Claude Certified Architect Foundations Practice Test – 360 Free Exam Questions with Answers

Correct answer: C. Document team testing standards, high-value test criteria, and reusable fixtures in project CLAUDE.md consumed by the CI job.

Project-level context files like CLAUDE.md are the correct place to define team-wide testing standards and available fixtures, ensuring consistent CI guidance. Relying on individual developer prompts or increasing test volume does not solve the underlying context gap and will still produce superficial or duplicate tests.

Correct answer: B. Run the review in a separate Claude Code invocation with the PR diff, repository context, and review criteria only.

Running the review in a separate Claude Code invocation provides a clean perspective focused solely on the diff and repository criteria, avoiding confirmation bias. Reusing the generation session makes the model defend its prior reasoning, which blinds it to subtle architectural flaws.

Correct answer: D. Share the failing test names, inputs, expected versus actual outputs, and stack traces, then iterate on minimal fixes.

Sharing specific failing test names, inputs, and expected versus actual outputs turns vague feedback into a concrete, test-driven loop. Subjective prompts like asking the model to be more careful do not provide actionable targets, leading to inconsistent fixes and unrelated code rewrites.

Correct answer: C. Create tests for expected behavior, edge cases, and performance constraints first, then iterate by sending Claude failing results.

Defining tests for expected behavior and edge cases before implementation creates objective acceptance criteria that catch subtle bugs early. Feeding specific failing test outputs back to the model ensures precise, incremental corrections instead of the uneven results caused by long prose specifications.

Correct answer: A. Send one detailed refinement request describing all observed failures, shared constraints, and expected behavior across the affected call paths.

When interacting failures share constraints across the same template, submitting one detailed refinement request gives the model the full context needed to avoid regressions. Fixing issues sequentially in separate sessions blinds the model to dependencies, often reintroducing previously resolved bugs.

Correct answer: C. Have the coordinator maintain shared investigation state and include relevant prior findings in each subsequent subagent prompt.

The correct answer makes the coordinator responsible for passing shared context into each subagent prompt. For the exam, remember that subagents do not automatically inherit prior conversational state, so the orchestrator must explicitly provide it.

Correct answer: C. Call escalate_to_human with the order context and brief summary, without attempting autonomous lookup or replacement processing first.

The correct choice honors an explicit human request by transferring immediately with context. Avoid the tempting autonomous lookup option, because forcing automation when a user explicitly requests a person breaks trust and violates core escalation policy design principles.

Correct answer: C. Rewrite the QA prompt to define reportable errors, acceptable variations, and skip conditions with concrete examples for each category.

Defining explicit reportable errors with concrete examples directly fixes high false positive rates. Vague instructions to be conservative fail because the model lacks objective boundaries for acceptable variations versus actual extraction mistakes.

Correct answer: B. Add a coordinator review loop that checks synthesis coverage, redelegates targeted follow-ups, and reruns synthesis until quality criteria are met.

Adding a coordinator review loop actively identifies gaps and triggers targeted follow-ups before report generation. Fixed iteration caps fail because they stop blindly regardless of remaining coverage gaps, while hiding low-coverage sections obscures rather than resolves the issue.

Correct answer: B. Move the required catalog server into project-scoped .mcp.json, and keep the experimental notes server in user-scoped ~/.claude.json.

Shared team servers belong in project-scoped configuration so they sync automatically via source control. User-scoped configuration isolates experimental tools, whereas manual setup commands or user-scoped files invite drift and configuration inconsistency across the team.

Correct answer: C. Return isError true with errorCategory business, isRetryable false, and guidance to request summaries or short excerpts instead.

Returning a non-retryable business error with actionable guidance tells the coordinator to change its approach. Treating this as transient or a credential failure wastes tokens on retries since the license restriction is a permanent business rule, not a technical fault.

Correct answer: C. Move the shared conventions into a project-level CLAUDE.md file and commit it so every clone loads them.

Committing a project-level configuration file guarantees that all cloned repositories receive the shared conventions automatically. Relying on manual copying or user memory files breaks consistency, while slash commands require manual invocation that engineers might forget.

Correct answer: A. Have Claude ask targeted clarification questions about operational constraints, ownership assumptions, and edge cases before proposing implementation changes.

Having the agent ask targeted clarification questions resolves hidden requirements before any code is written. Relying on code review to catch missing constraints is inefficient, and adding global instructions creates rigid rules that conflict with team-specific variations.

Correct answer: A. Randomly sample high-confidence claims across topic, source type, and report section, then track errors and new patterns by stratum.

Stratified random sampling exposes hidden error rates in specific topics or source types that aggregate metrics hide. Reviewing only low-confidence claims is risky because overconfidence in niche segments can let systematic defects reach production undetected by overall dashboards.

Correct answer: C. Preserve both statistics with source attribution, methodology notes, and a conflict annotation for the report generator to present transparently.

Preserving both statistics with methodology notes allows transparent reporting instead of forcing a false consensus. Averaging conflicting metrics is mathematically invalid when underlying definitions differ, and suppressing the smaller study hides critical context from readers.

Correct answer: A. Replace fetch_url with a load_document tool that accepts catalog document IDs or approved URLs and validates before fetching.

Replacing the generic tool with a constrained one directly prevents the subagent from fetching arbitrary URLs. Relying on prompt instructions to avoid links is a weaker safeguard because the model may still attempt to retrieve unapproved resources.

Correct answer: D. Require each subagent to return structured handoff records separating findings, file paths, symbols, line ranges, and evidence snippets.

Structured handoff records preserve attribution by separating findings from the metadata needed to verify them. Asking the synthesizer to reconstruct missing file paths from prose is unreliable because lost context cannot be recovered deterministically.

Correct answer: C. Use Grep to search file contents for the error string and function names, then Read the matching files.

Using Grep to search file contents is the correct first step because it efficiently targets the specific error string and function names. Reading files upfront or using Glob wastes context and ignores the actual file contents.

Correct answer: B. Use a fixed prompt chain for stable review rubrics, and let the agent adapt subtasks during open-ended investigations.

Stable workflows benefit from fixed prompt chains, while open-ended investigations require adaptive task planning. Using a single rigid checklist for both is an anti-pattern because bug discovery dictates subsequent steps.

Correct answer: D. Have Claude first ask targeted questions about invariants, invalidation triggers, audit requirements, and failure modes before proposing implementation.

Using the interview pattern surfaces hidden constraints and invalidation rules before implementation begins. Immediately writing code is a high-risk anti-pattern because Claude might fill compliance gaps with incorrect defaults.

Correct answer: A. Send a follow-up request containing the document, failed extraction, and exact validation errors for targeted self-correction.

Providing the exact validation errors alongside the failed extraction gives Claude the context needed for targeted self-correction. Blindly retrying the original prompt or lowering temperature fails because the model lacks the specific feedback required to fix semantic mistakes.

Correct answer: D. Run Claude Code with –output-format json and –json-schema, then map validated findings to inline PR comments.

Using explicit JSON output formats with a schema forces Claude to produce reliable, machine-readable data. Relying on regex parsing over markdown is a brittle anti-pattern because natural language formatting varies unpredictably across runs.

Correct answer: B. Add explicit human-review criteria with few-shot examples distinguishing absent optional fields, unreadable required fields, and unsupported document types.

Adding explicit human-review criteria with few-shot examples clarifies the decision boundaries between absent optional fields and unreadable required values. Global confidence thresholds fail because they cannot distinguish between structurally missing data and extraction errors.

Correct answer: C. Run focused document-level evidence checks first, then perform a separate cross-document consistency pass before synthesis and reporting.

Splitting the workflow into focused document-level checks followed by a cross-document consistency pass prevents attention dilution. Processing all sources in a single prompt causes lost-in-the-middle effects, leading to uneven validation quality.

Correct answer: B. Keep synchronous calls for onboarding forms, and use the Message Batches API for historical contracts with latency-tolerant processing.

Matching the API to the latency requirement means keeping synchronous calls for real-time onboarding screens. The Message Batches API is perfect for historical contracts because it reduces costs for asynchronous, latency-tolerant workloads.

Correct answer: B. Split the guidance into focused topic files under .claude/rules/, such as testing.md, api-conventions.md, and deployment.md.

Organizing topic-specific guidance in .claude/rules/ preserves shared repository-level configuration while avoiding a single oversized file. This directly addresses maintainability and review concerns without requiring developers to manually invoke commands or copy personal settings.

Correct answer: B. Revise coordinator planning to identify plausible entry points, then delegate distinct code areas to subagents before synthesis.

The symptoms show that subagents completed their assigned work, but the assignments covered only the HTTP path. Improving the coordinator's initial planning helps ensure broad codebase questions are decomposed across relevant entry points before synthesis begins.

Correct answer: C. Define an extraction tool with a JSON schema for target fields, then consume the structured tool_use input directly.

Tool use with a JSON schema is the most reliable way to obtain schema-compliant structured output from Claude. Instead of parsing generated text, the application reads the structured input from the tool_use block, eliminating common JSON syntax failures like markdown fences or trailing commentary.

Correct answer: C. Return failed tool calls with isError set true, plus concise diagnostic content the subagent can use for recovery.

MCP tools should distinguish failed executions from successful content using the isError pattern. This lets the agent treat the response as a tool failure rather than as research material, enabling retries, alternative searches, or coverage-gap annotations.

Correct answer: B. Use Glob with patterns like <code>**/*.fixture.json</code> to enumerate matching paths, then Read only the files that require inspection.

Glob is the appropriate built-in tool when the selection criterion is based on file names, extensions, or path patterns. It returns matching paths directly, reducing noise and allowing the agent to read only the relevant files.

Correct answer: D. Add targeted few-shot examples showing the fields extracted from cover-letter text, footnotes, and table captions with brief rationale.

Targeted few-shot examples are the best fit when extraction failures cluster around varied document structures. Showing examples where fields appear in cover letters or footnotes teaches Claude the intended judgment pattern, whereas forcing non-null values would encourage hallucinated data.

Correct answer: C. Run focused review passes on each packet section, then run a separate consistency pass across cross-section fields.

Running focused review passes on each packet section reduces attention dilution and improves coverage of local field errors. A separate consistency pass then focuses on relationships across sections, such as matching tax IDs and signatures, without competing with local checks.

Correct answer: B. Start a fresh session with a structured summary of still-valid findings, then re-run targeted searches and analyses for changed sources.

Starting fresh avoids carrying stale tool outputs and outdated subagent observations forward as if they were still authoritative. Injecting a structured summary preserves useful context while forcing updated evidence to be gathered where source validity has changed.

Correct answer: C. Expose the support policy index and knowledge-base catalog as MCP resources, while keeping tools for selected record actions.

MCP resources are the correct tool for exposing read-only reference catalogs, giving the agent immediate context without wasting turns on exploratory tool calls. A practical exam cue is to match static data to resources and dynamic state changes to tools, avoiding bloated system prompts or mandatory startup calls.

Correct answer: C. Use plan mode to explore dependencies, compare viable migration strategies, and produce an implementation plan before modifying files.

Using plan mode allows Claude to explore dependencies and design an implementation strategy before modifying files, which is critical for broad architectural changes. Jumping straight to execution risks breaking unrelated workflows because the model lacks a structured blueprint for complex multi-file migrations.

Correct answer: A. Rewrite each MCP tool description to specify purpose, accepted identifiers, returned fields, edge cases, and boundaries versus related tools.

Rewriting MCP tool descriptions to clarify their purpose, accepted inputs, and boundaries is the most direct fix for unreliable model-driven tool selection. Relying on hardcoded keyword routers or merging tools obscures functionality and defeats the purpose of granular, descriptive tool architecture.

Correct answer: D. Force the metadata extraction tool with tool_choice for the first request, then continue subsequent analysis in follow-up turns.

Forcing a specific tool call using tool choice is the most reliable way to guarantee a required initial step executes before subsequent model-driven analysis. Prompting alone is probabilistic, and setting tool choice to any does not ensure the correct prerequisite tool is selected.

Correct answer: A. Treat multiple customer matches as unresolved identity and request a disambiguating identifier before any order or refund operations.

Treating multiple customer matches as an unresolved identity forces the agent to request a disambiguating identifier before taking sensitive actions, safely preventing account mixups. Routing all duplicate records to humans breaks automation goals, while probabilistic heuristics remain too risky for refunds.

Correct answer: B. Add a structured detected_pattern field to each QA finding, then analyze acceptance rates by pattern and document type.

Adding a structured detected pattern field captures the specific trigger for each finding, making the feedback loop analyzable. Segmenting outcomes by pattern and document type reveals exactly which prompts or validators need refinement, whereas globally raising confidence thresholds suppresses useful categories indiscriminately.

Correct answer: D. Require summaries and handoffs to carry structured claim-source mappings with source type, excerpt or tool field, timestamp, and verification status.

Structured claim-source mappings preserve provenance as information moves through summarization and escalation, ensuring facts are distinguished from inferences. Relying on shorter narratives or confidence filtering increases provenance loss and hides uncertainty, making it an anti-pattern for multi-source workflows.

Correct answer: C. Require subagents to return structured handoff records with findings separated from file paths, symbols, line ranges, commands, and source excerpts.

Structured handoff records preserve both discovered content and the metadata needed to verify it later. Separating facts from attribution metadata helps subagents use findings accurately without losing file locations, whereas stripping source details or relying on narrative summaries degrades downstream reliability.

