PART II: EXAM DOMAIN NOTES

Domain 1: Agent Architecture and Orchestration (27%)

1.1 Designing Agentic Loops for Autonomous Task Execution

Key knowledge:

• Agent loop lifecycle: send a Claude request, check stop_reason ("tool_use" vs "end_turn"), execute tools, return results for the next iteration
• Tool results are appended to the conversation history so the model can decide the next action
• Model-driven decision making (Claude chooses the next tool) vs hard-coded decision trees

Key skills:

• Flow control: continue the loop when stop_reason = "tool_use" and stop on "end_turn"
• Appending tool results to context between iterations
• Anti-patterns to avoid: parsing assistant text for completion, using arbitrary iteration limits as the primary stopping mechanism

1.2 Orchestrating Multi-agent Systems (Coordinator–Subagent)

Key knowledge:

• Hub-and-spoke architecture: the coordinator owns all inter-agent communication, error handling, and routing
• Subagents operate with isolated context—they do not automatically inherit the coordinator’s history
• Coordinator responsibilities: task decomposition, delegation, result aggregation, dynamic selection of subagents
• Risk of overly narrow decomposition by the coordinator

Key skills:

• Split research coverage among subagents to minimize duplication
• Implement iterative refinement loops (coordinator evaluates synthesis and re-routes tasks)
• Route all communication through the coordinator for observability

1.3 Configuring Subagent Calls, Context Passing, and Spawning

Key knowledge:

• Task tool spawns subagents; the coordinator’s allowedTools must include "Task"
• Subagent context must be explicitly included in the prompt; subagents do not inherit parent context
• AgentDefinition configuration: descriptions, system prompts, tool constraints
• Session management via fork_session for exploring alternatives

Key skills:

• Include full outputs from prior agents in the subagent prompt
• Use structured formats to separate data from metadata when passing context
• Spawn parallel subagents via multiple Task calls in a single coordinator turn
• Write coordinator prompts in terms of goals and quality criteria rather than step-by-step instructions

1.4 Implementing Multi-step Workflows with Enforcement and Handoff Patterns

Key knowledge:

• The difference between programmatic enforcement (hooks, preconditions) and prompt guidance for ordering a workflow
• When you need deterministic guarantees (e.g., identity verification before financial operations), prompts alone are insufficient
• Structured handoff protocols during escalation (customer ID, reason, recommended action)

Key skills:

• Programmatic preconditions that block downstream calls until prior steps are complete (e.g., block process_refund until get_customer returns a verified ID)
• Decompose multi-aspect customer requests into separate items
• Produce structured summaries when escalating to a human

1.5 Agent SDK Hooks for Intercepting Tool Calls and Normalizing Data

Key knowledge:

• Hook patterns (e.g., PostToolUse) to intercept tool results before the model consumes them
• Hooks that intercept outgoing calls to enforce compliance rules (e.g., block refunds above a threshold)
• Hooks provide deterministic guarantees vs prompt instructions that provide probabilistic compliance

Key skills:

• PostToolUse hooks for normalizing data formats (Unix timestamps, ISO 8601, numeric status codes)
• Interception hooks to block policy-violating actions with redirection to escalation
• Choose hooks over prompts when business rules require guaranteed compliance

1.6 Task Decomposition Strategies for Complex Workflows

Key knowledge:

• Fixed pipelines (prompt chaining) vs dynamic adaptive decomposition based on intermediate results
• Prompt chaining: sequential steps (analyze each file separately, then run an integration pass)
• Adaptive investigation plans that generate subtasks based on what was discovered

Key skills:

• Use prompt chaining for predictable multi-aspect reviews; use dynamic decomposition for open-ended investigations
• Split large code reviews into per-file analysis plus a separate cross-file integration pass
• Decompose open-ended tasks: map structure first, then build a prioritized plan

1.7 Session State, Resuming, and Forking

Key knowledge:

• --resume <session-name> to continue named sessions
• fork_session to create independent investigation branches from shared context
• The importance of informing the agent about file changes when resuming sessions
• A new session with a structured summary can be more reliable than resuming with stale results

Key skills:

• Use --resume to continue named investigation sessions
• Use fork_session to compare approaches in parallel
• Choose between resuming (context still current) vs starting a new session (results stale)

Domain 2: Tool Design and MCP Integration (18%)

2.1 Designing Tool Interfaces with Clear Descriptions

Key knowledge:

• Tool descriptions are the primary mechanism an LLM uses to select tools; minimal descriptions lead to unreliable selection
• The importance of including input formats, example queries, edge cases, and applicability boundaries
• Ambiguous or overlapping descriptions cause misrouting
• System prompt wording can create unintended associations with tools

Key skills:

• Write descriptions that clearly distinguish each tool from similar alternatives
• Rename tools to eliminate functional overlap (e.g., analyze_content -> extract_web_results)
• Split general-purpose tools into specialized ones with clear input/output contracts

2.2 Implementing Structured Error Responses for MCP Tools

Key knowledge:

• The isError flag in MCP tool responses
• The difference between transient errors (timeouts), validation errors (bad input), business errors (policy violations), and access/permission errors
• Generic errors ("Operation failed") prevent correct recovery decisions
• The difference between retryable and non-retryable errors

Key skills:

• Return structured metadata such as errorCategory (transient/validation/permission), isRetryable, and a human-readable message
• Use retryable: false for business-rule violations with clear user-facing explanations
• Do local recovery inside subagents for transient failures; propagate only errors they cannot resolve
• Distinguish access failures (retry decision) from valid empty results (no matches)

2.3 Allocating Tools Across Agents and Configuring tool_choice

Key knowledge:

• Too many tools per agent (e.g., 18 instead of 4–5) reduces tool selection reliability
• Agents with tools outside their specialization tend to misuse them
• Scoped tool access: only role-relevant tools plus a limited set of cross-role utilities
• tool_choice: "auto", "any", and forced tool selection ({"type": "tool", "name": "..."})

Key skills:

• Restrict each subagent’s toolset to what is relevant for its role
• Replace general tools with constrained alternatives (e.g., fetch_url -> load_document)
• Use tool_choice: "any" to guarantee a tool call instead of a text answer
• Force a specific tool to ensure execution order

2.4 Integrating MCP Servers into Claude Code and Agent Workflows

Key knowledge:

• MCP server scope: project (.mcp.json) for teams vs user (~/.claude.json) for experiments
• Environment variable substitution in .mcp.json (e.g., ${GITHUB_TOKEN}) for secret management
• Tools from all connected MCP servers are discovered on connection and are available simultaneously
• MCP resources as “content catalogs” (task summaries, database schemas) to reduce exploratory tool calls

Key skills:

• Configure shared MCP servers in project .mcp.json with env-var-based tokens
• Keep personal/experimental servers in ~/.claude.json
• Prefer community MCP servers over custom servers for standard integrations

2.5 Selecting and Applying Built-in Tools (Read, Write, Edit, Bash, Grep, Glob)

Key knowledge:

