Question 280
A codebase contains tens of thousands of files, and Claude Code exhausts context by reading entire directories. What investigation pattern is more effective?
- Search narrowly, then read only relevant symbols
- Skip source inspection and rely on general framework familiarity
- Paste the entire repository into every prompt
- Modify files before learning which modules are involved
Explanation
Correct Answer
A:Selective retrieval keeps the active context focused. Directory maps and file references help explore deeply only where evidence warrants.
Incorrect Answers
B:Repository-specific behavior cannot be inferred reliably from generic knowledge. Actual code matters.
C:Unfiltered code consumes context and can bury the relevant paths. It also creates avoidable cost.
D:Premature edits risk breaking dependent components. Orientation should precede mutation.
Question 281
A developer wants the import chain for one public endpoint in a very large monorepo. Which exploration step should come first?
- Ask Claude to invent the likely call graph
- Locate the endpoint definition and follow only its referenced modules
- Read every module alphabetically before searching for the endpoint
- Perform a repo-wide refactor while mapping imports
Explanation
Correct Answer
B:A dependency-driven search narrows context to the relevant call graph. The investigator can expand only where a link requires it.
Incorrect Answers
A:A guess cannot establish actual dependencies in the current repository. Source references must be inspected.
C:Alphabetical full scans waste time and context. The entry point already provides a precise target.
D:Read-only mapping does not authorize changes. Mutation would introduce noise into the investigation.
Question 282
A coding agent frequently forgets earlier discovered file paths during a prolonged refactor. What artifact should it maintain?
- A list of filenames chosen before inspecting the code
- A claim that the entire repository has been understood
- A persistent map of files, dependencies and pending edits
- Only an unstructured copy of all previous tool outputs
Explanation
Correct Answer
C:A project map survives context compression and helps resume accurately. It should link to source rather than reproduce every file.
Incorrect Answers
A:Unverified guesses can misdirect subsequent steps. The map should be updated from actual exploration.
B:Unsupported certainty does not preserve practical navigation. Precise dependencies and gaps matter.
D:Raw transcripts are expensive and hard to search. A curated map better supports continuation.
Question 283
A team delegates code search to subagents but the coordinator is overwhelmed by verbose outputs. What return contract should be adopted?
- Unattributed code snippets with no repository location
- Full raw search logs from every worker on every turn
- Return concise findings with paths, evidence and blockers
- Only yes or no without any locations
Explanation
Correct Answer
C:The parent needs inspectable pointers rather than every search step. Scoped handoffs preserve context for synthesis and implementation.
Incorrect Answers
A:Contextless snippets cannot be checked reliably against current files. Provenance is part of the handoff.
B:Verbose logs duplicate information and consume the coordinator's context. They are better stored outside the active prompt.
D:Binary answers are insufficient for review and editing. Evidence paths make results actionable.
Question 284
A developer needs to understand how authentication flows through an unfamiliar project. What source-first strategy is best?
- Trace entry points, middleware and identity checks
- Read only files with auth in their names
- Assume every route inherits the same security behavior
- Disable tests to make exploration less expensive
Explanation
Correct Answer
A:Following the actual execution path reveals where authentication and authorization occur. It avoids assumptions based on filenames alone.
Incorrect Answers
B:Critical checks may reside in middleware or shared components with different names. A call-chain approach is broader and better grounded.
C:Routing and middleware exceptions can differ. Actual code should establish each boundary.
D:Removing test coverage does not help understand the flow and weakens later verification. Investigation should remain read-only.
Question 285
A patch spans several packages, and a context compaction drops the original acceptance tests. What should be preserved outside chat?
- Only the most recent changed file name
- A guess that all tests passed before context was lost
- A package-level task checklist tied to required tests
- Every redundant debug line from previous sessions
Explanation
Correct Answer
C:A durable progress artifact prevents loss of acceptance criteria. The agent can reconcile completed work after compaction or restart.
Incorrect Answers
A:A multi-package task needs cross-component requirements. One filename omits dependencies and verification.
B:Missing evidence is not proof of success. Test state must be recorded from actual runs.
D:Bulk transcripts are less useful than a focused acceptance ledger. Preserve what controls the next decision.
Question 286
A code exploration agent needs a summary of the repository's architecture, but not the contents of every file. What result format is suitable?
- A raw concatenation of the full repository
- A component map with entry points, boundaries and citations
- An unsupported list of modules invented for completeness
- A generic diagram based on the programming language alone
Explanation
Correct Answer
B:High-level structure with file anchors is useful for planning and can be verified selectively. It balances coverage against context budget.
Incorrect Answers
A:A giant dump is not an architecture summary and may exceed the context window. It obscures important relationships.
C:Fabricated components can mislead refactors. Unknown areas should be marked as such.
D:Language choice does not determine a project's actual architecture. The summary must reflect inspected code.
Question 287
An AI reviewer recommends deleting production customer records but says it is 95 percent confident. What control is most appropriate?
- Ask the model to raise its stated confidence to 100 percent
- Execute immediately because 95 percent sounds high
- Hide the proposed deletion from the change-review log
- Human approval and verification before irreversible deletion
Explanation
Correct Answer
D:High-impact destructive actions deserve accountable approval. Model confidence is not a substitute for permissions, evidence or impact assessment.
Incorrect Answers
A:Changing a label does not change the underlying uncertainty. The system needs external verification.
B:A numerical self-report does not certify safety. Even a small error risk is consequential for irreversible data loss.
C:Concealment undermines accountability and makes recovery harder. The decision must be auditable.
Question 288
A document extractor reports high confidence for some invoices but consistently fails on handwritten totals. How should confidence calibration be evaluated?
- Set every response confidence to high for consistency
- Check confidence against labeled handwritten examples
- Trust model confidence without measuring accuracy
- Remove difficult handwritten examples from tests
Explanation
Correct Answer
B:Empirical error rates reveal where confidence is misleading. Domain-specific evaluation can inform review thresholds.
Incorrect Answers
A:Uniform labels contain no useful risk information. Calibration should reflect observed reliability.
C:Self-reported certainty may be poorly calibrated on unusual input. A validation dataset is required.
D:Excluding the failure class hides the true production risk. The test set should represent realistic inputs.
Question 289
A team wants a human to review only invoices likely to contain material extraction errors. What routing design is suitable?
- Review only the longest invoices regardless of field quality
- Review none because the JSON schema passes
- Let the model secretly decide its own exceptions without audit
- Use risk thresholds, targeted review and periodic sampling
Explanation
Correct Answer
D:Review effort can focus on uncertain or high-impact records while samples monitor errors among auto-approved cases. Thresholds should come from measured performance.
Incorrect Answers
A:Length may correlate with difficulty but is not a complete risk signal. Evaluate evidence and business impact.
B:Shape checks cannot detect incorrect totals or missing source evidence. Semantic risk remains.
C:Unobservable routing cannot be calibrated or challenged. The decision criteria should be documented and measurable.
Question 290
A customer service assistant must always escalate potential fraud accusations before taking adverse action. Which mechanism is best?
- Hide the source evidence from the reviewer to avoid bias
- A mandatory review gate triggered by explicit risk criteria
- Allow the assistant to bypass review for familiar customers
- Treat the model's fraud label as an automatic ban
Explanation
Correct Answer
B:A human checkpoint can assess evidence and consequences before account action. The agent can gather facts but not make irreversible accusations alone.
Incorrect Answers
A:Reviewers need relevant facts to assess the claim. Data should be scoped and presented with provenance.
C:Familiarity is not a documented exception and can create inconsistent treatment. The gate should apply according to policy.
D:An unverified classifier can produce false positives. Human review is the stated requirement.