Correct answer: A. Split the request into distinct concern threads, investigate them in parallel with shared constraints, then synthesize one unified cross-sector report.

Splitting the request into distinct concern threads ensures balanced coverage while keeping shared constraints consistent across the investigation. A final unified synthesis step resolves contradictions before report generation, avoiding the attention dilution and conflicting recommendations caused by single-pass sequential processing.

Correct answer: D. Update each subagent's output contract to attach dates, source locations, and methodological notes to every extracted finding.

Updating output contracts to attach dates, source locations, and methodological notes preserves critical context at the finding level. This prevents important qualifiers from being lost during handoffs, whereas passing full transcripts risks attention problems and cleanup passes might hallucinate missing provenance.

Correct answer: C. Add a prerequisite gate that rejects process_refund until verified customer, eligible order, and computed refund amount are present in state.

Adding a prerequisite gate creates a deterministic control point for workflow ordering, preventing financially risky actions until state requirements are met. Relying on system prompts or few-shot examples remains probabilistic, whereas a hardcoded gate guarantees compliance regardless of model reasoning.

Correct answer: D. Analyze accuracy by document type and extracted field, then automate only segments meeting validated thresholds while routing others to review.

Analyzing accuracy by document type and field exposes weak performance clusters that strong aggregate scores hide. Segment-level validation lets you safely automate reliable areas while preserving human review for underperforming slices, avoiding the false confidence of relying strictly on global metrics.

Correct answer: B. Return structured MCP errors with categories, retryability, and user-safe details so the agent can choose recovery, clarification, or escalation.

Returning structured MCP errors with categories and retryability gives the agent the exact context needed to choose appropriate local recovery steps. Uniform failure text forces the model to guess, causing unnecessary escalations and frustrating retries on non-retryable failures.

Correct answer: C. Use plan mode to map relevant flows and compare implementation points before allowing file modifications.

Using plan mode lets the agent map complex flows and identify the correct abstraction layer before modifying files. Proceeding with direct execution relies on trial and error, which risks breaking shared dependencies and causes avoidable rollback work on legacy systems.

Correct answer: B. Run a representative pilot batch, analyze validation failures and null patterns, refine the prompt, then submit the full workload.

Running a representative pilot batch exposes real formatting variations and failure modes before they impact the full workload. Refining the prompt based on pilot validation results maximizes first-pass success, whereas bulk submission or increasing max_tokens ignores underlying template issues.

Correct answer: B. Send a follow-up request containing the source document, failed extraction, and precise validator messages for targeted correction.

Targeted correction feeds specific validator feedback back to the model, directly addressing the defect. Loosening the schema weakens validation, while blind retries lack the context needed to fix semantic errors.

Correct answer: D. Add a PostToolUse hook that converts each tool result into canonical dates, status labels, and tier fields before Claude sees it.

A PostToolUse hook deterministically normalizes tool outputs before the model reasons over them. Relying on prompt instructions to translate formats is fragile and risks hallucinations during complex support cases.

Correct answer: B. Analyze each source in a focused pass with structured findings, then run a separate cross-source integration pass before reporting.

Separating local analysis from cross-source synthesis uses prompt chaining to reduce attention dilution. A dedicated integration pass resolves contradictions and preserves caveats before report generation begins.

Correct answer: A. Run a separate Claude review instance with the source document, extracted fields, and rubric, excluding the original extraction reasoning.

A separate review instance prevents the model from inheriting assumptions from the original extraction rationale. Self-critique in the same conversation often reinforces prior reasoning instead of objectively catching subtle placement errors.

Correct answer: B. Add targeted examples showing complete actionable findings with document location, affected field, issue description, severity, and suggested correction.

Targeted few-shot examples teach the exact structure and specificity required for actionable reviewer notes. General instructions alone fail because they lack the concrete formatting needed to ensure consistent downstream ticketing.

Correct answer: C. Have subagents report unresolved failures with attempted section, failure type, partial fields, and recovery suggestions for coordinator handling.

Reporting failures with recovery suggestions lets the coordinator handle errors without discarding successful partial work. Returning empty objects is an anti-pattern because it falsely masks unresolved failures as successful extractions.

Correct answer: B. Add a detected_pattern field to each structured finding, then analyze dismissal rates by pattern during feedback review.

A structured pattern field preserves the specific source construct or extraction trigger behind each finding. When reviewers dismiss findings, the team can aggregate by pattern and identify which prompts, schemas, or validations need improvement.

Correct answer: D. Resume the session, identify the changed policy and tool files, and request targeted re-analysis before continuing implementation decisions.

Resuming the session preserves useful investigation context while explicitly invalidating parts likely to be stale. Targeted re-analysis lets Claude refresh its understanding of the changed policy and tool schema before continuing implementation decisions.

Correct answer: D. Assign each subagent a distinct issue boundary and shared case facts, then synthesize their non-overlapping findings centrally.

Assigning distinct boundaries and shared facts leverages parallel specialization without redundancy. On the exam, remember that uncontrolled subagent messaging or broad prompts cause overlapping loops and conflicting recommendations.

Correct answer: A. Return MCP tool errors with isError plus category, retryability, and safe messages distinguishing transient, validation, business, and permission failures.

Categorized tool errors give the agent signals to choose appropriate recovery paths like retrying timeouts or clarifying invalid inputs. Generic error messages fail because they force the model to guess recovery steps blindly.

Correct answer: A. Include verified customer identity, relevant order records, prior findings, and exact investigation goal in each subagent prompt.

Passing verified identity, relevant data, and the exact goal directly in the prompt gives subagents the context they need to act accurately. Subagents do not automatically inherit the parent conversation history or prior tool results.

Correct answer: A. Add a coordinator-side prerequisite that blocks report-generation Task calls until synthesis returns a structured citation manifest and success status.

Enforcing a deterministic prerequisite ensures the report agent cannot start until synthesis provides a structured manifest. Relying on the report agent to guess if synthesis completed creates fragile workflows and risks generating incomplete reports.

Correct answer: C. Run an agentic loop that inspects stop_reason, executes requested tools, appends tool results, and repeats until end_turn.

Implementing the core agentic loop lets Claude inspect the stop reason, execute requested tools, and append results until the turn naturally ends. Templatized responses fail because they bypass the model's reasoning capabilities over newly acquired tool context.

Correct answer: A. Retry with the source document, rejected extraction, and exact validator messages, instructing Claude to produce a corrected record only.

Providing the source document, the rejected extraction, and exact validator messages supports targeted self-correction. Asking the model to fix errors without the original text forces it to guess the correct values, leading to plausible but unsupported hallucinations.

Correct answer: B. Have subagents report successful empty searches separately from access failures, including attempted paths, queries, partial results, and suggested next steps.

The correct answer ensures subagents explicitly differentiate between a successful search with no matches and a failed search due to access errors. A practical exam cue is to select the option that maximizes error visibility and context recovery, avoiding silent failures where missing data mimics clean results.

Correct answer: C. Pass the relevant current test files into the review context and ask Claude to target only uncovered behaviors.

Providing existing test files gives Claude the necessary context to identify actual coverage gaps instead of suggesting redundant tests. Adding a prompt rule without context fails because Claude cannot know existing coverage without seeing the tests.

Correct answer: C. Return error details that include failure type, attempted order lookup, partial customer data, cached summaries, and suggested fallback actions.

Returning detailed error context enables the coordinator to distinguish isolated failures from total outages, allowing safe usage of cached data. Converting timeouts to empty results is a strong distractor because it masks actual errors and risks processing incomplete backend states.

Correct answer: C. Commit project-level CLAUDE.md guidance describing valuable test criteria, fixture conventions, and reusable helpers for CI-invoked Claude Code.

Committing project-level guidance ensures Claude Code consistently loads team preferences from ephemeral CI runners. Local configurations are a strong distractor because ephemeral environments lack access to maintainer laptops, making prompt-only adjustments ineffective for persistent context.

Correct answer: D. Rewrite the MCP tool descriptions to detail support-specific capabilities, accepted identifiers, returned fields, and when to use them instead of generic search.

Rewriting MCP tool descriptions gives Claude the precise context needed to prefer specialized tools over generic file searches. Disabling generic file search is a strong distractor because it removes useful agent flexibility instead of addressing the underlying tool selection signal.

Correct answer: C. Use plan mode to investigate dependencies and choose the migration path, then switch to direct execution for the agreed edits.

The correct answer uses plan mode to map dependencies safely before switching to direct execution for mechanical edits. When tackling workflow questions, remember that staying in plan mode for mechanical execution wastes time, while direct execution risks costly rework during discovery.

Correct answer: B. Route all subagent requests and results through the coordinator, which records context, errors, and final resolution inputs.

Routing communication through the coordinator preserves specialization while giving one component responsibility for observability, error handling, and final aggregation. This directly addresses the missing retry rationale and bypassed escalation notes because all handoffs become visible to the coordinator.

Correct answer: B. Have the coordinator assign each subagent a distinct concern and shared verified facts, then combine findings into one customer response.

Having the coordinator assign distinct concerns and shared facts keeps subagents focused on their specialization without redundant tool calls. Allowing direct messaging bypasses centralized observability, which often leads to missed issues and overlapping summaries during complex investigations.

Correct answer: D. Assign a stable custom_id derived from each document record, persist it with request metadata, and reconcile results using that identifier.

Message Batches support request and response correlation through custom_id values. Persisting a stable identifier with the submitted request lets the ingestion job map results correctly, whereas relying on output ordering fails when responses are processed after retries or partial failures.

Correct answer: B. Drive the loop from stop_reason, execute requested tools when it is tool_use, and return only when it is end_turn.

The agentic loop should be driven by stop_reason, executing tools when it equals tool_use and returning when end_turn occurs. Stopping the loop based on natural language text is fragile because Claude often outputs text alongside tool calls before actually finishing.

Correct answer: A. Move the accessibility guidance into a .claude/rules/ file with YAML frontmatter paths matching React component globs.

Path-specific rule files allow Claude Code to load guidance only when file paths match the declared glob patterns. Using generic root instructions or manual commands fails because it relies on the model correctly guessing file contexts rather than applying deterministic system rules.

Correct answer: A. Route every subagent handoff and result through the coordinator, which aggregates state, applies escalation rules, and delegates follow-up work.

A coordinator-controlled hub-and-spoke pattern preserves observability, consistent error handling, and controlled information flow across specialized subagents. Letting subagents communicate directly bypasses central escalation rules and creates blind spots in the execution trace.

Correct answer: D. Populate escalate_to_human with a structured handoff including customer ID, verified facts, root cause, amounts, attempted actions, and recommended next steps.

Structured escalation handoffs are essential when the human reviewer lacks access to the original conversation or tool outputs. The agent should send a compact, explicit payload including identifiers, verified facts, and recommended next steps. Relying on transcript links fails when reviewers lack system access.

Correct answer: A. Rewrite delegated prompts to state the support objective, known context, quality criteria, constraints, and required structured return format.

Subagents perform better when they understand the objective, constraints, and required return format rather than following brittle procedural scripts. Expanding step-by-step branches fails because it cannot anticipate every edge case, whereas goal-oriented prompts allow dynamic tool adaptation.

Correct answer: C. Grant the refund-policy subagent scoped access to lookup_order, while keeping process_refund and escalation routed through the coordinator.

Granting scoped read-only tool access to specialized subagents reduces latency without compromising safety or central oversight. Giving the subagent state-changing tools would violate least privilege principles and bypass the coordinator's ability to handle sensitive actions reliably.

Correct answer: C. Add enum values for ambiguous cases and a fallback category paired with a required free-text detail field.

Adding fallback enum values alongside a required free-text detail field preserves stable machine-readable categories without forcing inaccurate classifications. Unrestricted strings break downstream analytics, while closed enums force inaccurate predictions when handling unexpected inputs.

Correct answer: C. Detect source absence after validation, stop automated retries, mark extraction incomplete, and request the missing approval letter.

Retries cannot recover facts that are absent from the provided source context. Marking the extraction incomplete and requesting the missing approval letter prevents unsupported hallucinations from entering downstream systems, which is a core architectural principle.

Correct answer: D. Break down validation results by artifact category and generated field, then reduce review only for consistently reliable segments.

Aggregate acceptance rates can mask weak performance in specific artifact categories or fields. Segmenting validation results by artifact type allows you to reduce human review safely only where consistency is proven, rather than trusting uncalibrated global confidence scores.

Correct answer: A. Require each extracted obligation to carry source identifiers and excerpts, then merge duplicates while retaining every supporting reference.

The correct answer preserves data provenance by attaching source identifiers and excerpts to each extracted obligation, enabling safe consolidation. A strong exam cue is to always look for the option that maintains verifiable audit trails while allowing downstream data merging. Batch-level metadata lacks granular tracing.

Correct answer: B. Rewrite the review prompt to report demonstrable bugs, security risks, and test gaps, while skipping stylistic preferences without documented impact.

The correct answer aligns the prompt with objective issues like bugs and security risks, successfully filtering out subjective noise. When tackling CI integration questions, always prioritize options that narrow the evaluation scope to actionable, demonstrable problems over mere threshold adjustments.

Correct answer: A. Create a repository rule file with YAML path globs matching component and story files, then remove duplicated directory instructions.

The correct answer uses path-scoped repository rules to apply conventions dynamically based on file globs, eliminating drift from duplicated files. Root configurations require manual slicing, whereas custom commands fail if developers forget to invoke them during active development.