• Grep: search within file contents (function names, error messages, imports)
• Glob: find files by name/extension patterns
• Read/Write: full-file operations; Edit: precise changes via unique text matches
• If Edit fails due to non-unique matches, fall back to Read + Write

Key skills:

• Use Grep for content search and Glob for file discovery by patterns
• Build understanding incrementally: Grep entry points, then Read to trace flows
• Trace function usage through wrapper modules

Domain 3: Claude Code Configuration and Workflows (20%)

3.1 Configuring CLAUDE.md with Hierarchy, Scope, and Modular Organization

Key knowledge:

• CLAUDE.md hierarchy: user (~/.claude/CLAUDE.md), project (.claude/CLAUDE.md or root CLAUDE.md), and directory-level (CLAUDE.md in subdirectories)
• User-level settings apply only to one user and are not shared via VCS
• @path syntax for referencing external files (e.g., @./standards/coding-style.md) to modularize CLAUDE.md
• The .claude/rules/ directory for topic-focused rule files instead of a monolithic CLAUDE.md

Key skills:

• Diagnose hierarchy issues (a new team member misses instructions because they are user-level instead of project-level)
• Use @path (e.g., @./standards/testing.md) to selectively include standards in each package’s CLAUDE.md
• Split large CLAUDE.md into multiple .claude/rules/ files (testing.md, api-conventions.md, deployment.md)

3.2 Creating and Configuring Custom Slash Commands and Skills

Key knowledge:

• Project commands in .claude/commands/ (shared via VCS) vs user commands in ~/.claude/commands/
• Skills in .claude/skills/ with SKILL.md frontmatter: context: fork, allowed-tools, argument-hint
• context: fork runs the skill in an isolated subagent context so it does not pollute the main session
• Personal skill variants can live in ~/.claude/skills/ under different names

Key skills:

• Store project slash commands in .claude/commands/ so the whole team gets them
• Use context: fork to isolate skills with verbose output
• Use allowed-tools to restrict what tools a skill can use
• Use argument-hint to prompt developers for required parameters

3.3 Using Path-specific Rules for Conditional Convention Loading

Key knowledge:

• .claude/rules/ files can include YAML frontmatter paths to activate rules based on glob patterns
• Path-scoped rules load only when editing matching files, saving context and tokens
• Glob-based path rules can be preferable to directory-level CLAUDE.md when conventions apply across many directories (e.g., tests)

Key skills:

• Create .claude/rules/ files with paths: ["terraform/**/*"] to load only when working on matching files
• Use glob patterns (**/*.test.tsx) to apply conventions by file type regardless of location
• Prefer path-specific rules over directory-level CLAUDE.md when conventions span the codebase

3.4 Deciding When to Use Planning Mode vs Direct Execution

Key knowledge:

• Planning mode: for complex tasks with large changes, multiple viable approaches, and architectural decisions
• Direct execution: for simple, well-understood changes (e.g., adding a single validation)
• Planning mode enables safe exploration of the codebase before making changes
• Explore subagent isolates verbose discovery output

Key skills:

• Use planning mode for tasks with architectural consequences (microservices, migrations touching 45+ files)
• Use direct execution for fixes with a clear stack trace and a single file
• Use Explore subagent to prevent context-window exhaustion in multi-phase tasks
• Combine approaches: plan for discovery, then execute for implementation

3.5 Iterative Refinement for Progressive Improvement

Key knowledge:

• Concrete input/output examples are the most effective way to communicate expectations
• Test-driven iteration: write tests first, then iterate based on failures
• The “interview” pattern: Claude asks questions to surface non-obvious design considerations
• When to provide all issues in one message (interdependent) vs sequentially (independent)

Key skills:

• Provide 2–3 concrete input/output examples to clarify transformation requirements
• Build test sets with expected behavior, edge cases, and performance requirements before implementation
• Use the interview pattern to surface design aspects (cache invalidation, failure modes)
• Provide concrete test cases with sample inputs and expected outputs for edge cases

3.6 Integrating Claude Code into CI/CD Pipelines

Key knowledge:

• The -p (or --print) flag for non-interactive mode in automated pipelines
• --output-format json and --json-schema for structured output in CI
• CLAUDE.md provides project context (testing standards, review criteria) for CI-triggered Claude Code
• Session context isolation: the same session that generated code is less effective at reviewing it than an independent instance

Key skills:

• Run Claude Code in CI with -p to avoid hanging on interactive input
• Use --output-format json + --json-schema for structured results (e.g., inline PR comments)
• Include prior review results when re-running after new commits (report only new/unfixed issues)
• Document testing standards and available fixtures in CLAUDE.md to improve test generation quality
• Include existing test files in context when generating new tests to avoid duplication and keep style consistent

Domain 4: Prompt Engineering and Structured Output (20%)

4.1 Designing Prompts with Explicit Criteria to Improve Accuracy

Key knowledge:

• Explicit criteria are more effective than vague instructions (e.g., “flag comments only when they contradict code” vs “check comment accuracy”)
• Generic guidance like “be more conservative” works worse than concrete categorical criteria
• The effect of false positives on developer trust: high false-positive rates in some categories undermine trust in accurate categories

Key skills:

• Define review criteria: what to report (bugs, security) vs what to ignore (minor style)
• Temporarily disable categories with high false-positive rates
• Define explicit severity criteria with code examples for each level

4.2 Using Few-shot Prompting to Improve Output Consistency

Key knowledge:

• Few-shot examples are the most effective method for producing consistently formatted, actionable output
• Few-shot can demonstrate handling of ambiguous cases (tool selection, gaps in test coverage)
• Few-shot helps the model generalize to new patterns rather than just repeating defaults
• Few-shot can reduce hallucinations in extraction tasks

Key skills:

• Provide 2–4 targeted examples for ambiguous scenarios with rationale
• Include few-shot examples that demonstrate the output format (location, issue, severity, suggested fix)
• Provide examples that distinguish acceptable code patterns from real issues
• Provide examples of correct extraction from documents with different structures

4.3 Enforcing Structured Output with tool_use and JSON Schemas

Key knowledge:

• tool_use with JSON Schemas is the most reliable way to guarantee schema-conformant output and eliminate JSON syntax errors
• With tool_choice: "auto" the model can return text; with "any" it must call a tool; forced selection chooses a specific tool
• Strict JSON Schemas eliminate syntax errors but do not prevent semantic errors (totals don’t add up; values in wrong fields)
• Schema design: required vs optional fields; enums with “other” plus a detail string for extensibility

Key skills:

• Define extraction tools with JSON Schemas and parse data from tool_use results
• Use tool_choice: "any" to guarantee structured output when multiple schemas exist
• Force a specific tool call: tool_choice: {"type": "tool", "name": "extract_metadata"}
• Make fields optional/nullable when the source may not contain information to avoid fabricating values
• Use enum values like "unclear" and "other" plus detail fields for extensible categorization

4.4 Implementing Validation, Retries, and Feedback Loops for Extraction Quality

Key knowledge:

