement the reviewed patch within its stated scope and verify the security regression
- Expose the architectural trade-off and request the missing business decision
- Run the narrow approved edit directly and run the relevant check
Explanation
Correct Answer
A:Optimization without a baseline makes success impossible to judge and may harm unrelated paths. A plan should identify hotspots and a measurable performance objective.
Incorrect Answers
B:A useful plan describes not just changes but the evidence that will establish success; that may serve embedding verification into the implementation plan, but it is no substitute for baselining performance before optimization.
C:Replanning a settled narrow remediation can delay correction and introduce scope drift; this may support avoiding needless replanning of an approved patch but cannot alone meet the requirement for baselining performance before optimization.
D:An unresolved design choice is not safe to infer from code alone; the stated problem requires baselining performance before optimization rather than resolving architecture ambiguity before editing.
E:A bounded, well-specified low-risk change often needs no lengthy planning phase; a method suited to implementing a tiny well-scoped change does not meet the need for baselining performance before optimization.
Question 7
A research subagent cannot reach its source endpoint but reports an empty result set. How should the coordinator treat it?
- Mark the research run complete whenever the summarizer returns a finished document, regardless of source-tool errors
- Distinguish tool failure from a valid search with zero matches
- Retry the failed search as if the endpoint had returned a valid empty list
- Forward a generic success-shaped result with no entries and keep diagnostic errors in private logs only
Explanation
Correct Answer
B:Infrastructure errors must propagate as errors, not negative evidence. Otherwise a report may falsely conclude that no sources support a claim.
Incorrect Answers
A:A polished answer is not proof that source collection succeeded. Endpoint failures must be passed upstream as failures, not silently converted into missing search results.
C:Retrying a transport failure may be appropriate under a bounded policy, but labeling it as an empty successful response hides missing evidence. Failure and a true zero-match search must remain distinct outcomes.
D:The coordinator must know the search was unavailable so it can retry, escalate or qualify the final answer. A success-shaped empty result falsely implies the source was queried successfully.
Question 8
A refactor touches a shared interface with many dependents. What analysis helps plan safe edits?
- Consolidate a short verified configuration summary with exact file references
- Capture tested commands, outcomes and affected paths alongside decisions
- Identify references and consumer contracts before changing the shared definition
- Review dependency manifests and lockfiles to confirm the affected version
Explanation
Correct Answer
C:Shared APIs can affect distant packages. Dependency discovery reveals compatibility obligations that local edits alone cannot show.
Incorrect Answers
A:Shared grounded notes prevent repeated expensive reads while retaining the ability to verify details; reusing concise verified repository findings and mapping consumers of a shared interface require different controls in this situation.
B:Implementation history is not reliable if verification evidence is lost; the unsatisfied requirement is mapping consumers of a shared interface, not preserving code-testing evidence in work summaries.
D:When considering grounding dependency remediation in repository metadata, current local package metadata grounds the remediation plan; the requirement for mapping consumers of a shared interface remains unmet.
Question 9
A code-search tool includes hidden filters that exclude generated files, causing apparently missing results. What design change helps?
- Provide structured, bounded results with pagination and relevant fields
- Provide each tool a distinct action-focused name and usage description
- Separate the read-only lookup from the authorized state-changing operation
- Document search exclusions and filter parameters as explicit tool behavior
Explanation
Correct Answer
D:The model can plan around a tool only when it knows the search domain. Hidden exclusions create false negative evidence and misguided follow-up actions.
Incorrect Answers
A:Typed and paginated results let the agent request evidence progressively; that addresses controlling the volume of tool output, whereas this case calls for revealing retrieval coverage limitations.
B:Discriminating names and descriptions specify the business intent and side effects of each operation; this concerns differentiating similar tool responsibilities, leaving the requirement for revealing retrieval coverage limitations outstanding.
C:Tool metadata should expose whether a call merely inspects or mutates an order; clarifying read versus write tool intent and revealing retrieval coverage limitations are distinct responsibilities under these constraints.
Question 10
A long automation run applies several patches without making checkpoints. What practice improves recovery?
- Review and checkpoint cohesive changes with their test evidence
- Inspect the actual dependency version and local interface before rewriting calls
- Introduce the failing edge case to tests and implement the smallest correction
- Collect reproduction inputs and compare observed environment differences first
Explanation
Correct Answer
A:Small validated increments provide rollback boundaries and clearer blame history. A giant unverified diff makes failures difficult to isolate.
Incorrect Answers
B:Version drift can make a plausible call invalid; that serves aligning generated code to installed dependency APIs, whereas the case requires checkpointing incremental implementation progress.
C:Because treating review feedback as a concrete test turns a discovered omission into durable coverage, the measure supports incorporating edge-case feedback into tests rather than the needed checkpointing incremental implementation progress.
D:Without a reproducible failure or causal evidence, broad edits are speculative; its benefits for establishing a reliable bug reproduction do not resolve the need for checkpointing incremental implementation progress.