Correct answer: D. Have Claude maintain a repository scratchpad file recording discovered schemas, validators, edge cases, and file paths for future reference.

The correct answer uses a persistent scratchpad file to retain key architectural discoveries outside the volatile session context. When evaluating context degradation fixes, avoid relying on conversation compression tools, as these often omit critical implementation details needed for later success.

Correct answer: A. Use Glob with targeted filename patterns to list candidate policy and refund test files, then Read only the relevant matches.

Glob is the appropriate built-in tool when the selection criterion is a file path or filename pattern rather than file contents. Listing candidate paths first keeps context focused and avoids loading unrelated files before Claude knows which files matter.

Correct answer: A. Resume the named session, identify the changed policy and tool-schema files, and request targeted revalidation before implementing fixes.

Resuming the session while revalidating changed files preserves useful investigation context and explicitly corrects stale assumptions. Asking for a single confidence score is a fragile cue because it lacks deterministic file verification, which might lead to implementing fixes based on outdated schemas.

Correct answer: D. Add a PostToolUse hook that transforms lookup_order outputs into a canonical support-case structure before Claude receives them.

A PostToolUse hook intercepts tool results and transforms them into a canonical structure before the model reasons over them. Expanding the system prompt with regional mappings relies on probabilistic model compliance and increases context burden, whereas the hook centralizes normalization deterministically.

Correct answer: A. Temporarily suppress the ambiguous payment terms category while refining its criteria and examples before re-enabling it.

Temporarily suppressing the noisy category targets the specific issue damaging reviewer trust without weakening accurate finding categories. Raising global confidence thresholds or adding broad conservative instructions reduces the visibility of accurate findings, worsening overall system reliability.

Correct answer: A. Expose the read-only policy, reason-code, and queue catalogs as MCP resources, while keeping customer-changing operations as tools.

MCP resources are appropriate for read-only catalogs that provide context about available data or policy options. This reduces exploratory tool calls while preserving tools for model-driven actions such as looking up orders, processing refunds, or escalating cases.

Correct answer: D. Add Task to coordinator allowedTools and invoke specialists through it, while keeping backend MCP tools scoped to appropriate agents.

In the Agent SDK, a coordinator invokes subagents through the Task tool, so allowedTools must include Task. Prompting the coordinator to mention specialist names is only natural-language guidance, not a real subagent invocation, and granting broad tool permissions breaks least privilege.

Correct answer: B. Create a focused rule file with YAML frontmatter paths for both migration globs, containing only the migration review checklist.

Using a focused rule file with YAML frontmatter path globs ensures instructions load only for matching files. Path-scoped rules prevent irrelevant warnings in TypeScript reviews while keeping automated guidance active for SQL migrations.

Correct answer: C. Add a tool-call gate that rejects process_refund until get_customer and lookup_order return matching, refund-eligible records.

A programmatic tool-call gate deterministically enforces the required validation sequence before refunds execute. Relying solely on system prompt instructions is risky for business-critical safety, whereas a gate guarantees eligibility checks without blocking valid cases.

Correct answer: C. Store the command markdown under .claude/commands/ in the repository, then commit it with the CI workflow changes.

Storing custom commands inside the project's dot-Claude directory ensures they are version-controlled and available from a fresh checkout. Keeping them in a user's home directory prevents standard CI runners and new contributors from accessing the workflow.

Correct answer: D. Before launching each new investigation wave, create a structured findings brief and include it in every subagent's initial prompt.

Providing a structured findings brief in each subagent prompt explicitly passes accumulated context across isolated boundaries. Relying on subagents to rediscover earlier constraints wastes effort and risks inconsistent conclusions, so explicit context passing is essential.

Correct answer: C. Revise each tool description to state accepted identifiers, example customer phrases, edge cases, and boundaries versus similar tools.

Clear tool descriptions provide the necessary context for the model to select the right tools during ambiguous conversations. Hard-coded decision trees or keyword routers are brittle, whereas detailed descriptions help the model distinguish overlapping features naturally.

Correct answer: C. Use fork_session from the verified case baseline, explore each resolution path independently, then compare branch summaries before deciding.

Forking a verified session preserves the shared factual baseline while exploring divergent reasoning paths in isolated contexts. Exploring multiple paths in a single conversation causes assumptions to bleed between branches, leading to flawed comparisons.

Correct answer: C. Return distinct structured categories for transient, validation, business, and permission failures, with retryability flags and customer-safe recovery guidance.

Returning distinct structured categories for each failure type gives the model actionable context to choose the right recovery path. Collapsing all errors into a single result prevents the agent from distinguishing transient timeouts from hard policy denials.

Correct answer: B. Keep tool-based structured output, then add semantic validation against diff metadata, severity rules, and finding consistency before publishing comments.

Semantic validation against pull request diff metadata ensures that model outputs meet actual review policies before publishing. Structured JSON schemas only enforce formatting compliance, so logic checks are still needed to catch invalid line numbers or logical contradictions.

Correct answer: B. Have the coordinator compare the draft resolution with the original request, assign targeted follow-up subagent work, then rerun synthesis.

An explicit coordinator review step compares draft resolutions against original requests to identify incomplete answers. This allows targeted follow-up queries to subagents, ensuring completeness without unnecessarily escalating complex cases to humans.

Correct answer: A. Separate findings by category and severity, tighten criteria for noisy checks, and prioritize high-precision issues in reviewer queues.

Separating findings by category and severity directly fixes reviewer fatigue by preventing noisy alerts from burying high-precision flags. Tightening criteria for ambiguous categories ensures reviewers maintain trust and focus on critical confirmed mismatches.

Correct answer: B. Require confidence per critical record field, then calibrate routing thresholds using a labeled set of previously reviewed changes.

Field-level confidence scores expose hidden uncertainty in high-risk areas like rollback and security. Calibrating routing thresholds against labeled review data prevents a high overall score from masking critical blind spots, whereas a single aggregate score lacks the needed precision.

Correct answer: B. Create a .claude/rules/ migration file with paths matching all migration patterns so guidance loads only for those files.

Path-scoped rule files are ideal for enforcing conventions by file pattern rather than directory hierarchy. This setup keeps migration guidance automatic for matching files across the monorepo while preventing irrelevant context from loading during unrelated edits.

Correct answer: C. Use a prompt chain that reviews each order or charge independently, then runs a final account-level consistency pass.

A prompt chain matches this predictable workflow by isolating each local item before running a final account-level synthesis. Separating per-item review from broader reconciliation reduces attention dilution and catches interactions across issues before responding.

Correct answer: C. Use small package-level CLAUDE.md files that reference the relevant standards documents with @import, keeping shared root guidance minimal.

Using small package-level files with @import keeps package guidance close to the code while avoiding copied instruction drift. This allows maintainers to load only relevant content, reducing context bloat and ensuring standards remain canonical and updatable.

Correct answer: B. Configure tool_choice to force extract_pr_metadata on the first request, then run review findings in a follow-up turn.

Forcing a specific tool choice is the safest mechanism when a structured extraction must happen before later steps. It guarantees the first response calls the metadata tool, allowing the pipeline to use those results in a follow-up request for generating findings.

Correct answer: D. Return successful empty results only for completed searches, and structured retryable errors when order lookup cannot access the backend.

Separating valid empty results from access failures provides the necessary state for correct control flow. Distinguishing between a completed search with zero matches and a backend outage prevents false denials and enables proper retry logic or escalation.

Correct answer: B. Have agents maintain concise scratchpad files with key findings, file paths, and decisions, then reference them in later prompts.

Persistent scratchpad files preserve important discoveries outside the model's shifting conversation context. Referencing these files in later prompts helps counteract context degradation and keeps the investigation grounded in previously verified facts.

Correct answer: C. Use Glob with **/*.policy-case.json, review the matched paths, then Read only the relevant fixture files.

Using Glob is the most direct approach when the selection criterion is a filename pattern rather than file contents. Reviewing matched paths before reading keeps the context focused and avoids unnecessary token usage from loading irrelevant files.

Correct answer: B. Move specialized guidance into .claude/rules/ files with YAML paths globs so only rules relevant to changed files load.

Path-scoped rule files load conditionally based on matching file paths, keeping irrelevant guidance out of the active context. This directly reduces token usage and false positives while preserving automatic rule application during continuous integration review.

Correct answer: D. Run the review as a separate Claude Code invocation using only the diff, repository context, and review criteria.

A separate review invocation prevents the model from carrying forward implementation reasoning that might make it less likely to question its own choices. Providing the diff and repository context preserves necessary information without inheriting generator assumptions.

Correct answer: D. Create a project skill in .claude/skills/test-gap-review/SKILL.md, configuring argument-hint and allowed-tools frontmatter, then include the reusable checklist.

A project skill packages reusable task-specific workflows directly in the repository. Frontmatter elements like argument-hint and allowed-tools control execution behavior, making this approach superior to slash commands or relying on generic CLAUDE.md rules for constrained tool access.

Correct answer: D. Add a preprocessing layer that stores full tool outputs externally and appends only schema-relevant fields and provenance to conversation history.

Storing full tool outputs externally and appending only schema-relevant fields to the context prevents verbose data from degrading model performance. This approach solves context bloat and lost-in-the-middle effects, whereas simply increasing max tokens or using prose summaries fails to fix the architectural issue.

Correct answer: C. Rewrite the affected tool descriptions to specify purposes, required inputs, side effects, examples, and boundaries between refund and escalation workflows.

Rewriting tool descriptions to specify boundaries, required inputs, and side effects directly addresses inconsistent tool selection. Claude relies heavily on these descriptions to differentiate between similar actions, making explicit boundaries far more reliable than adding brittle keyword routers or merging distinct workflows.

Correct answer: A. Configure tool_choice to any so Claude must call one available tool while retaining model-driven selection among applicable tools.

Configuring tool_choice to any guarantees the agent will execute a backend action rather than returning generic text. This setting enforces tool usage while still allowing the model to select the most contextually appropriate tool, avoiding the brittleness of forcing a specific hardcoded tool.

Correct answer: A. Keep tool_use schema enforcement, add deterministic checks for line ranges and file membership, then retry with those failures included.

Adding deterministic checks for file membership and line ranges catches semantic inconsistencies that JSON schema validation alone misses. Including these specific failures in a targeted retry provides actionable feedback to the model, which is much more effective than running blind retries or tightening schema enums.

Correct answer: C. Update each MCP tool to return categorized failures with retryability and customer-safe guidance tailored to the specific backend condition.

Updating tools to return categorized failures with retryability and customer-safe guidance gives the agent the context needed to choose correct recovery paths. This allows the agent to distinguish between temporary timeouts and permanent policy denials, which generic error messages or hard-coded retry loops cannot do.

Correct answer: D. Run Claude Code with –output-format json and –json-schema defining required fields for each pull request finding.

Using structured JSON output with an explicit schema gives the CI script a stable contract for parsing findings. Relying on regex to parse markdown is a brittle anti-pattern, whereas structured JSON directly enforces required fields like severity and line ranges for reliable automation.

Correct answer: D. Have the coordinator assess request scope, delegate only to relevant specialist subagents, then aggregate their findings into one response.

Having the coordinator assess scope and delegate only to relevant specialists eliminates unnecessary work for simple requests. This pattern preserves multi-specialist aggregation for complex cases while avoiding the fragility of keyword routing and the anti-pattern of subagents messaging each other directly.

Correct answer: C. Populate each Task prompt with the verified customer facts, relevant order records, prior findings, and needed decision criteria.

Subagents operate with isolated context and cannot automatically view the coordinator's transcript. Populating the Task prompt with verified facts and decision criteria ensures the subagent has the necessary information, preventing duplicate tool calls and redundant questions to the customer.

Correct answer: A. Replace vague confidence guidance with concrete reportable and non-reportable categories, severity definitions, and code examples for each finding type.

Replacing vague confidence guidance with concrete categories and severity definitions gives the model operational boundaries for deciding what to report. A 95% certainty threshold is too restrictive and eliminates useful findings, whereas specific examples calibrate the reviewer effectively.

Correct answer: B. Append each executed tool result to the conversation history using the matching tool_use_id before sending the next model request.

Claude can only reason over information explicitly included in the messages sent on the next request. Appending each tool result with its matching tool use ID bridges the context gap, whereas keeping data solely in application state leaves the model blind.

Correct answer: B. Define the shared server in the repository's .mcp.json, using environment variable placeholders for credentials rather than committed secrets.

Defining the server in the repository .mcp.json file makes shared tooling travel with the codebase. Using environment variable placeholders allows secure local credential injection, whereas documenting manual local setup steps breaks reproducibility for the team.

Correct answer: A. Update guidance to acknowledge frustration, proceed with supported resolution steps, and escalate only if the customer again asks for a human.

Frustration alone is not a reliable signal that a case requires human escalation. When the issue is within the agent's capability and the customer has not explicitly requested a human, the agent should attempt the supported resolution while remaining ready to escalate if the customer states that preference.

Correct answer: A. Rewrite delegated Task prompts to state the support objective, constraints, case facts, and resolution quality criteria, leaving investigation paths flexible.

This approach improves reliability by giving the subagent the context and success criteria needed to adapt its investigation to unusual case details. The coordinator still defines the goal and quality bar, but it avoids over-constraining the subagent with brittle procedural steps.

Correct answer: D. Run an interview-style refinement pass where Claude asks maintainers targeted questions, then fold clarified criteria into the CI review prompt.