• Retry-with-error-feedback: include concrete validation errors in the retry prompt to guide corrections
• Retries are ineffective when the information is simply absent from the source
• Feedback loop design: track the pattern that triggered a finding (detected_pattern)
• Semantic errors (totals don’t reconcile) vs syntax errors (addressed by tool_use)

Key skills:

• Follow-up prompts with the original document, an incorrect extraction, and specific validation errors
• Identify when retry will be ineffective (the required info is only in an external document)
• Include detected_pattern fields in findings to analyze false positives
• Design self-correction by extracting both calculated_total and stated_total to detect discrepancies

4.5 Designing Efficient Batch Processing Strategies

Key knowledge:

• Message Batches API: 50% savings, up to 24-hour processing window, no latency SLA guarantees
• Batch processing is suitable for non-blocking tasks (overnight reports, audits) and not suitable for blocking tasks (pre-merge checks)
• Batch API does not support multi-turn tool calling within a single request
• custom_id fields correlate request/response within batches

Key skills:

• Use synchronous API for blocking checks; use Batch API for overnight/weekly workloads
• Plan batch submission cadence based on SLA needs (e.g., 4-hour windows for a 30-hour guarantee with 24-hour processing)
• Handle failures by re-submitting only failed documents (identified by custom_id)
• Iterate on prompts using a sample before running large-scale processing

4.6 Designing Multi-instance and Multi-pass Review Architectures

Key knowledge:

• Self-review limitations: the model retains its reasoning context and is less likely to challenge its own decisions
• Independent review instances (without generation context) are better at finding subtle issues
• Multi-pass review: per-file local analysis plus a cross-file integration pass to avoid attention dilution

Key skills:

• Use a second independent Claude instance to review changes without generation context
• Split multi-file reviews into per-file passes plus integration passes for cross-file dataflow analysis
• Use verification passes with self-rated confidence to route reviews in a calibrated way

Domain 5: Context Management and Reliability (15%)

5.1 Managing Conversation Context to Preserve Critical Information

Key knowledge:

• Risks of progressive summarization: numeric values, percentages, and dates get condensed into vague summaries
• Lost-in-the-middle effect: models reliably process the start and end of long inputs, but may miss findings from the middle
• Tool outputs can accumulate in context disproportionately to relevance (40+ fields when 5 are needed)
• The importance of sending the full conversation history in subsequent API requests

Key skills:

• Extract transactional facts into a persistent “case facts” block outside the summarized history
• Trim verbose tool outputs down to relevant fields
• Place key findings at the beginning of aggregated data with explicit section headings
• Require subagents to include metadata (dates, sources) in structured outputs

5.2 Designing Effective Escalation Patterns and Resolving Ambiguity

Key knowledge:

• Suitable escalation triggers: explicit request for a human, policy gaps/exceptions, inability to make progress
• Immediate escalation (explicit request) vs attempt-to-resolve (within agent scope)
• Sentiment analysis and model confidence self-ratings are unreliable proxies for case complexity
• Multiple customer matches require asking for additional identifiers, not heuristic guessing

Key skills:

• Explicit escalation criteria with few-shot examples in the system prompt
• Execute explicit requests for a human immediately without additional investigation
• Escalate when policy is ambiguous or silent for a specific request
• Ask for additional identifiers when tool results contain multiple matches

5.3 Implementing Error Propagation Strategies in Multi-agent Systems

Key knowledge:

• Structured error context (failure type, query, partial results, alternatives) enables smarter coordinator recovery
• Distinguish access failures (timeouts require a retry decision) from valid empty results (no matches)
• Generic error statuses (“search unavailable”) hide valuable context from the coordinator
• Silent suppression or aborting the whole workflow on a single failure are both anti-patterns

Key skills:

• Return structured error context: failure type, what was attempted, partial results, possible alternatives
• Distinguish access failures from valid empty results
• Perform local recovery in subagents for transient failures; propagate only non-recoverable errors with partial results
• Annotate coverage in synthesis: what is well-supported vs where gaps remain

5.4 Managing Context Efficiently When Investigating Large Codebases

Key knowledge:

• Context degradation in long sessions: the model starts producing unstable answers and referring to “typical patterns” instead of specific classes
• Scratchpad files preserve key findings across context boundaries
• Delegating to subagents isolates verbose discovery output
• Structured state persistence enables crash recovery

Key skills:

• Spawn subagents for specific questions while keeping high-level coordination in the main agent
• Use scratchpad files to store key findings and reference them later
• Summarize key findings before spawning next-phase subagents
• Use /compact to reduce context usage during long investigations

5.5 Designing Workflows with Human Oversight and Confidence Calibration

Key knowledge:

• Aggregate metrics (e.g., 97% overall accuracy) can mask poor performance on specific document types or fields
• Stratified random sampling measures error rates in high-confidence extractions
• Field-level confidence calibration using labeled validation sets
• Validate accuracy by document type and field segment before automating

Key skills:

• Implement stratified random sampling to detect new error patterns
• Analyze accuracy by document type and field to validate stable performance
• Output field-level confidence scores and calibrate review thresholds using labeled data
• Route low-confidence or ambiguous-source extractions to human review

5.6 Preserving Provenance and Handling Uncertainty in Multi-source Synthesis

Key knowledge:

• Attribution is lost during summarization without preserving “claim → source” mappings
• Structured mappings must be preserved during aggregation
• Handle conflicting statistics by annotating conflicts with attribution rather than arbitrarily choosing one value
• Include publication/collection dates to avoid misreading temporal differences as contradictions

Key skills:

• Require subagents to output “claim → source” mappings (URL, document name, quotes)
• Structure reports to separate stable findings from disputed ones
• Preserve conflicting values with annotations and pass them to the coordinator for reconciliation
• Include publication dates for correct temporal interpretation
• Render content by type: financial data as tables, news as prose, technical findings as structured lists

Examples of Exam Questions with Explanations

Question 1 (Scenario: Customer Support Agent)

Situation: Data shows that in 12% of cases the agent skips get_customer and calls lookup_order using only the customer’s name, which leads to incorrect refunds.

Which change is most effective?

• A) Add a programmatic precondition that blocks lookup_order and process_refund until an ID is obtained from get_customer [CORRECT]
• B) Improve the system prompt
• C) Add few-shot examples
• D) Implement a routing classifier

Why A: When critical business logic requires a specific tool sequence, software provides deterministic guarantees that prompt-based approaches (B, C) cannot. D addresses availability, not tool ordering.

Question 2 (Scenario: Customer Support Agent)

Situation: The agent often calls get_customer instead of lookup_order for order-related questions. Tool descriptions are minimal and similar.

What is the first step?

• A) Few-shot examples
• B) Expand each tool’s description with input formats, examples, and boundaries [CORRECT]
• C) Add a routing layer
• D) Merge the tools

Why B: Tool descriptions are the model’s primary selection mechanism. This is the lowest-effort, highest-impact fix. A adds tokens without addressing the root cause. C is overengineering. D requires more effort than justified.

Question 3 (Scenario: Customer Support Agent)

Situation: The agent resolves only 55% of issues with a target of 80%. It escalates simple cases and tries to handle complex policy exceptions autonomously.

How do you improve calibration?

• A) Add explicit escalation criteria with few-shot examples [CORRECT]
• B) Self-rated confidence (1–10) with automatic escalation
• C) A separate classifier trained on historical data
• D) Sentiment analysis

Why A: It directly addresses the root cause—unclear decision boundaries. B is unreliable (the model can be confidently wrong). C is overengineering. D solves a different problem (mood != complexity).

Question 4 (Scenario: Code Generation with Claude Code)

Situation: You need a custom /review command for standard code review that is available to the whole team when they clone the repository.

Where should you create the command file?

• A) .claude/commands/ in the project repository [CORRECT]
• B) ~/.claude/commands/
• C) Root CLAUDE.md
• D) .claude/config.json

Why A: Project commands stored in .claude/commands/ are version-controlled and automatically available to everyone. B is for personal commands. C is for instructions, not command definitions. D does not exist.

Question 5 (Scenario: Code Generation with Claude Code)

Situation: You need to restructure a monolith into microservices (dozens of files, service-boundary decisions).

What approach should you use?

• A) Planning mode: explore the codebase, understand dependencies, design an approach [CORRECT]
• B) Direct execution incrementally
• C) Direct execution with detailed up-front instructions
• D) Direct execution and switch to planning when it gets hard

Why A: Planning mode is designed for large changes, multiple possible approaches, and architectural decisions. B risks expensive rework. C assumes you already know the structure. D is reactive.

Question 6 (Scenario: Code Generation with Claude Code)

Situation: A codebase has different conventions across areas (React, API, database). Tests are co-located with code. You want conventions to be applied automatically.

What approach should you use?

• A) .claude/rules/ files with YAML frontmatter and glob patterns [CORRECT]
• B) Put everything in the root CLAUDE.md
• C) Skills in .claude/skills/
• D) CLAUDE.md in every directory

Why A: .claude/rules/ with glob patterns (e.g., **/*.test.tsx) enables automatic convention application based on file paths—ideal for tests spread across the codebase. B relies on model inference. C is manual/on-demand. D does not work well when relevant files are in many directories.

Question 7 (Scenario: Multi-agent Research System)

Situation: The system researches “AI impact on creative industries,” but reports cover only visual art. The coordinator decomposed the topic into: “AI in digital art,” “AI in graphic design,” “AI in photography.”

What’s the cause?

• A) The synthesis agent does not detect gaps
• B) The coordinator decomposed the task too narrowly [CORRECT]
• C) The web search agent does not search thoroughly enough
• D) The document analysis agent filters out non-visual sources

Why B: The logs show the coordinator decomposed “creative industries” only into visual subtopics, completely missing music, literature, and film. Subagents executed correctly—the issue is what they were assigned.

Question 8 (Scenario: Multi-agent Research System)

Situation: A web-search subagent times out while researching a complex topic. You need to design how error information is passed back to the coordinator.

Which error propagation approach best enables intelligent recovery?

• A) Return structured error context to the coordinator: failure type, query, partial results, and alternatives [CORRECT]
• B) Implement automatic retries with exponential backoff inside the subagent, then return a generic “search unavailable” status
• C) Catch the timeout inside the subagent and return an empty result set marked as success
• D) Propagate the timeout exception to a top-level handler that terminates the whole workflow

Why A: Structured error context gives the coordinator what it needs to decide whether to retry with a modified query, try an alternative approach, or continue with partial results. B hides context behind a generic status. C masks failure as success. D aborts the entire workflow unnecessarily.

Question 9 (Scenario: Multi-agent Research System)

Situation: The synthesis agent often needs to verify specific claims while merging results. Currently, when verification is needed, the synthesis agent hands control back to the coordinator, which calls the web-search agent and then re-runs synthesis with the new results. This adds 2–3 extra round trips per task and increases latency by 40%. Your assessment shows that 85% of these checks are simple fact checks (dates, names, statistics), while 15% require deeper investigation.

How do you reduce overhead while maintaining reliability?

• A) Give the synthesis agent a limited verify_fact tool for simple checks, and continue routing complex verification through the coordinator [CORRECT]
• B) Accumulate all verification needs into a batch and return them to the coordinator at the end
• C) Give the synthesis agent full access to all web-search tools
• D) Proactively cache additional context around each source

Why A: This applies the principle of least privilege: the synthesis agent gets exactly what it needs for the 85% common case (simple fact checks) while preserving the coordinator-mediated path for complex investigations. B introduces blocking dependencies (later synthesis steps may depend on earlier verified facts). C breaks separation of responsibilities. D relies on speculative caching that cannot reliably predict needs.

Question 10 (Scenario: Claude Code for CI)

Situation: A pipeline runs claude "Analyze this pull request for security issues", but hangs waiting for interactive input.

What is the correct approach?

• A) Use the -p flag: claude -p "Analyze this pull request for security issues" [CORRECT]
• B) Set CLAUDE_HEADLESS=true
• C) Redirect stdin from /dev/null
• D) Use --batch

Why A: -p (or --print) is the documented way to run Claude Code in non-interactive mode. It processes the prompt, prints to stdout, and exits. The other options are either non-existent features or Unix workarounds.

Question 11 (Scenario: Claude Code for CI)

Situation: The team wants to reduce API cost for automated analysis. Claude currently serves two workflows in real time: (1) a blocking pre-merge check that must complete before developers can merge a PR, and (2) a tech-debt report generated overnight for morning review. A manager proposes moving both to the Message Batches API to save 50%.

How should you evaluate this proposal?

• A) Use batch processing only for tech-debt reports; keep real-time calls for pre-merge checks [CORRECT]
• B) Move both workflows to batch processing and poll for completion
• C) Keep real-time calls for both to avoid ordering issues in batch results
• D) Move both to batch processing with a fallback to real time if a batch takes too long

Why A: The Message Batches API saves 50%, but processing time can be up to 24 hours with no guaranteed latency SLA. That makes it unsuitable for blocking pre-merge checks where developers are waiting, but ideal for overnight batch workloads like tech-debt reports.

Question 12 (Scenario: Multi-file Code Review)

Situation: A pull request changes 14 files in an inventory tracking module. A single-pass review of all files produces inconsistent results: detailed comments for some files but superficial ones for others, missed obvious bugs, and contradictory feedback (a pattern is flagged as problematic in one file but approved in identical code in another file).

How should you restructure the review?

• A) Split into focused passes: analyze each file individually for local issues, then run a separate integration pass for cross-file data flows [CORRECT]
• B) Require developers to split large PRs into submissions of 3–4 files
• C) Switch to a higher-tier model with a larger context window to review all 14 files in one pass
• D) Run three independent full-PR review passes and report only issues found in at least two runs

Why A: Focused passes directly address the root cause—attention dilution when processing many files at once. Per-file analysis ensures consistent depth, and a separate integration pass catches cross-file issues. B shifts burden to developers without improving the system. C is a misconception: larger context does not fix attention quality. D suppresses real bugs by requiring consensus across inconsistent detections.