An interview-style refinement pass is the best next step because the core problem is underspecified requirements, not just prompt wording. Having the model ask maintainers targeted questions surfaces hidden criteria and domain-specific risks before the workflow becomes mandatory.

Correct answer: D. Route subagent messages, errors, and follow-up requests through the coordinator, which manages delegation, recovery, and information flow.

A coordinator-centered pattern gives one agent responsibility for routing information, tracking failures, and deciding what should happen next. This improves observability because every handoff, retry, and aggregation step passes through a single control point, unlike a peer-to-peer mesh.

Correct answer: B. Add a stateful prerequisite gate that blocks process_refund unless verified customer and order owner IDs match, then escalates mismatches.

Adding a stateful prerequisite gate creates a deterministic control point outside the model's probabilistic reasoning. Since the required workflow depends on verified state from prior tool results, the gate directly prevents the unsafe downstream action when the prerequisite condition is not satisfied.

Correct answer: A. Restrict each subagent to role-relevant tools, adding only narrow cross-role utilities for frequent simple checks like source metadata lookup.

Scoped tool access keeps each subagent focused on its specialization and reduces tool selection confusion. Allowing only narrow cross-role utilities for common simple needs preserves efficiency without letting subagents take over responsibilities better handled by other agents or the coordinator.

Correct answer: D. Require each extracted obligation to carry source excerpts and locations, then preserve and merge that evidence into each synthesized matrix row.

Multi-source synthesis needs provenance that survives every handoff. The system should require each extracted obligation to include its supporting excerpt and location metadata, then merge those records into the final matrix row, rather than relying on a final bibliography or reconstructing citations later.

Correct answer: D. Keep universal review principles in project CLAUDE.md, and move specialized CI workflows into project skills invoked by relevant jobs.

Separating always-applicable guidance from task-specific procedures reduces irrelevant context during routine reviews. Use project CLAUDE.md for universal standards, and move specialized workflows into project skills invoked only by relevant jobs, preventing false positives on documentation-only pull requests.

Correct answer: A. Use Grep to search repository contents for the exact error string and refund handler names, then Read matching files.

Grep is the optimal tool to locate exact error strings and handler names across a repository. Glob only matches filenames, while reading entire directories wastes context and risks missing the targeted error message.

Correct answer: D. Configure Claude Code to emit structured JSON matching a schema containing required fields for inline review comments.

Configuring structured JSON output creates a predictable contract for reliable downstream automation. Markdown and raw text are fragile formats that require complex parsing, which inevitably breaks when wording slightly changes.

Correct answer: D. Add errorCategory and isRetryable fields to tool failures, retrying only failures marked retryable and preserving user-facing explanations.

Adding explicit error categories and retry flags gives the agent structured semantics for correct recovery routing. Relying on keyword parsing is brittle, and blanket escalation undermines the goal of automated first-contact resolution.

Correct answer: A. Define reportable evidence thresholds for performance findings, including unbounded input growth, hot paths, or repeated database calls inside loops.

Defining concrete evidence thresholds replaces vague caution with specific conditions the model can evaluate. Asking the model for longer reasoning or higher confidence scores fails to eliminate speculative findings because it lacks measurable constraints.

Correct answer: A. Review each changed file separately for local prompt and schema defects, then run a final pass checking cross-file consistency.

Reviewing individual files isolates local logic without exhausting the context window, preventing superficial later analysis. A final integration pass ensures schemas and validators match, whereas bulk reviews suffer from lost-in-the-middle attention failures.

Correct answer: A. Extend the enum with unclear and other, plus a required detail field when other is selected.

Extending the schema with fallback categories handles ambiguous data without breaking validation rules. Forcing existing enum values creates inaccurate reporting, while using unrestricted strings destroys the reliable structure required by downstream systems.

Correct answer: A. Create a .claude/rules/migrations.md file with YAML paths matching all migration locations and keep the conventions there.

Using a .claude/rules file with YAML path matching correctly applies guidance dynamically based on file locations. Placing everything in the root CLAUDE.md would load irrelevant instructions for every CI run, increasing noise and defeating your targeted scoping goals.

Correct answer: B. Add a structured field capturing the triggering code pattern, then analyze dismissal rates by that field during prompt iterations.

Capturing the specific code construct that triggered a finding makes false-positive analysis actionable. Aggregating dismissals by this field pinpoints problematic patterns, whereas asking the model to self-report confidence just masks bad logic without providing measurable data.

Correct answer: A. Configure the skill with context: fork so verbose exploration runs separately and returns only a concise summary.

Configuring the skill with context fork runs the verbose exploration in an isolated sub-agent context. This prevents setup noise from filling the main conversation window, whereas simply instructing the prompt to ignore earlier notes fails to prevent token bloat.

Correct answer: C. Rewrite the MCP tool description to specify coverage, accepted queries, returned citation metadata, and when agents should prefer it.

Tool descriptions act as the primary routing signal for Claude, so expanding the description clarifies when to use it. Renaming the tool with buzzwords fails to convey actual constraints or output expectations, leaving the agents confused about its intended coverage.

Correct answer: B. Move narrowly scoped guidance into topic-specific files under .claude/rules/, leaving root CLAUDE.md with only universal CI review principles.

Splitting massive files into focused topic-specific rule files significantly improves maintainability and reduces context noise. Leaving only universal CI principles in the root CLAUDE.md prevents unrelated standards from blending together, whereas broad headings fail to stop bloat.

Correct answer: B. Add 2 to 4 representative input and ideal output examples showing concise obligations, evidence boundaries, and owner wording.

Few-shot examples are the strongest next step when detailed instructions alone fail to stabilize formatting and actionability. Representative input and output pairs clearly demonstrate exactly how ambiguous source text should be extracted, outperforming rigid word limits.

Correct answer: B. Create separate issue records after customer verification, investigate independent concerns using shared case facts, then synthesize one customer-facing resolution.

Creating separate issue records preserves boundaries between distinct concerns while utilizing shared verified facts. Synthesizing one unified response resolves the multiple issues cleanly, whereas forcing sequential replies or splitting tickets worsens the customer experience.

Correct answer: B. Classify missing-baseline failures separately, retry only format or placement errors, and add the baseline artifact before re-extracting.

Retries are useful when the model has the needed information but produced it in the wrong structure or format. When the evidence is absent from the provided context, the correct recovery is to supply the missing source material or classify the case as unresolved rather than asking the model to invent evidence.

Correct answer: A. Rewrite returns_helper MCP tool description to specify eligibility inputs, returned fields, examples, and boundaries with refund tools.

Tool descriptions are the primary signal Claude uses to decide when and how to call MCP tools. Making the tool's purpose, inputs, outputs, examples, and boundaries explicit directly improves model-driven tool selection without adding brittle orchestration.

Correct answer: D. Add a tool call interception hook that blocks process_refund above $150 and redirects the case to human escalation.

A tool call interception hook provides deterministic enforcement before the refund action reaches the backend system. This is the right approach when a business rule must be guaranteed rather than merely encouraged through instructions.

Correct answer: B. Begin with a brief fact-gathering pass, then let the coordinator create targeted follow-up subtasks as tool findings reveal the dispute type.

This matches an open-ended investigation where the correct next steps depend on what is discovered during the case. A lightweight initial pass establishes the facts, then the coordinator can adapt the plan and delegate only the investigations needed for the actual dispute pattern.

Correct answer: B. Configure the review helper as a skill with context: fork, returning only a concise summary to the main CI review.

Skills can run in an isolated sub-agent context when configured with context fork, which prevents verbose exploratory work from polluting the main conversation. Returning only a concise summary keeps the final review focused on actionable findings while preserving the reusable workflow.

Correct answer: D. Decompose the request into concern-specific tasks, run relevant subagents in parallel with shared context, then synthesize one integrated report.

A multi-concern request should be split into distinct work items so each concern receives focused investigation. Running independent concerns in parallel improves latency, while shared context and final synthesis preserve coherence across the report.

Correct answer: C. Use Glob with the known path patterns to enumerate candidate template files, then Read only the selected matches.

This approach uses the purpose-built file path matching tool before consuming file contents. It minimizes context usage by separating file discovery from content inspection, then reading only the files that are actually relevant.

Correct answer: C. Run verbose repository discovery in an Explore subagent, returning concise findings before the main CI review produces structured comments.

This isolates the noisy discovery phase from the main review context while still preserving the important conclusions. The main review can then focus on applying judgment and producing actionable findings instead of carrying every intermediate search result and file excerpt.

Correct answer: A. Change intermediate handoffs so each extracted fact carries its source document, location, and supporting excerpt into final synthesis.

This preserves provenance at the same granularity as the extracted information, so downstream systems can trace each final field back to supporting evidence. It directly addresses the failure introduced by compression, where a summary keeps the conclusion but drops the source relationship needed for auditability.

Correct answer: C. Add targeted examples showing varied existing test structures and genuine gaps, including why each should be reported or skipped.

Targeted examples teach the model how to interpret ambiguous coverage evidence across different test organization patterns. Showing both acceptable existing coverage and genuine gaps helps the model generalize judgment instead of assuming unfamiliar test structures represent missing tests.

Correct answer: D. Start a new session, provide a structured summary of durable findings, and explicitly identify changed sources for fresh analysis.

Starting fresh avoids carrying forward stale tool results that remain embedded in the prior conversation context. A structured summary preserves durable decisions and research scope while forcing current sources to be reloaded and reanalyzed, rather than hoping the model ignores old data.

Correct answer: B. Have each subagent export structured state to a known location; the coordinator reloads and injects it when resuming.

Exporting structured state creates an explicit recovery mechanism independent of fragile conversational memory. On the exam, remember that asking Claude to reconstruct progress from transcripts is an anti-pattern; externalizing state is required for reliable agent resumption.

Correct answer: D. Return errorCategory and isRetryable per failure, marking permanent source issues non-retryable and transient timeouts retryable with clear messages.

Structured retry metadata allows agents to distinguish transient failures from permanent roadblocks. Marking permanent conditions as non-retryable prevents wasted latency, whereas asking the coordinator to infer retryability from prose introduces brittle natural language parsing.

Correct answer: A. Include the relevant verified customer facts, order details, and prior tool outputs directly in every delegated Task prompt.

Subagents operate with isolated context, meaning they only process information explicitly passed during invocation. Injecting verified facts and tool outputs directly into the task prompt guarantees accurate analysis, whereas relying on session memory or heuristic selection causes failures.

Correct answer: B. Use stratified random sampling across issue type, order age, seller channel, and resolution action for high-confidence auto-closures.

Stratified random sampling guarantees that rare but critical segments are represented in the review set. This allows you to measure hidden error rates in high-confidence closures, whereas simply increasing the uniform sample size or skipping high-confidence cases hides edge case failures.

Correct answer: C. Have each exploration subagent write key findings to scratchpad files at known paths, and require later prompts to reference them.

Writing key findings to scratchpad files creates durable state outside of noisy conversational context. Requiring later prompts to reference these files ensures exact knowledge reuse, whereas increasing the token budget fails to prevent degraded memory and hallucinations.

Correct answer: A. Run Claude Code with the -p or –print flag so it processes the prompt and exits after output.

Claude Code runs interactively by default, waiting for further input which causes CI timeouts. The dash p or print flag forces the script to process the prompt and exit immediately, whereas appending prompt instructions to avoid follow-up questions is unreliable.

Correct answer: D. Add explicit fallback enum values with companion detail fields so uncertain or novel categories are captured without forced misclassification.

Adding explicit fallback enum values with companion detail fields preserves schema validity while safely capturing uncertainty. This prevents semantic misclassification, whereas endlessly expanding enums or using free text degrades downstream reliability and breaks constraints.

Correct answer: A. Redesign the schema so unavailable source-dependent fields are nullable or optional, with instructions to return null unless directly evidenced.

Making source-dependent fields nullable directly addresses hallucination by allowing the model to represent missing information. Relying solely on stronger prompt wording often fails to prevent fabricated values when strict schema constraints force an output.

Correct answer: C. Generate a structured handoff with research topic, disputed claims, source citations, analysis performed, and recommended reviewer decisions.

A structured handoff gives the human reviewer the context needed to continue from the system's work instead of restarting the investigation. It captures the unresolved decision point, supporting evidence, prior analysis, and recommended next steps in a compact, actionable format.

Correct answer: A. When the API response has stop_reason tool_use, execute all tool_use blocks, append matching results, and continue until end_turn.

Executing all tool use blocks implements the agentic loop around the model's explicit control signal rather than assuming a single tool per response. Returning all matching results lets the coordinator reason over complete subagent outputs before synthesis.

Correct answer: C. Add a tool call interception hook that blocks process_refund above $500 and returns guidance to call escalate_to_human.

A tool call interception hook provides deterministic enforcement before the financial operation executes. It preserves normal autonomous refunds under the threshold while preventing policy-violating calls and redirecting the workflow to escalation.

Correct answer: A. Organize conclusions by evidentiary status, retaining source characterization and methodological context for each major claim in the final report.

Organizing conclusions by evidentiary status preserves uncertainty by making the strength of evidence visible to readers. Separating stronger conclusions from tentative ones prevents the report from flattening source quality into a single authoritative narrative.

Correct answer: D. Have the coordinator issue separate Task calls for independent research branches in one response, then aggregate their returned findings before synthesis.

Independent subagent work can be launched concurrently when the coordinator emits multiple tool calls in the same response. This preserves the coordinator as the aggregation point while reducing wall-clock time for branches that do not depend on each other.