Practice Test

60 questions across 4 scenarios. Format and difficulty match the real exam.

Alternatively, you can practice these questions in an exam-like HTML file: Practical Test (EN)

Scenario: Multi-agent Research System

Question 1 (Scenario: Multi-agent Research System)

Situation: A document analysis agent discovers that two credible sources contain directly contradictory statistics for a key metric: a government report states 40% growth, while an industry analysis states 12%. Both sources look credible, and the discrepancy could materially affect the research conclusions. How should the document analysis agent handle this situation most effectively?

Which approach is most effective?

• A) Apply credibility heuristics to pick the most likely correct number, finish analysis with that value, and add a footnote mentioning the discrepancy.
• B) Include both numbers in the analysis output without marking them as conflicting, letting the synthesis agent decide which to use based on broader context.
• C) Stop analysis and immediately escalate to the coordinator, asking it to decide which source is more authoritative before continuing.
• D) Complete analysis with both numbers, explicitly annotate the conflict with source attribution, and let the coordinator decide how to reconcile the data before passing to synthesis. [CORRECT]

Why D: This approach preserves separation of responsibilities: the analysis agent completes its core work without blocking, preserves both conflicting values with clear attribution, and correctly passes reconciliation to the coordinator, which has broader context.

Question 2 (Scenario: Multi-agent Research System)

Situation: The web-search and document-analysis agents have completed their tasks and returned results to the coordinator. What is the next step for creating an integrated research report?

Which next step is most appropriate?

• A) Each agent sends its results directly to the report-writing agent, bypassing the coordinator.
• B) The document analysis agent requests web-search results and merges them internally.
• C) The coordinator passes both sets of results to the synthesis agent for a unified integration. [CORRECT]
• D) The coordinator concatenates the raw outputs from both agents and returns them as the final result.

Why C: In a coordinator–subagent architecture, the coordinator forwards both result sets to the synthesis agent for centralized integration, preserving control and ensuring high-quality merging.

Question 3 (Scenario: Multi-agent Research System)

Situation: A document analysis subagent frequently fails when processing PDF files: some have corrupted sections that trigger parsing exceptions, others are password-protected, and sometimes the parsing library hangs on large files. Currently, any exception immediately terminates the subagent and returns an error to the coordinator, which must decide whether to retry, skip, or fail the whole task. This causes excessive coordinator involvement in routine error handling. What architectural improvement is most effective?

Which improvement is most effective?

• A) Create a dedicated error-handling agent that monitors all failures via a shared queue and decides recovery actions, sending restart commands directly to subagents.
• B) Configure the subagent to always return partial results with a success status, embedding error details in metadata; the coordinator treats all responses as successful.
• C) Make the coordinator validate all documents before sending them to the subagent, rejecting documents that might cause failures.
• D) Implement local recovery in the subagent for transient failures and escalate to the coordinator only errors it cannot resolve, including attempted steps and partial results. [CORRECT]

Why D: Handle errors at the lowest level capable of resolving them. Local recovery reduces coordinator workload while still escalating truly unrecoverable issues with full context and partial progress.

Question 4 (Scenario: Multi-agent Research System)

Situation: After running the system on “AI impact on creative industries,” you observe that every subagent completes successfully: the web-search agent finds relevant articles, the document analysis agent summarizes them correctly, and the synthesis agent produces coherent text. However, final reports cover only visual art and completely miss music, literature, and film. In the coordinator logs, you see it decomposed the topic into three subtasks: “AI in digital art,” “AI in graphic design,” and “AI in photography.” What is the most likely root cause?

What is the most likely root cause?

• A) The synthesis agent lacks instructions to detect coverage gaps.
• B) The document analysis agent filters out non-visual sources due to overly strict relevance criteria.
• C) The coordinator’s task decomposition is too narrow, assigning subagents work that does not cover all relevant areas. [CORRECT]
• D) The web-search agent’s queries are insufficient and should be broadened to cover more sectors.

Why C: The coordinator decomposed a broad topic only into visual-art subtasks, missing music, literature, and film entirely. Since subagents executed their assignments correctly, the narrow decomposition is the obvious root cause.

Question 5 (Scenario: Multi-agent Research System)

Situation: The web-search subagent returns results for only 3 of 5 requested source categories (competitor sites and industry reports succeed, but news archives and social feeds time out). The document analysis subagent successfully processes all provided documents. The synthesis subagent must produce a summary from mixed-quality upstream inputs. Which error-propagation strategy is most effective?

Which error-propagation strategy is most effective?

• A) Continue synthesis using only successful sources and produce an output without mentioning which data was unavailable.
• B) The synthesis subagent returns an error to the coordinator, triggering a full retry or task failure due to incomplete data.
• C) The synthesis subagent asks the coordinator to retry timed-out sources with a longer timeout before starting synthesis.
• D) Structure the synthesis output with coverage annotations that indicate which conclusions are well-supported and where gaps exist due to unavailable sources. [CORRECT]

Why D: Coverage annotations implement graceful degradation with transparency, preserving value from completed work while propagating uncertainty to enable informed decisions about confidence.

Question 6 (Scenario: Multi-agent Research System)

Situation: The document analysis subagent encounters a corrupted PDF file that it cannot parse. When designing the system’s error handling, what is the most effective way to handle this failure?

Which approach is most effective?

• A) Return an error with context to the coordinator agent, allowing it to decide how to proceed. [CORRECT]
• B) Silently skip the corrupted document and continue processing the remaining files to avoid interrupting the workflow.
• C) Automatically retry parsing the document three times with exponential backoff before reporting a failure.
• D) Throw an exception that terminates the entire research workflow.

Why A: Returning an error with context to the coordinator is the most effective approach because it lets the coordinator make an informed decision—skip the file, try an alternative parsing method, or notify the user—while maintaining visibility into the failure.

Question 7 (Scenario: Multi-agent Research System)

Situation: Production logs show a persistent pattern: requests like “analyze the uploaded quarterly report” are routed to the web-search agent 45% of the time instead of the document analysis agent. Reviewing tool definitions, you find that the web-search agent has a tool analyze_content described as “analyzes content and extracts key information,” while the document analysis agent has a tool analyze_document described as “analyzes documents and extracts key information.” How should you fix the misrouting problem?

How should you fix the misrouting problem?

• A) Add a pre-routing classifier that detects whether the user refers to uploaded files or web content before the coordinator decides on delegation.
• B) Rename the web-search tool to extract_web_results and update its description to “processes and returns information retrieved from web search and URLs.” [CORRECT]
• C) Add few-shot examples to the coordinator prompt showing correct routing: “User uploads a quarterly report → document analysis agent” and “User asks about a web page → web-search agent.”
• D) Expand the document analysis tool description with usage examples like “Use for uploaded PDFs, Word docs, and spreadsheets,” leaving the web-search tool unchanged.