Correct answer: D. Create separate fork_session branches from the completed analysis baseline, develop each framing independently, then compare outputs through the coordinator.

Forking from a shared baseline preserves the same verified starting context while isolating divergent exploration paths. This fits the need to compare independent approaches without allowing assumptions from one branch to contaminate the other.

Correct answer: C. Use Glob with targeted filename patterns, then Read only matching policy and test files before editing.

Using Glob is the appropriate technique when the selection criterion is a specific file path or naming pattern. It narrows the search space before reading files, which avoids loading unrelated repository content into context unnecessarily.

Correct answer: A. Run Claude Code with claude -p "Review the changed files for security regressions" so it prints output and exits.

The print flag is the documented way to run the CLI in non-interactive mode for automation pipelines. It processes the prompt, writes the response to standard output, and exits instead of opening an interactive session that halts the CI job.

Correct answer: D. Split the request into distinct industry and jurisdiction workstreams, pass shared scope definitions, investigate in parallel, then synthesize one report.

Splitting the request into distinct workstreams gives each subagent a focused slice of the prompt while preserving shared constraints. Parallel investigation improves coverage, and the final synthesis step resolves overlaps into a unified report rather than fragmented outputs.

Correct answer: D. Create a targeted test suite for expected behavior and edge cases, then iterate by feeding Claude the specific failing results.

Creating an executable test suite establishes an objective feedback loop for the model. Vague prompts like asking to be exhaustive lack measurable targets, whereas feeding exact failing test results directs the model to make targeted corrections.

Correct answer: D. Load the full file with Read, identify the intended refund block, then Write the complete corrected file.

Reading the full file provides necessary context when an edit operation fails due to duplicate code blocks. Writing the complete corrected file ensures the specific handler updates safely without risking unintended modifications to similar paths.

Correct answer: C. Have each subagent export structured state to fixed paths, then load a resume manifest into restarted coordinator prompts.

Exporting structured state to fixed paths preserves progress outside the volatile conversation window. Loading this resume manifest allows the coordinator to inject exact handoff details into new prompts, avoiding the token limits of session histories.

Correct answer: A. Modify upstream subagents to return compact structured records: key facts, citations, relevance scores, and only synthesis-relevant details.

Modifying upstream agents to return compact structured records solves context dilution by reducing token noise. Increasing the context limit fails because the synthesis agent still struggles to track relevant facts buried inside long narratives.

Correct answer: D. Add targeted examples showing correct extraction from multiple document layouts, including inline citations, appendix tables, and embedded methodology details.

Few-shot examples teach Claude how to map differently placed evidence into the same schema fields across varied layouts. Making these fields required is a risky anti-pattern because it pressures the model into fabricating values when the data is genuinely absent.

Correct answer: D. Create focused .claude/rules/ files with YAML frontmatter paths globs targeting generated clients, migrations, and frontend files.

Path-specific rule files are designed for conditional convention loading based on file paths. YAML frontmatter paths fields with glob patterns let Claude Code load only the relevant guidance for matching files, reducing irrelevant context and false positive review comments.

Correct answer: D. Use Read to load the full template, modify the intended section, then Write the complete corrected file back.

When an edit target is not unique, loading the file and rewriting the complete contents is the reliable fallback. As a practical cue, relying on broader text matching or scripting shell commands risks modifying the wrong occurrence and damaging file integrity.

Correct answer: A. Add a small set of labeled examples showing acceptable unit variants and genuinely ambiguous entries, including expected fields and reasoning.

Few-shot examples are highly effective when a model must distinguish acceptable variations from genuine data issues. Showing both positive and negative cases provides a reliable decision boundary, whereas merely expanding prose rules or increasing thresholds fails to fix core logic.

Correct answer: D. Replace broad caution language with concrete report and skip criteria for each review category, including representative true-positive and non-issue examples.

Broad caution language fails to define the actual decision boundary for the model. Providing concrete category-specific criteria with true-positive and non-issue examples operationalizes the review process, whereas relying on self-reported confidence scores leaves the judgments noisy.

Correct answer: D. Add an orchestration-level prerequisite that rejects report-generation delegation until required results and source mappings are present.

An orchestration-level prerequisite reliably enforces workflow ordering by checking shared state before allowing downstream delegation. Prompt instructions are probabilistic and can be ignored, making them insufficient for mandatory dependency checks between subagents.

Correct answer: D. Configure a PreToolUse hook for the publishing tool that blocks calls missing the required citation manifest.

A PreToolUse hook intercepts the tool call before execution, providing deterministic enforcement to block unauthorized side effects. A PostToolUse hook fires after execution completes, meaning the draft would already be published before the check could reject it.

Correct answer: D. Provide the missing methodology sources, and allow an unavailable value when no cited source is in context.

The extraction failed because the methodology information was completely absent from the provided context. Structured output and strict tool use enforce schema compliance for existing text, but they cannot manufacture missing facts or recover unread documents.

Correct answer: C. Add a PreToolUse hook for Bash that rejects publishing commands missing an approved ticket ID.

A PreToolUse hook intercepts Bash commands before execution, deterministically blocking publishing commands that lack a ticket ID. Relying on context rules or few-shot examples remains probabilistic and cannot guarantee that unauthorized commands are prevented.

Correct answer: B. Append the assistant tool_use response, then send one following user message containing both tool_result blocks matched by tool_use_id.

The application must preserve the assistant turn containing the tool use requests, then return the outputs as tool result blocks in the next user message. Matching each result to its original tool use ID ensures Claude correctly interprets the execution feedback.

Correct answer: D. Add diverse few-shot examples that show issue and non-issue reviews, with brief reasoning that marks the decision boundary.

Providing diverse few-shot examples teaches Claude the underlying judgment pattern rather than just memorizing surface wording. Including both issues and non-issues establishes clear decision boundaries, helping the model apply the criteria accurately to novel refactors.

Correct answer: D. Update the lookup tool description to require clarification when multiple candidate tickets are returned.

Updating the tool description to require clarification makes the ambiguity explicit to Claude. Selecting the newest or highest-ranked ticket relies on risky heuristics, whereas asking for a unique identifier guarantees the correct ticket is targeted.

Correct answer: D. Move a concise findings-and-source map to the start of the synthesis prompt and keep detailed evidence under labeled sections.

Moving a concise findings map to the start directly mitigates the lost-in-the-middle effect where models ignore buried details. Simply increasing output length or adding reminders addresses adjacent symptoms without fixing the core context positioning problem.

Correct answer: D. Return a tool_result with is_error: true and JSON content containing errorCategory, isRetryable, a clear message, and a suggested valid key format.

Returning a tool result with is_error true and structured JSON provides the exact recovery metadata needed. Generic failures do not give Claude enough context to correct the malformed input and retry successfully.

Correct answer: C. Add 3 to 5 relevant, diverse, structured examples showing varied source formats and the expected extracted output.

Adding diverse few-shot examples teaches Claude the extraction patterns needed across varied document structures. Simply lowering temperature or demanding no nulls does not solve the core comprehension gap when parsing tables and release notes.

Correct answer: B. Add a PreToolUse command hook for Bash that validates target paths and blocks out-of-workspace commands with feedback.

A blocking PreToolUse hook enforces the workspace boundary deterministically before the Bash command executes. PostToolUse hooks or system prompts react after the fact, which violates the strict requirement to stop out-of-workspace actions before they run.

Correct answer: C. Have the main agent spawn the reviewers with Agent, pass each reviewer its needed context, collect summarized results, and synthesize follow-ups centrally.

Claude Code subagents run in isolated context windows and return summarized results to the caller. Keeping synthesis and routing in the main agent matches the documented hub and spoke pattern, avoiding the pitfalls of cross-chatting subagents.

Correct answer: D. Re-add the server with –scope project and commit the generated .mcp.json using environment-variable placeholders.

Project scope writes an MCP server configuration file to the repository root for easy team sharing. Using environment variable placeholders inside this file keeps sensitive credentials out of source control while standardizing the setup across user machines.

Correct answer: C. Turn the workflow into a prompt template, run representative edge-case evals with the same structured-output request shape, refine and re-run the prompt, then batch only after a synchronous dry run succeeds.

Running representative edge-case evaluations addresses both quality and parsing issues before submitting large batches. Relying solely on strict schema validation or blind prompt improvements ignores real-world failures that a synchronous dry run easily catches.

Correct answer: D. Use explicit prompt chaining with file-level review calls, then run a separate integration call over the local findings.

Explicit prompt chaining breaks the review into smaller, inspectable steps, which reduces attention dilution from large contexts. A final integration call focuses on cross-file consistency after each file receives dedicated local analysis.

Correct answer: B. Delegate code search and file discovery to the Explore subagent, and return concise findings to the main conversation.

Delegating code search to an exploration subagent moves high-volume output into a separate context window. This preserves the main conversation for high-level coordination without losing important architecture details to verbose exploration logs.

Correct answer: A. Have the main thread issue the independent Agent calls in one assistant turn, then return all matching tool_result blocks together in the next user message.

Issuing multiple independent Agent tool calls in a single assistant turn allows the orchestrator to execute them concurrently. Grouping the returned tool result blocks in the next user message completes the parallel cycle, preventing the main thread from blocking on sequential subagent responses.

Correct answer: B. Define an escalate_to_human client tool whose input_schema requires customer_id, root_cause, refund_amount, and recommended_action.

Defining an escalation client tool with a strict input schema forces the model to generate structured arguments for every handoff. This guarantees downstream systems receive consistent data, whereas relying on prompt instructions for ticket content remains purely probabilistic.

Correct answer: D. Start plan mode with all failing fixtures, shared schema constraints, and the validation command.

Because these parser failures share dependencies, plan mode allows Claude to inspect the codebase and propose a coherent strategy addressing all cross-effects upfront. Fixing independent issues sequentially often pollutes the context window and degrades results for unrelated tasks.

Correct answer: D. Add a short examples block showing representative finding inputs and ideal cited-bullet outputs, including edge cases.

Providing concrete input and output examples clarifies the exact transformation boundary that prose instructions leave ambiguous. Broad wording or configuration files fail to enforce strict structural mapping, and hooks are better suited for blocking side effects than formatting text.

Correct answer: D. Run claude -p "Continue the refactor investigation from yesterday." –resume "$session_id"

Passing the captured session ID to the resume flag explicitly restores the exact prior conversation context before sending the new prompt. Putting the identifier inside the prompt text only passes a string to the model, failing to direct the CLI to load that session.

Correct answer: C. Run a separate review request that receives the source document, extracted JSON, and explicit discrepancy criteria.

Providing the source document, extracted output, and explicit discrepancy criteria creates an isolated review pass that objectively audits the original extraction. A follow-up prompt in the same context risks confirmation bias, as the model often assumes its previous output is already correct.

Correct answer: D. Add three to five tagged source snippet and expected JSON examples covering typical and edge-case missing values.

Providing concrete input and output examples is the most reliable way to enforce strict formatting rules and handle edge cases. Expanding abstract prose instructions often fails to correct nuanced mapping inconsistencies because the model lacks specific examples to anchor its behavior.

Correct answer: B. Segment the evaluation by document type and extracted field before changing review thresholds.

Segmenting the evaluation by document type and field reveals concentrated failures that a high aggregate accuracy score easily masks. Relying solely on an overall metric might cause you to approve a release with critical errors in rarer document categories.

Correct answer: C. Return a CallToolResult with isError set and structuredContent describing the error category, retryability, affected fields, and correction hint.

Returning a tool result with the error flag set and structured content gives Claude machine-readable context for recovery. Treating failures as normal text results causes loops because the model lacks the clear signal needed to understand the validation failure.

Correct answer: D. Give the synthesis subagent a narrow verify_fact tool and route broader investigation requests through the coordinator.

Providing a narrow verification tool allows the synthesis subagent to quickly resolve minor evidence questions without overstepping its role. Giving it broad exploration tools would duplicate work, while removing all tools causes unnecessary coordinator delays.

Correct answer: B. Return structured tool_result errors with failure category, retryability, message, and corrective guidance.

Structured tool results provide the exact failure category, retryability, and corrective guidance Claude needs to choose the right next step. Generic error messages force the model to guess, causing inappropriate retries or escalations that break the workflow.

Correct answer: B. Pass the relevant findings and source metadata directly in the synthesis task message.

Passing the relevant findings and source metadata directly in the task message gives the subagent the required data. Subagents operate with isolated context windows, so they cannot automatically read the previous research transcripts stored by the coordinator.

Correct answer: B. Add a PreToolUse prerequisite gate that denies Bash until the project root check has been recorded.

A PreToolUse gate intercepts the request before execution, allowing you to programmatically block unsafe tool calls. System prompts and few-shot examples are behavioral suggestions, meaning the model might still bypass the required ordering under complex conditions.

Correct answer: D. Connect each required MCP server for the subagent, list tools after each connection, and expose a unified registry while routing calls to the owning server.

Connecting to all required servers simultaneously allows the host to aggregate tools into a unified registry for the subagent. Switching between single active servers creates unnecessary connection overhead and limits the agent from chaining tools effectively.

Correct answer: B. Have the search subagent retry transient failures locally, then return attempted queries, unresolved error context, and any partial sources.

Local retries keep transient recovery inside the worker while returning partial sources and errors gives the coordinator enough context to proceed. Avoid having the parent inspect intermediate tool calls, because that breaks subagent isolation and defeats the purpose of delegation.

Correct answer: B. Generate a schema-validated escalation packet with Structured outputs as the final reviewer handoff.

Structured outputs generate schema-validated JSON as the final response, satisfying the queue requirements without relying on hidden tool history. Option A fails because a tool call yields arguments rather than the required final text response.

Correct answer: B. Use Grep to search repository file contents for the exact insured_party reference.

Use Grep to search file contents when you know the target string but not the specific files. Glob fails here because it only matches file paths and names, while Bash is less efficient than dedicated built-in tools.

Correct answer: D. Send a new Messages API request with XML-tagged sections for the original document, rejected JSON, validator errors, and instructions to return only corrected structured data.

Because the API is stateless, a retry must include the original document, rejected output, and exact errors so Claude can repair against concrete evidence. Using XML tags separates these mixed content types, which is much more reliable than simply asking the model to be more careful.

Correct answer: C. Have the investigation agent maintain a concise investigation-notes.md file of confirmed dependencies and read it before follow-up questions.

Writing a durable notes file gives Claude a stable place to store confirmed findings outside the active context window. Reading this file before follow-up questions reliably restores specific state after long sessions or compaction, whereas plan mode does not preserve tool outputs.

Correct answer: D. Move the patent landscape procedure into a SKILL.md-based skill and keep only universal research conventions in CLAUDE.md.

Move specialized procedures into a skill file so the context loads only when that specific workflow is triggered. Keep universal project rules in your root memory file to prevent bloating unrelated sessions with unnecessary tokens.

Correct answer: C. Define a client tool in the tools array with input_schema, set strict: true, and force that tool with tool_choice.

Forcing a strict tool call guarantees the output exactly matches your schema and prevents the model from returning standard text. Relying on prompt instructions or prefills still leaves room for formatting errors and hallucinated commentary.

Correct answer: C. Use POST /v1/messages for the merge gate, and submit the weekly audit through the Message Batches API with custom_id values.

The merge gate needs low latency, so use a synchronous message call. The massive weekly audit is asynchronous, making the batches API with custom tracking IDs the correct choice to handle high volume without timing out.

Correct answer: D. Have Claude explore the existing invoice extraction flow, then propose a migration plan for changing how ambiguous totals are represented.

Plan mode is ideal when a change might impact multiple parts of a system, like altering data representation across parsing logic and schemas. Skip plan mode for trivial single-line edits like fixing typos or adding a basic log statement.

Correct answer: D. Route explicit human requests directly to review, and offer automated repair for straightforward correction requests.

Always route explicit human requests to review immediately to respect user intent. Offer automated repair for straightforward corrections when escalation is not demanded, rather than relying strictly on confidence or sentiment scores.

Correct answer: D. Run /memory in the new developer's session and inspect the loaded instruction files.

The /memory command reveals exactly which CLAUDE.md files are loaded in a session, making it the ideal tool for this diagnosis. Use /status only when troubleshooting settings files rather than memory files.

Correct answer: D. Commit a repository-root CLAUDE.md containing the testing standards, fixture conventions, and review criteria.

Committing a repository root memory file ensures that both local developers and automated CI runs automatically load the shared project rules. Workflow prompt inputs are better suited for transient instructions rather than durable team standards.

Correct answer: C. Replace the broad keyword rule with a narrower instruction to use search tools only when repository or MCP data must be inspected.

Replacing the broad keyword rule directly fixes the system prompt instruction that is incorrectly forcing tool use. Broad keyword triggers can easily override clear tool descriptions, so narrowing the instruction to actual data lookup ensures tools are only called when needed.

Correct answer: A. Persist previous PR findings and pass them into each rerun, instructing Claude to report only new or still-unaddressed issues.

Persisting previous findings gives each review run the context needed to distinguish resolved, still-open, and newly introduced issues. Matching only by line number or ignoring files is too brittle and can easily hide newly introduced defects in previously reviewed code.

Correct answer: D. Put shared conventions in root CLAUDE.md or .claude/CLAUDE.md, put report-agent guidance in that directory's CLAUDE.md, and keep personal preferences in ~/.claude/CLAUDE.md.

This correctly uses the documented Claude Code memory hierarchy by placing team conventions in the root file and personal preferences in the user file. Lower-level files do not replace parent files but are combined, so relying on overrides or shared user files fails.

Correct answer: C. Start a fresh session and inject a structured summary of validated findings, decisions, and sources to recheck.

Starting a fresh session avoids carrying stale tool results and outdated documents back into the model context. Injecting a structured summary preserves validated findings without risking hallucination from instructing the model to ignore old context.

Correct answer: A. Append the assistant message with the tool requests, then append one user message containing both matching tool results before the next request.

Appending the assistant message and then a single user message with both tool results maintains the exact conversation state required. The API expects all tool results for a given turn to be returned together in one user message so the model can map them correctly.

Correct answer: D. Rewrite each tool description to define use cases, non-use cases, input formats, example queries, returned data, and edge cases.

Tool descriptions are the highest-leverage signal Claude uses when deciding which tool to call. Broad instructions or tool renames do not directly repair the model's decision boundary between overlapping tools like explicit descriptions do.

Correct answer: D. Add Agent to the coordinator's allowedTools or allowed_tools configuration.

The documented tool for spawning a subagent is Agent, and SDK configurations must include it in allowed tools for auto-approval. Simply allowing the subagent names directly will fail to trigger the native delegation tool without explicit invocation setups.

Correct answer: A. Define the tools with clear descriptions and schemas, then run a loop that executes each tool_use request and returns matching tool_result blocks.

This uses the documented tool use contract where Claude decides which client tools to call, while the application executes them. Fixed controllers or forced tool choices prevent the model from dynamically adapting its path based on the document.

Correct answer: B. Require each subagent to return structured evidence records with the claim, source location, document date, and method context.

Requiring structured evidence records directly improves the handoff by preserving provenance and methodology before summarization compresses the details. Downstream synthesis only receives what the coordinator passes forward, so prose summaries often lose critical context.

Correct answer: B. Submit the review prompts as batch items, retrieve any tool_use results by custom_id, run the requested client tools, and send follow-up requests with tool_result blocks.

The Message Batches API processes asynchronously, so your application must execute any client tools after retrieving the initial results. You then send follow-up requests containing the tool results to complete the workflow. Batch workers cannot execute local code directly.

Correct answer: B. Create ~/.claude/skills/my-code-review/SKILL.md and invoke it as /my-code-review.

Placing a uniquely named skill directory under the personal tilde slash claude path ensures it remains private to you and avoids overriding the shared project skill. Project-level files are shared, whereas user-level files apply only to your local machine.

Correct answer: A. Run claude -p with –output-format json and –json-schema, then parse structured_output for the PR comment payload.

Using print mode with a JSON schema strictly enforces the required fields for the downstream script by returning guaranteed valid data in structured output. Plain text outputs lack programmatic reliability and break easily in automated pipelines.

Correct answer: A. Delegate focused read-only investigations to subagents, then use their returned summaries in the main conversation.

Delegating focused read-only investigations to subagents isolates the noisy exploration logs from the main coordination thread. Pasting full transcripts directly into the main conversation defeats this goal by consuming context window space with irrelevant details.

Correct answer: C. Run Claude Code with -p, –output-format json, and –json-schema, then read the validated result from structured_output.

Print mode with JSON output and a schema ensures the PR bot receives strictly validated data inside structured output while retaining top-level metadata. Without schema enforcement, relying on prompt instructions alone risks posting malformed comments.

Correct answer: C. Maintain a structured scratchpad of exact facts that is updated separately from compacted conversation history.

Maintaining a separate structured scratchpad of exact facts protects critical numerical values from being degraded during conversation compaction. Relying on generalized summaries inevitably leads to broken logic when precise constants are required.

Correct answer: D. Replace the vague flagging instruction with measurable rules for when parsed line items conflict with stated totals.

Replacing subjective instructions with specific, measurable rules eliminates ambiguity and ensures consistent flagging of arithmetic mismatches. Broad instructions cause the model to guess inconsistently, leading to false positives and missed errors.

Correct answer: A. Check stop_reason, retry truncated generations with a larger output limit, then return specific semantic validation errors for correction.

Checking the stop reason handles truncation by prompting a retry with an expanded token limit. Relying solely on syntax validation fails here because semantic validation is required to catch nonexistent line numbers and correct the model.

Correct answer: A. Read the current file, construct the complete desired file contents, then overwrite it with Write.

Reading the file and using the Write tool to overwrite it with the exact desired contents provides a reliable last-resort edit when exact matching fails. Using replace all incorrectly modifies every matching snippet instead of making a single targeted change.

Correct answer: D. Add a PreToolUse command hook that denies approval commands when the passing artifact is absent.

A PreToolUse command hook runs before execution, providing deterministic control to block forbidden actions. PostToolUse fires too late, while relying on system prompts leaves enforcement up to unpredictable model compliance.

Correct answer: D. Change the tool wrapper to return compact, review-specific tool_result payloads with only the fields needed.

Tool results consume context window space, so shaping tool output before returning it preserves room for critical evidence. Prompt caching reduces latency, but cached tokens still count against the context limit.

Correct answer: D. Add an application-side guard that blocks over-threshold process_refund calls and routes the case to escalate_to_human.

The application boundary is the correct place to enforce non-negotiable business rules for client-executed tools. Validating the outgoing tool call before execution deterministically prevents high-impact errors that relying on prompts cannot guarantee.

Correct answer: B. Sample high-confidence findings within strata for change type, component, language, and finding category, then label errors and novel patterns separately.

Stratified sampling preserves visibility into high-confidence outputs while focusing limited reviewer capacity on meaningful production segments. Sampling by change type or language helps reveal hidden pockets of poor performance that aggregate metrics miss.

Correct answer: B. Return categorized tool_result errors with retryability, cause, attempted input, and suggested recovery.

Claude reasons directly from tool result content. Categorized errors allow the coordinator to distinguish retryable failures from valid empty results. Masking failures as empty sets removes the context needed for proper recovery.

Correct answer: C. Add three representative before-and-after examples, including edge cases, in a clearly tagged examples section.

Concrete input and output examples teach the desired transformation boundary much better than prose alone. Plan mode outlines strategy but does not enforce the specific formatting rules needed for consistent transformations.

Correct answer: D. Run Claude Code with the -p or –print flag and provide the review prompt non-interactively.

The print flag processes a prompt and exits without waiting for terminal input. Configuration files tell Claude what to know but do not prevent unattended pipelines from hanging when the system requests human interaction.

Correct answer: D. Move topic-specific guidance into Markdown files under .claude/rules/ and add paths frontmatter for rules that apply only to matching files.

Path scoped rules load only when Claude works with matching file patterns preventing deployment noise during test reviews. Imports still expand into context at launch so they organize files for humans without reducing actual context size.

Correct answer: D. Define a programmatic agents map keyed by subagent type, set each description and prompt, and give the documentation subagent a read-only tools allowlist.

The AgentDefinition uses the map key for routing and the tools field as a strict allowlist. Top level settings auto approve tools but do not restrict the subagent from using others making it unsafe for read only tasks.

Correct answer: C. Pass each prior subagent's complete structured findings directly into the review synthesis subagent's prompt.

Subagents are isolated instances so you must pass required facts into their prompts explicitly. Inferring details from parent history fails because the synthesis agent lacks access to that earlier conversation context.

Correct answer: B. Include the changed files, key diff excerpts, failing test output, and review criteria in the subagent prompt.

Non fork subagents start with isolated context so the delegation prompt must carry the necessary facts. Compacting the parent conversation does not pass context down and the agent cannot read transcripts automatically.

Correct answer: D. Run an explicit prompt chain with source-level review passes, then run a separate integration pass over the reviewed findings.

Prompt chaining creates inspectable intermediate artifacts for logging and dedicated conflict reconciliation. Relying on confidence scores or duplicating overloaded reviewers does not solve the underlying context rot degrading accuracy.

Correct answer: A. Return an insufficient information result and require a supporting quote before posting the finding.

Returning an insufficient information result correctly handles the missing supporting quote, as the provided source lacks transcript policy. Retrying the prompt cannot create policy evidence that was never supplied in the context.

Correct answer: A. Return tool results for both calls, flag lookup_order with is_error: true, and continue from partial context.

Flagging the timed-out tool with is_error true preserves the tool result protocol while giving Claude accurate context to escalate or retry. Masking the timeout as an empty list tricks the model into assuming no orders exist, causing flawed logic.

Correct answer: A. Run independent reviewer agents on section-specific claim-source bundles, then run a separate integration pass for conflicts and coverage gaps.

Running independent reviewer agents on section-specific bundles reduces self-reinforcing context errors. Relying on a single integration pass by the synthesis agent misses cross-section contradictions because it lacks independent verification.

Correct answer: D. Rewrite each tool definition with detailed descriptions covering use cases, non-use cases, inputs, returned data, and caveats.

Rewriting tool definitions with detailed use cases and boundaries is the highest-leverage fix for incorrect tool selection among similar capabilities. Shorter names or isolated examples do not provide the clear decision boundaries the model needs.

Correct answer: C. Require each subagent to return structured claim records with source location, publication date, evidence excerpt, and methodology notes.

Requiring structured claim records forces subagents to pass provenance and methodology to the synthesizer explicitly. Relying on concise summaries permanently destroys the underlying evidence needed to weigh conflicting claims.

Correct answer: A. Add Task to the coordinator SDK run's allowedTools or allowed_tools list.