Why B: Renaming the web-search tool to extract_web_results and updating its description to explicitly reference web search and URLs directly removes the root cause by eliminating semantic overlap between the two tool names and descriptions. This makes each tool’s purpose unambiguous, enabling the coordinator to reliably distinguish document analysis from web search.

Question 8 (Scenario: Multi-agent Research System)

Situation: A colleague proposes that the document analysis agent should send its results directly to the synthesis agent, bypassing the coordinator. What is the main advantage of keeping the coordinator as the central hub for all communication between subagents?

What is the main advantage of keeping the coordinator as the central hub?

• A) The coordinator can observe all interactions, handle errors uniformly, and decide what information each subagent should receive. [CORRECT]
• B) The coordinator batches multiple requests to subagents, reducing total API calls and overall latency.
• C) Routing through the coordinator enables automatic retry logic that direct inter-agent calls cannot support.
• D) Subagents use isolated memory, and direct communication would require complex serialization that only the coordinator can perform.

Why A: The coordinator pattern provides centralized visibility into all interactions, uniform error handling across the system, and fine-grained control over what information each subagent receives—these are the primary advantages of a star-shaped communication topology.

Question 9 (Scenario: Multi-agent Research System)

Situation: The web-search subagent times out while researching a complex topic. You need to design how information about this failure is returned to the coordinator. Which error-propagation approach best enables intelligent recovery?

Which error-propagation approach best enables intelligent recovery?

• A) Return structured error context to the coordinator including the failure type, the query executed, any partial results, and potential alternative approaches. [CORRECT]
• B) Catch the timeout within the subagent and return an empty result set marked as successful.
• C) Implement automatic exponential-backoff retries inside the subagent, only returning a generic “search unavailable” status after exhausting retries.
• D) Propagate the timeout exception directly to the top-level handler, terminating the entire research workflow.

Why A: Returning structured error context—including failure type, executed query, partial results, and alternative approaches—gives the coordinator everything needed to make intelligent recovery decisions (e.g., retry with a modified query or continue with partial results). It preserves maximum context for informed coordination-level decision-making.

Question 10 (Scenario: Multi-agent Research System)

Situation: In your system design, you gave the document analysis agent access to a general-purpose tool fetch_url so it could download documents by URL. Production logs show this agent now frequently downloads search engine results pages to perform ad hoc web search—behavior that should be routed through the web-search agent—causing inconsistent results. Which fix is most effective?

Which fix is most effective?

• A) Replace fetch_url with a load_document tool that validates that URLs point to document formats. [CORRECT]
• B) Remove fetch_url from the document analysis agent and route all URL fetching through the coordinator to the web-search agent.
• C) Implement filtering that blocks fetch_url calls to known search engine domains while allowing other URLs.
• D) Add instructions to the document analysis agent prompt that fetch_url should only be used to download document URLs, not to search.

Why A: Replacing a general-purpose tool with a document-specific tool that validates URLs against document formats fixes the root cause by constraining capability at the interface level. This follows the principle of least privilege, making undesired search behavior impossible rather than merely discouraged.

Question 11 (Scenario: Multi-agent Research System)

Situation: While researching a broad topic, you observe that the web-search agent and the document analysis agent investigate the same subtopics, leading to substantial duplication in their outputs. Token usage nearly doubles without a proportional increase in research breadth or depth. What is the most effective way to address this?

What is the most effective way to address this?

• A) Allow both agents to finish in parallel, then have the coordinator deduplicate overlapping results before passing them to the synthesis agent.
• B) The coordinator explicitly partitions the research space before delegating, assigning each agent distinct subtopics or source types. [CORRECT]
• C) Implement a shared-state mechanism where agents log their current focus area so other agents can dynamically avoid duplication during execution.
• D) Switch to sequential execution where document analysis runs only after web search completes, using web-search results as context to avoid duplication.

Why B: Having the coordinator explicitly partition the research space before delegating is most effective because it addresses the root cause—unclear task boundaries—before any work begins. It preserves parallelism while preventing duplicated effort and wasted tokens.

Question 12 (Scenario: Multi-agent Research System)

Situation: During research, the web-search subagent queries three source categories with different outcomes: academic databases return 15 relevant papers, industry reports return “0 results,” and patent databases return “Connection timeout.” When designing error propagation to the coordinator, which approach enables the best recovery decisions?

Which approach enables the best recovery decisions?

• A) Aggregate the results into a single success-percentage metric (e.g., “67% source coverage”) with detailed logs available on demand.
• B) Report both “timeout” and “0 results” as failures requiring coordinator intervention.
• C) Retry transient failures internally and report only persistent errors.
• D) Distinguish access failures (timeout) that require a retry decision from valid empty results (“0 results”) that represent successful queries. [CORRECT]

Why D: A timeout (access failure) and “0 results” (valid empty result) are semantically different outcomes requiring different responses. Distinguishing them allows the coordinator to retry the patent database while accepting the industry reports “0 results” as a valid, informative finding.

Question 13 (Scenario: Multi-agent Research System)

Situation: Production monitoring shows inconsistent synthesis quality. When aggregated results are ~75K tokens, the synthesis agent reliably cites information from the first 15K tokens (web-search headlines/snippets) and the last 10K tokens (document analysis conclusions), but often misses critical findings in the middle 50K tokens—even when they directly answer the research question. How should you restructure the aggregated input?

How should you restructure the aggregated input?

• A) Summarize all subagent outputs to under 20K tokens before aggregation to keep content within the model’s reliable processing range.
• B) Stream subagent results to the synthesis agent incrementally, processing web-search results first to completion, then adding document analysis results.
• C) Place a key-findings summary at the start of the aggregated input and organize detailed results with explicit section headings for easier navigation. [CORRECT]
• D) Implement rotation that alternates which subagent’s results appear first across research tasks to ensure both sources get equal top positioning over time.

Why C: Putting a key-findings summary at the start leverages primacy effects so critical information sits in the most reliably processed position. Adding explicit section headings throughout helps the model navigate and attend to mid-input content, directly mitigating the “lost in the middle” phenomenon.

Question 14 (Scenario: Multi-agent Research System)

Situation: In testing, the combined output of the web-search agent (85K tokens including page content) and the document analysis agent (70K tokens including chains of thought) totals 155K tokens, but the synthesis agent performs best with inputs under 50K tokens. Which solution is most effective?

Which solution is most effective?

• A) Modify upstream agents to return structured data (key facts, quotes, relevance scores) instead of verbose content and reasoning. [CORRECT]
• B) Add an intermediate summarization agent that condenses findings before passing them to synthesis.
• C) Have the synthesis agent process findings in sequential batches, maintaining state between calls.
• D) Store findings in a vector database and give the synthesis agent search tools to query during its work.

Why A: Modifying upstream agents to return structured data fixes the root cause by reducing token volume at the source while preserving essential information. It avoids passing bulky page content and reasoning traces that inflate tokens without improving the synthesis step.