Adding Task to the allowed tools list authorizes the coordinator to execute subagent delegation without pausing. Adding Agent only to worker subagents fails because the interactive pause happens before workers are ever spawned.

Correct answer: B. Rename the tools with source-specific names, rewrite each description with use and non-use boundaries, and update the dispatcher for the new names.

Renaming the tools with source-specific names and defining clear use boundaries directly resolves overlapping description ambiguity. Merging them into a generic tool violates the requirement to maintain separate output contracts and audit logs.

Correct answer: C. Declare the server's resources capability, expose catalog metadata through resources/list, and read selected URI contents with resources/read.

Using MCP resources lets the host application control context attachment via list and read primitives. Forcing Claude to call discovery tools wastes conversation turns and directly violates the team's context loading requirement.

Correct answer: C. Use dynamic adaptive decomposition so Claude inspects the module, creates subtasks from findings, and synthesizes the test plan.

Dynamic adaptive decomposition is optimal for unfamiliar legacy code because subtasks depend on discovered dependencies. Fixed prompt chaining fails because it assumes a rigid sequence of steps that may not apply to the specific module.

Correct answer: B. Add .claude/rules/database-migrations.md with YAML frontmatter paths: ["**/migrations/**/*.{sql,ts}"] and place the migration review checklist there.

Using YAML frontmatter in a rules file correctly scopes the migration checklist to relevant file paths, preventing false positives. Placing rules in the root CLAUDE.md applies them globally, which continues to generate irrelevant review noise.

Correct answer: D. Have the coordinator split the request into dimension-specific subtasks, pass shared constraints to each, invoke independent specialists together, and synthesize their findings.

Splitting the request into dimension-specific subtasks allows independent specialists to execute in parallel, drastically reducing latency and improving coverage. Keeping one broad subtask or disabling parallel tool use creates unnecessary sequential bottlenecks for independent workstreams.

Correct answer: B. Prepend a concise findings map before the aggregated outputs, then group detailed evidence under explicit source and topic headers.

Prepending a concise findings map mitigates lost-in-the-middle attention limits by placing critical cues at the top, guiding the model through the entire bundle. Explicit source and topic headers further help the model navigate dense context without dropping provenance.

Correct answer: B. Add downstream business-rule validation before refund execution and feed specific failures back for correction.

Tool schemas guarantee the correct JSON shape, but they cannot enforce business logic. You must add downstream code to validate domain rules like field semantics and totals, returning failures to the model for correction.

Correct answer: B. Return a CallToolResult with isError: true and place custom errorCategory, isRetryable, and policy text in structuredContent.

Using the official MCP error indicator clearly signals the agent that execution failed, preventing blind retries. Placing custom metadata in structured content allows the agent to evaluate retryability and understand the specific policy failure.

Correct answer: B. Start with a discovery pass that maps the topic structure and high-impact areas, then assign prioritized subtasks that can change as findings reveal dependencies.

Open-ended research benefits from separating exploration from execution. A discovery pass maps the problem space, enabling the coordinator to create dynamic subtasks and adapt when early findings reveal gaps or dependencies.

Correct answer: B. Persist and resend the relevant messages array, including earlier assistant outputs, tool results, and the current user turn.

The Messages API is stateless, so Claude cannot implicitly remember prior turns. Persisting and resending the relevant messages array provides the historical context needed for the model to synthesize findings consistently.

Correct answer: C. Emit per-field confidence values in structured output, calibrate thresholds on labeled examples, and route low-confidence fields to reviewers.

Emitting per-field confidence values targets reviewer attention precisely where risk is highest. Calibrating thresholds using a labeled validation set ensures raw model scores accurately reflect observed accuracy for each specific field.

Correct answer: A. Add a PostToolUse hook that reads tool_response and returns hookSpecificOutput.updatedToolOutput with normalized fields.

A PostToolUse hook runs after a successful tool call, making it the right place to transform data. Using the updatedToolOutput field replaces the raw response with normalized fields, ensuring Claude reasons over canonical formats.

Correct answer: B. Run a prompt chain that reviews files in focused passes, then runs a separate integration pass over the collected findings.

A prompt chain reduces attention dilution by reviewing files in focused passes before running a dedicated integration pass. Running independent reviewers without a synthesis step fails because no one reconciles the cross-file findings.

Correct answer: A. Commit a .mcp.json file that references Bearer ${GITHUB_TOKEN} in headers, and set GITHUB_TOKEN in each CI and developer environment.

Project scoped configuration belongs in the tracked file while environment variables keep secrets out of version control. Option C fails because Claude Code requires the specific variable syntax to expand headers properly during the connection.

Correct answer: C. Run an interactive planning session where Claude uses AskUserQuestion, writes a self-contained SPEC.md, then implement it from a fresh session.

Holding an interactive planning session lets Claude ask clarifying questions and produce a self-contained specification before any code is written. Avoid jumping straight into implementation for ambiguous features, because starting a fresh session from a detailed spec keeps execution focused.

Correct answer: A. Configure a maintained Jira MCP server for Jira operations, and implement a custom MCP server for the release-readiness workflow.

Using a maintained Model Context Protocol server for standard tools reduces integration debt, while building a custom server encapsulates complex private business logic. Avoid hardcoding REST procedures via Bash, because standard servers provide a much cleaner and more maintainable interface.

Correct answer: A. Track precision and recall by category, then tighten or disable the noisy style criteria using flagged and non-flagged examples.

Tracking precision and recall by category isolates the noisy style criteria responsible for eroding developer trust. A strong exam cue is to avoid aggregate metrics here, because an overall acceptance rate hides the specific category causing false positives.

Correct answer: A. Resume the baseline session with the SDK fork option enabled for each strategy, then track the new session IDs.

Forking a specific session creates a new branch that starts with a copy of the original history, leaving the baseline unchanged. Remember that simply resuming normally modifies the tracked history, whereas forking is the explicit mechanism for isolated branching.

Correct answer: D. Add context: fork in SKILL.md and keep the audit as explicit task instructions.

Using the fork context setting isolates verbose discovery work inside a subagent so the main conversation only receives the useful result. Remember that standard compaction or memory instructions do not prevent context flooding the way forking a side task does.

Correct answer: B. Persist an application-defined JSON state manifest, then load it on restart and inject it into coordinator and subagent prompts.

Persisting an application-defined state manifest and injecting it into the prompts explicitly recovers the verified facts after a crash. Do not rely on local transcript files or session memory for critical recovery, because replacement workers often cannot access them.

Correct answer: C. Move source-check.md into .claude/commands/ in the repository and commit that file.

Committing the command file inside the project commands directory makes it a version-controlled asset that all teammates receive upon cloning. Personal commands stored in the home directory are always private to the individual user and invisible to the repository.

Correct answer: C. Create a .claude/rules/testing.md file with paths globs for test files and store the shared testing conventions there.

Using a rules file with path globs ensures these conventions are loaded dynamically only when matching test files are read. A key exam cue is to avoid duplicating instructions in multiple files or placing them globally in the root memory, which wastes context.

Correct answer: A. Define testable report and skip rules that focus comments on production bugs, security flaws, broken edge cases, and subtle regressions.

Defining strict report and skip rules creates clear category boundaries, filtering out subjective stylistic feedback in favor of high-impact bugs. Requiring numerical confidence scores fails because it still leaves the model trying to quantify inherently subjective stylistic opinions.

Correct answer: B. Create .claude/skills/case-review/SKILL.md with context: fork, allowed-tools for Read and Grep, and argument-hint: [case-id].

Claude Code skills require a dedicated folder containing a SKILL.md file to properly register custom workflows and parse YAML frontmatter. Setting context to fork creates an isolated subagent for background note generation, while allowed tools prevents permission prompts and argument hint guides developers.

Correct answer: C. Add a coordinator step that inspects the case, invokes only needed subagents, and passes each a self-contained task summary.

Introducing a coordinator step allows the system to dynamically route simple requests away from unnecessary specialized subagents, immediately reducing latency. Because subagents operate with isolated context, the coordinator must always pass a self-contained summary to invoke them correctly.

Correct answer: B. Add 3 to 5 structured examples that map varied guide formats to the expected JSON fields.

Adding diverse few-shot examples directly teaches the model how to map varied input formats like shell blocks and checklists into the required schema. Strengthening the negative instruction alone fails because the model currently lacks the structural mapping needed to extract the values.

Correct answer: C. Return a CallToolResult with isError: true and include structured failure details in the result content.

Returning a CallToolResult with the error flag set to true explicitly signals an application-level failure to the agent, preventing false success claims. Returning a normal result with an error field risks the model ignoring the structured payload and hallucinating a positive outcome.

Correct answer: A. Temporarily omit the noisy category from the active prompt, keep accurate categories enabled, and rebuild it with examples, definitions, and precision/recall evals.

Temporarily removing the noisy category immediately restores trust in the automated pipeline while keeping the accurate signals active. Rebuilding the prompt with positive and negative examples alongside precision metrics addresses the underlying classification failure directly.

Correct answer: D. Send an immediate tool_result for that tool_use_id with is_error true and the concrete validation failures.

Returning a tool result with the error flag set provides the exact validation failures needed for correction. Specific feedback is far more effective than generic prompt tweaks or random temperature retries.

Correct answer: B. Resume the saved session ID with the fork option for each attempt, and checkpoint files before edits.

Forking a saved session creates separate histories from the same baseline without altering the original. Since a session is not a filesystem snapshot, checkpointing files ensures safe reversion for overlapping edits.

Correct answer: C. Use Grep for refund and escalation identifiers, then Read only matching files and follow imports from those entry points.

Using Grep to search for identifiers and then reading only the matching files builds understanding incrementally. This prevents context bloat better than reading entire directories or pasting raw search outputs.

Correct answer: B. Add 2-4 tagged examples showing similar ambiguous requests, the selected response path, and concise reasoning against plausible alternatives.

Tagged few-shot examples teach the model the exact decision boundary for ambiguous requests. Explicit reasoning alongside these examples resolves judgment calls better than long written checklists.

Correct answer: B. Run each requested Task tool call, send matching tool_result blocks, and keep calling the Messages API until stop_reason is end_turn.

The agent loop must execute requested tools and return their results before continuing. A tool use request is not a final answer, so the coordinator stops prematurely if the loop lacks result handling.

Correct answer: B. Add per-field confidence, ambiguity, contradiction, and evidence-excerpt fields to the JSON schema, then queue flagged records before clean records.

Defining explicit application-level signals for ambiguity and contradiction catches the specific failure modes that a successful stop reason cannot detect. On the exam, remember that model generation metadata like stop reason or logprobs is never a safe substitute for validating structured extraction logic.

Correct answer: D. Separate review and test-generation steps, expose only role-specific tools to each step, and force the review step's reporting tool with tool_choice.

Separating the steps and restricting tool access ensures the review phase only sees its specific reporting tool, preventing irrelevant actions. For the exam, remember that setting tool_choice to a specific named tool is the safest way to force structured output.

Correct answer: A. Enter plan mode, have Claude inspect the affected areas, and review the proposed plan before choosing execution mode.

Plan mode directly separates codebase exploration from actual edits, allowing Claude to propose an approach without risking unintended changes. For large migrations, starting in bypass permissions or accept edits risks immediate regressions before the maintainers even approve a design.

Correct answer: C. Submit accumulated requests every 4 hours, using custom_id to correlate each case outcome.

A thirty-hour service level agreement paired with a twenty-four-hour maximum batch window leaves only six hours for safe accumulation. Submitting every four hours keeps the worst-case timeline within the strict audit limit while providing a crucial two-hour safety buffer.

Correct answer: C. Create a Markdown rule file under .claude/rules/ with YAML frontmatter using paths glob patterns for the test file names.

Using path-specific rule files applies instructions only to matching files. This targets conventions precisely, whereas root or user-level files apply globally and load unnecessary context for production files.

Correct answer: A. Have the subagent retry transient command failures locally, then return attempts, partial findings, and unresolved errors in its final result.

A subagent should retry transient failures locally and summarize partial findings in its final result. This isolates messy command logs from the parent context while preserving useful details for planning.

Correct answer: B. Add REVIEW.md guidance that uses prior findings during re-reviews and reports only new or still-unresolved issues.

Configuring review guidance ensures the system uses prior findings during re-reviews. This prevents duplicate comments by shaping behavior directly, rather than relying on manual cleanup by maintainers.

Correct answer: C. Use batch for initial analysis or tool requests, then execute returned client tools and continue each case in later Messages requests.

Batch processing is asynchronous and cannot execute client tools directly. The batch should perform initial analysis, then your application handles the tools and continues in standard requests.

Correct answer: D. Run the SDK-driven agent loop so each MCP tool invocation is dispatched, results are returned, and processing continues until a final response.

The agent runtime must dispatch tool requests and return observations until the model finishes. Claude cannot execute backend operations itself, so a missing runtime loop causes the session to halt prematurely.

Correct answer: B. Set tool_choice to {"type": "any"} while providing all extraction tools.

Setting tool_choice to any forces Claude to use one of the provided tools, resolving the conversational-summary failure mode. Auto allows conversational responses, while forcing a specific schema breaks extraction when the document type is unknown.

Correct answer: B. Declare a client extraction tool with an input_schema for the release note fields, require that tool call, and read the tool_use input object.

Declaring an extraction tool with a JSON schema ensures Claude emits structured arguments in a reliable tool_use block. Parsing raw assistant text often fails because the model might include unwanted conversational commentary before the data.

Correct answer: B. Continue the loop whenever stop_reason is "tool_use", return matching tool_result blocks, and use the response only when stop_reason is "end_turn".

The client tool loop must continue whenever stop_reason is tool_use, returning matching tool_result blocks. Treating a tool_use response as the final answer breaks the execution chain and causes the agent to miss required steps.

Correct answer: C. Start Claude Code in plan mode, have it inspect the affected areas, review the proposed plan, then approve an editing mode.

Plan mode lets Claude research dependencies and propose a design before modifying source files. For risky migrations, starting in an editing mode like acceptEdits often causes broad, unintended rewrites that require costly rework.

Correct answer: B. Add an application-level gate that blocks submit_record unless validate_record has succeeded for the same record.

Adding an application-level gate enforces the required validation step before submission, making the workflow deterministic. Relying on system instructions is unreliable because the model might still attempt to guess values and submit prematurely.

Correct answer: A. Define a strict client tool for the resolution record with an input_schema and force that tool with tool_choice.

Strict tool use constrains the output to the declared schema, guaranteeing valid JSON. Forcing the tool call via tool_choice prevents Claude from falling back to free text, which easily breaks parsing when models add conversational prose around raw JSON payloads.

Correct answer: B. Route follow-up work through the coordinator and require subagents to return structured final results for any additional delegation.

The coordinator pattern keeps the parent agent as the single control point for task routing and error handling. Letting subagents message each other directly creates circular dependencies and hidden state changes that make debugging and provenance tracking nearly impossible.

Correct answer: B. Route each request through a coordinator step that selects only the specialists whose descriptions match the current issue.

A routing coordinator analyzes the request and delegates only to relevant specialists. This prevents context window pollution from unnecessary account-recovery details while maintaining the separation of concerns needed to resolve ambiguous billing issues accurately.

Correct answer: B. Delegate the exploratory search to the built-in Explore subagent with explicit scope and conventions, then implement from its summarized findings.

The Explore subagent runs in a separate context window to handle verbose search operations. It returns a concise summary to the parent conversation, preserving the main context window for the actual implementation work and preventing token limits from losing target schemas.

Correct answer: A. Route the nightly archive reprocessing through the Message Batches API, and keep interactive validation on synchronous Messages API calls.

The Message Batches API provides significant cost savings for asynchronous workloads that can tolerate delayed processing. Keeping interactive validation on synchronous API calls ensures users receive immediate validation feedback without waiting for batch processing windows.

Correct answer: A. Keep shared conventions in the root CLAUDE.md and let each package owner add @ imports for only that package's relevant standards in its package CLAUDE.md.

Layering root instructions with package-level configurations uses the import feature to load only relevant domain standards automatically. When Claude Code starts in a package directory, it reads the local files and ancestors without loading sibling package configurations, preventing context bloat.

Correct answer: B. Add allowed-tools: Read Grep Glob to the skill's SKILL.md frontmatter.

The allowed-tools field in the SKILL.md frontmatter enforces hard execution limits at the system level. Relying on prompt instructions to avoid edits is a soft constraint, whereas explicitly listing read-only tools guarantees the skill cannot modify files or run shell commands.

Correct answer: D. Require field-level confidence outputs and calibrate review thresholds using labeled validation examples.

Field-level confidence allows routing only the riskiest parts of a generated change to human reviewers. Aggregate acceptance rates hide specific failures, so you must calibrate against a labeled validation set to turn raw scores into a reliable gating signal.

Correct answer: D. Update the tool and field descriptions with accepted input formats, example queries, edge cases, and usage boundaries.

Tool descriptions and field descriptions are the strongest levers for helping Claude choose and call tools reliably. Forcing tool use or routing more traffic would only increase incorrect calls, so you must clarify acceptable inputs and boundaries directly in the tool metadata.

Correct answer: C. Run /memory in each affected session and compare the listed CLAUDE.md, CLAUDE.local.md, user-level, and managed policy memory files.

Running the /memory command explicitly lists all active CLAUDE.md and policy files loaded for the current working directory. If the package-specific file is missing from this list, the model cannot see those instructions, which immediately explains the inconsistent behavior.

Correct answer: A. Restrict each agent to role-specific tools and route extra lookups through the coordinator.

Restricting each subagent to role-specific tools directly enforces boundaries, preventing the synthesis agent from executing exploratory web searches. If additional lookups are genuinely required, the coordinator can delegate that work to a designated research agent with the appropriate tool access.

Correct answer: C. Ask Claude Code to use AskUserQuestion to gather requirements about implementation constraints, edge cases, and failure modes before planning code changes.

Using an interview pattern to gather requirements leverages Claude to surface hidden design considerations before any code is written. This is critical for complex features involving cache invalidation, ensuring the model understands failure modes and escalation rules prior to planning.

Correct answer: A. Return MCP error results with isError set and structured content that distinguishes protected-branch failures from temporary service failures.

Returning structured MCP error results with the isError flag provides the model with actionable context about the failure type. A protected branch requires a workflow change, so the model stops retrying, whereas a temporary timeout can still trigger a safe retry attempt.

Correct answer: C. Use Grep with the helper name as pattern, scope with glob: "**/*.ts", and set output_mode: "content" for matched lines.

The Grep tool is designed specifically to search file contents and locate call sites within a codebase. By scoping the search with a TypeScript glob pattern and setting output mode to content, the command returns the exact matching lines and file paths needed for navigation.

Correct answer: C. Invoke Claude Code with claude -p "…" or claude –print "…" in the job.

The print flag configures Claude Code to run in a non-interactive mode explicitly designed for CI pipelines and automated scripts. This ensures the process executes the prompt, writes the output, and terminates cleanly without waiting for interactive human input.

Correct answer: A. Return structured subagent results with error codes, attempted queries, and any available source metadata.

Returning structured results with error codes and attempted queries preserves the context the coordinator needs to choose an appropriate recovery path. A generic failure message collapses actionable scenarios into a single vague state, removing the provenance required for downstream reasoning.

Correct answer: D. Configure a PostToolUse hook matching Bash that returns updatedToolOutput with the Bash result shape.

A PostToolUse hook runs after execution but before the model processes the result, allowing you to intercept and normalize the Bash output. You must return the updatedToolOutput object matching the expected structured Bash result shape so the replacement integrates seamlessly into the context.

Correct answer: A. Maintain a structured research state file with the discovered taxonomy and source mappings, then load it into later synthesis prompts.

Maintaining a structured research state file preserves critical early findings explicitly, preventing context degradation over long agent runs. Relying on prompt caching or sending the full transcript does not guarantee the model will recall specific details buried deep in the history.

Correct answer: A. Run section-level extraction passes, then run a reconciliation pass that validates schema, deduplicates entities, and checks cross-section consistency.

A fixed sequential pipeline using section-level extraction reduces attention dilution before a global reconciliation pass resolves duplicates and cross-section relationships. Single-pass extraction on massive documents often misses global context, while dynamic agent loops introduce unpredictable latency and cost.

Correct answer: A. Create a representative evaluation set, refine the templated prompt against clear criteria, and rerun it before submitting the full batch.

Building a representative evaluation set and refining the prompt against clear criteria establishes a necessary evaluation loop before scaling batch operations. Relying on automated prompt improvers or processing bad rows blindly wastes asynchronous budget and hides systematic logic failures.

Correct answer: D. Run a separate verification pass that returns schema-constrained findings with confidence, severity, evidence, and routing fields.

A separate verification pass preserves recall during the initial finding stage while adding structured confidence and severity fields for routing. Relying on free-form prose or non-existent API confidence scores prevents reliable, machine-readable downstream automation.

Correct answer: D. Return the funding source as unknown and include an evidence-gap note when no citable source passage supports it.

Returning an unknown value with an evidence-gap note safely handles missing evidence without triggering hallucinations from endless retries. Retries cannot generate missing citations, and document metadata fields are not valid citable sources for grounded claims.

Correct answer: A. Set tool_choice to {"type": "tool", "name": "extract_metadata"} for the first document-analysis request.

Setting tool_choice to a specific tool type explicitly forces the model to call the named tool immediately. Any other tool_choice setting leaves the selection up to the model's discretion, failing the strict sequential workflow requirement.

Correct answer: A. Append the assistant tool_use response, then send a user message containing the matching tool_result block.

The model needs the raw tool output mapped directly back via a user message containing the matching tool_result block. Passing plain text summaries or logging the result externally breaks the tool use loop, leaving the model blind to the actual data.

Correct answer: C. Maintain a structured project facts file with key decisions, constraints, and open questions, and have Claude Code update and reference it during the session.

Maintaining a structured project facts file preserves critical architectural constraints outside the degrading context window of a long transcript. Prompt caching optimizes cost and latency but does not prevent the model from forgetting earlier instructions as the conversation grows.

Correct answer: C. Run a stop_reason driven loop with tool_choice: {"type":"auto"} and let Claude choose tool use or a final reply each turn.

Using a stop_reason driven loop with automatic tool choice lets Claude dynamically select the necessary tools based on the current context. Predefined chains fail because they cannot adapt to the varying requirements of individual billing disputes.

Correct answer: B. Add unclear and other enum values, plus an other_detail field, and instruct when to use each.

Adding fallback categories like unclear and other with a detail field provides valid choices for ambiguous or unanticipated inputs. This maintains the integrity of your structured schema without forcing the model into incorrect categorizations.

Correct answer: C. Enter plan mode so Claude Code researches the codebase and proposes changes for approval before editing files.

Plan mode allows Claude Code to research the unfamiliar codebase and propose a structural plan for engineer approval before editing. Direct execution is a poor choice here because large architectural refactors require human review to validate the approach first.

Correct answer: C. Add 3-5 XML-wrapped examples that show relevant claim-evidence cases, edge cases, and ideal report phrasing.

Adding structured few-shot examples directly demonstrates the intended judgment boundary for nuanced phrasing decisions. This targeted approach fixes reasoning inconsistencies better than simply adjusting instructions or enforcing structural validation rules.

Correct answer: B. Run query() with resume=baseline_session_id and fork_session=True for each strategy, then store each new session_id.

Using the fork session parameter creates a new branch of the original history for independent exploration. This ensures the baseline session remains completely unchanged while generating two distinct paths you can resume later.

Correct answer: D. Run a separate Claude session with fresh context and provide the requirements, diff, and review checklist.

A separate reviewer session with fresh context provides an unbiased audit free from the original generator rationalizations. The strongest distractor fails because compacting history does not eliminate the self-validation bias inherent to self review.

Correct answer: D. Return no-match lookups as successful empty results, and return timeout failures as error tool results with retry context.

Successful empty results represent valid negative evidence while timeouts represent missing evidence that requires explicit error context. Collapsing both into empty arrays causes the model to assume the data is absent rather than unreachable.

Correct answer: C. Run focused per-file review passes first, then run a separate integration pass over the summarized local findings and cross-file data flow.

Focused per-file reviews prevent attention dilution before a final integration pass checks cross-file data flow. Reviewing everything at once causes context overload, making the model miss localized defects while trying to track global dependencies.

Correct answer: C. Maintain a structured case facts block that is updated after each relevant turn and made available on every relevant turn.

Maintaining a structured case facts block preserves critical transactional data outside the summarized conversation history. Relying on conversational memory or frequent summarization risks data loss, so exact state must live in a durable structured layer.

Correct answer: D. Condense prior exploration into a source-linked summary and include it in each next subagent's delegation prompt.

Condensing prior exploration into a source-linked summary gives the next subagent the relevant state without carrying verbose artifacts forward. Attaching full transcripts wastes context, while keeping delegations independent causes duplicated effort.

Correct answer: B. Create .claude/rules/testing.md with paths frontmatter matching the test file patterns, then place the testing guidance below it.

Path-scoped rule files apply instructions only to matching files using glob patterns in the frontmatter. Duplicating guidance across multiple directories creates maintenance overhead, and root instructions apply too broadly.

Correct answer: D. Rewrite the system prompt and tool description with narrower use boundaries and ordinary wording for audit reports.

Rewriting the system prompt and tool descriptions removes the forceful language causing the tool to trigger on common words. Tool choice configuration cannot fix bad trigger conditions, so defining clear boundaries is the highest-leverage fix.

Correct answer: C. Require each research subagent to return schema-conformant JSON that includes source dates and a date basis for every substantive finding.

Subagents operate in isolated contexts and return only their final output to the parent. Reverting to schema JSON with explicit source dates ensures the parent agent receives the exact metadata needed to distinguish between current information and outdated compatibility notes.

Correct answer: B. Define all three servers in mcp_servers, add one matching mcp_toolset per server in tools, and let Claude choose among the discovered tools.

The Messages API requires defining each server connection in the servers block and enabling its tools through an MCP toolset. This allows Claude to dynamically choose from all available tools during a single request rather than requiring rigid pre-selection per document.

Correct answer: D. Measure precision and recall by finding category, then narrow or pause the noisy dead-code checks.

Measuring precision and recall by finding category identifies the specific class causing the false positives. Narrowing the noisy dead-code checks prevents developers from completely ignoring the automated review output, preserving trust in the highly accurate defect findings.

Correct answer: C. Append the assistant tool use turn, then add an immediate user turn with the matching tool result block.

The application must preserve the assistant message that requested the tool, then send the result back as a user message. The tool result must reference the original tool use ID so the model can connect the output to the requested action.

Correct answer: B. Use a maintained Jira MCP server for standard issue operations, and build a focused custom MCP server for the proprietary workflow.