Question 15 (Scenario: Multi-agent Research System)

Situation: In testing, you observe that the synthesis agent often needs to verify specific claims while merging results. Currently, when verification is needed, the synthesis agent returns control to the coordinator, which calls the web-search agent and then re-invokes synthesis with the results. This adds 2–3 extra loops per task and increases latency by 40%. Your assessment shows 85% of these verifications are simple fact checks (dates, names, stats) and 15% require deeper research. Which approach most effectively reduces overhead while preserving system reliability?

Which approach is most effective?

• A) Give the synthesis agent access to all web-search tools so it can handle any verification need directly without coordinator loops.
• B) Have the synthesis agent accumulate all verification needs and return them as a batch to the coordinator at the end, which then sends them all to the web-search agent at once.
• C) Have the web-search agent proactively cache extra context around each source during initial research in anticipation of synthesis needing verification.
• D) Give the synthesis agent a limited-scope verify_fact tool for simple checks, while routing complex verifications through the coordinator to the web-search agent. [CORRECT]

Why D: A limited-scope fact-verification tool lets the synthesis agent handle 85% of simple checks directly, eliminating most loops, while preserving the coordinator delegation path for the 15% of complex verifications. This applies least privilege while significantly reducing latency.

Scenario: Claude Code for Continuous Integration

Question 16 (Scenario: Claude Code for Continuous Integration)

Situation: Your CI pipeline runs the Claude Code CLI (in --print mode) using CLAUDE.md to provide project context for code review, and developers generally find the reviews substantive. However, they report that integrating findings into the workflow is difficult—Claude outputs narrative paragraphs that must be manually copied into PR comments. The team wants to automatically post each finding as a separate inline PR comment at the relevant place in code, which requires structured data with file path, line number, severity level, and suggested fix. Which approach is most effective?

Which approach is most effective?

• A) Add an “Output Format for Review” section to CLAUDE.md with examples of structured findings so Claude learns the expected format from project context.
• B) Use the CLI flags --output-format json and --json-schema to enforce structured findings, then parse the output to post inline comments via the GitHub API. [CORRECT]
• C) Include explicit formatting instructions in the review prompt requiring each finding to follow a parseable template like [FILE:path] [LINE:n] [SEVERITY:level] ....
• D) Keep narrative review format but add a summarization step that uses Claude to generate a structured JSON summary of findings.

Why B: Using --output-format json with --json-schema enforces structured output at the CLI level, guaranteeing well-formed JSON with the required fields (file path, line number, severity, suggested fix) that can be reliably parsed and posted as inline PR comments via the GitHub API. It leverages built-in CLI capabilities designed specifically for structured output.

Question 17 (Scenario: Claude Code for Continuous Integration)

Situation: Your team uses Claude Code for generating code suggestions, but you notice a pattern: non-obvious issues—performance optimizations that break edge cases, cleanups that unexpectedly change behavior—are only caught when another team member reviews the PR. Claude’s reasoning during generation shows it considered these cases but concluded its approach was correct. Which approach directly addresses the root cause of this self-check limitation?

Which approach directly addresses the root cause?

• A) Run a second independent instance of Claude Code to review the changes without access to the generator’s reasoning. [CORRECT]
• B) Enable extended thinking mode for the generation stage to allow more thorough deliberation before producing suggestions.
• C) Add explicit self-review instructions to the generation prompt asking Claude to critique its own suggestions before finalizing output.
• D) Include full test files and documentation in prompt context so Claude better understands expected behavior during generation.

Why A: A second independent Claude Code instance without access to the generator’s reasoning directly addresses the root cause by avoiding confirmation bias. This “fresh eyes” perspective mirrors human peer review, where another reviewer catches issues the author rationalized.

Question 18 (Scenario: Claude Code for Continuous Integration)

Situation: Your code review component is iterative: Claude analyzes the changed file, then may request related files (imports, base classes, tests) via tool calls to understand context before providing final feedback. Your application defines a tool that lets Claude request file contents; Claude calls the tool, gets results, and continues analysis. You’re evaluating batch processing to reduce API cost. What is the primary technical limitation when considering batch processing for this workflow?

What is the primary technical limitation?

• A) Batch processing does not include correlation IDs to map outputs back to input requests.
• B) The asynchronous model cannot execute tools mid-request and return results for Claude to continue analysis. [CORRECT]
• C) The Batch API does not support tool definitions in request parameters.
• D) The batch processing latency of up to 24 hours is too slow for pull request feedback, although the workflow would otherwise function.

Why B: A “fire-and-forget” asynchronous Batch API model has no mechanism to intercept a tool call during a request, execute the tool, and return results for Claude to continue analysis. This is fundamentally incompatible with iterative tool-calling workflows that require multiple tool request/response rounds within a single logical interaction.

Question 19 (Scenario: Claude Code for Continuous Integration)

Situation: Your CI/CD system runs three Claude-based analyses: (1) fast style checks on every PR that block merging until completion, (2) comprehensive weekly security audits of the entire codebase, and (3) nightly test-case generation for recently changed modules. The Message Batches API offers 50% savings but processing can take up to 24 hours. You want to optimize API cost while maintaining an acceptable developer experience. Which combination correctly matches each task to an API approach?

Which combination is correct?

• A) Use the Message Batches API for all three tasks to maximize 50% savings, configuring the pipeline to poll for batch completion.
• B) Use synchronous calls for PR style checks; use the Message Batches API for weekly security audits and nightly test generation. [CORRECT]
• C) Use synchronous calls for all three tasks for consistent response times, relying on prompt caching to reduce costs across workloads.
• D) Use synchronous calls for PR style checks and nightly test generation; use the Message Batches API only for weekly security audits.

Why B: PR style checks block developers and require immediate responses via synchronous calls, while weekly security audits and nightly test generation are scheduled tasks with flexible deadlines that can tolerate up to a 24-hour batch window—capturing 50% savings for both.

Question 20 (Scenario: Claude Code for Continuous Integration)

Situation: Your automated reviews find real issues, but developers report the feedback is not actionable. Findings include phrases like “complex ticket routing logic” or “potential null pointer” without specifying what exactly to change. When you add detailed instructions like “always include concrete fix suggestions,” the model still produces inconsistent output—sometimes detailed, sometimes vague. Which prompting technique most reliably produces consistently actionable feedback?

Which prompting technique is most reliable?

• A) Further refine instructions with more explicit requirements for each part of the feedback format (location, issue, severity, proposed fix).
• B) Expand the context window to include more surrounding codebase so the model has enough information to propose concrete fixes.
• C) Implement a two-pass approach where one prompt identifies issues and a second generates fixes, allowing specialization.
• D) Add 3–4 few-shot examples showing the exact required format: identified issue, location in code, concrete fix suggestion. [CORRECT]

Why D: Few-shot examples are the most effective technique for achieving consistent output format when instructions alone produce variable results. Providing 3–4 examples that show the exact desired structure (issue, location, concrete fix) gives the model a concrete pattern to follow, which is more reliable than abstract instructions.

Question 21 (Scenario: Claude Code for Continuous Integration)

Situation: Your CI pipeline includes two Claude-based code review modes: a pre-merge-commit hook that blocks PR merge until completion, and a “deep analysis” that runs overnight, polls for batch completion, and posts detailed suggestions to the PR. You want to reduce API cost using the Message Batches API, which offers 50% savings but requires polling and can take up to 24 hours. Which mode should use batch processing?

Which mode should use batch processing?

• A) Only the pre-merge-commit hook.
• B) Only the deep analysis. [CORRECT]
• C) Both modes.
• D) Neither mode.

Why B: Deep analysis is an ideal candidate for batch processing because it already runs overnight, tolerates delay, and uses a polling model before publishing results—matching the asynchronous, polling-based architecture of the Message Batches API while capturing 50% savings.

Question 22 (Scenario: Claude Code for Continuous Integration)

Situation: Your automated review analyzes comments and docstrings. The current prompt instructs Claude to “check that comments are accurate and up to date.” Findings often flag acceptable patterns (TODO markers, simple descriptions) while missing comments describing behavior the code no longer implements. What change addresses the root cause of this inconsistent analysis?

What change addresses the root cause?

• A) Include git blame data so Claude can identify comments that predate recent code changes.
• B) Add few-shot examples of misleading comments to help the model recognize similar patterns in the codebase.
• C) Filter TODO, FIXME, and descriptive comment patterns before analysis to reduce noise.
• D) Specify explicit criteria: flag comments only when the behavior they claim contradicts the code’s actual behavior. [CORRECT]

Why D: Explicit criteria—flagging comments only when claimed behavior contradicts actual code behavior—directly addresses the root cause by replacing a vague instruction with a precise definition of what constitutes a problem. This reduces false positives on acceptable patterns and misses of truly misleading comments.

Question 23 (Scenario: Claude Code for Continuous Integration)

Situation: Your automated code review system shows inconsistent severity ratings—similar issues like null pointer risks are rated “critical” in some PRs but only “medium” in others. Developer surveys show growing distrust—many start dismissing findings without reading because “half are wrong.” High-false-positive categories erode trust in accurate categories. Which approach best restores developer trust while improving the system?

Which approach best restores developer trust?

• A) Temporarily disable high-false-positive categories (style, naming, documentation) and keep only high-precision categories while improving prompts. [CORRECT]
• B) Keep all categories enabled but display confidence scores with each finding so developers can decide what to investigate.
• C) Keep all categories enabled and add few-shot examples to improve accuracy for each category over the next few weeks.
• D) Apply a uniform strictness reduction across all categories to bring the overall false-positive rate down.

Why A: Temporarily disabling high-false-positive categories immediately stops trust erosion by removing noisy findings that cause developers to dismiss everything, while preserving value from high-precision categories like security and correctness. It also creates space to improve prompts for problematic categories before re-enabling them.

Question 24 (Scenario: Claude Code for Continuous Integration)

Situation: Your automated review generates test-case suggestions for each PR. Reviewing a PR that adds course completion tracking, Claude suggests 10 test cases, but developer feedback shows that 6 duplicate scenarios already covered by the existing test suite. What change most effectively reduces duplicate suggestions?

What change is most effective?

• A) Include the existing test file in context so Claude can determine what scenarios are already covered. [CORRECT]
• B) Reduce the requested number of suggestions from 10 to 5, assuming Claude prioritizes the most valuable cases first.
• C) Add instructions directing Claude to focus exclusively on edge cases and error conditions rather than success paths.
• D) Implement post-processing that filters suggestions whose descriptions match existing test names via keyword overlap.

Why A: Including the existing test file fixes the root cause of duplication: Claude can only avoid suggesting already-covered scenarios if it knows what tests already exist. This gives Claude the information needed to propose genuinely new, valuable tests.

Question 25 (Scenario: Claude Code for Continuous Integration)

Situation: After an initial automated review identifies 12 findings, a developer pushes new commits to address issues. Re-running review produces 8 findings, but developers report that 5 duplicate previous comments on code that was already fixed in the new commits. What is the most effective way to eliminate this redundant feedback while maintaining thoroughness?

What is the most effective way to eliminate redundant feedback?

• A) Run review only when the PR is created and in the final pre-merge state, skipping intermediate commits.
• B) Add a post-processing filter that removes findings that match previous ones by file paths and issue descriptions before posting comments.
• C) Restrict review scope to files changed in the most recent push, excluding files from earlier commits.
• D) Include previous review findings in context and instruct Claude to report only new or still-unresolved issues. [CORRECT]

Why D: Including prior review findings in context lets Claude distinguish new problems from those already addressed in recent commits. This preserves review thoroughness while using Claude’s reasoning to avoid redundant feedback on fixed code.

Question 26 (Scenario: Claude Code for Continuous Integration)

Situation: Your pipeline script runs claude "Analyze this pull request for security issues", but the job hangs indefinitely. Logs show Claude Code is waiting for interactive input. What is the correct approach to run Claude Code in an automated pipeline?

What is the correct approach?

• A) Add a --batch flag: claude --batch "Analyze this pull request for security issues".
• B) Add the -p flag: claude -p "Analyze this pull request for security issues". [CORRECT]
• C) Redirect stdin from /dev/null: claude "Analyze this pull request for security issues" < /dev/null.
• D) Set the environment variable CLAUDE_HEADLESS=true before running the command.

Why B: The -p (or --print) flag is the documented way to run Claude Code non-interactively. It processes the prompt, prints the result to stdout, and exits without waiting for user input—ideal for CI/CD pipelines.

Question 27 (Scenario: Claude Code for Continuous Integration)

Situation: A pull request changes 14 files in an inventory tracking module. A single-pass review that analyzes all files together produces inconsistent results: detailed feedback on some files but shallow comments on others, missed obvious bugs, and contradictory feedback (a pattern is flagged in one file but identical code is approved in another file in the same PR). How should you restructure the review?

How should you restructure the review?

• A) Run three independent full-PR review passes and flag only issues that appear in at least two of the three runs.
• B) Split into focused passes: review each file individually for local issues, then run a separate integration-oriented pass to examine cross-file data flows. [CORRECT]
• C) Require developers to split large PRs into smaller submissions of 3–4 files before running automated review.
• D) Switch to a larger model with a bigger context window so it can pay sufficient attention to all 14 files in one pass.

Why B: Focused per-file passes address the root cause—attention dilution—by ensuring consistent depth and reliable local issue detection. A separate integration-oriented pass then covers cross-file concerns such as dependency and data-flow interactions.

Question 28 (Scenario: Claude Code for Continuous Integration)

Situation: Your automated code review averages 15 findings per pull request, and developers report a 40% false-positive rate. The bottleneck is investigation time: developers must click into each finding to read Claude’s rationale before deciding whether to fix or dismiss it. Your CLAUDE.md already contains comprehensive rules for acceptable patterns, and stakeholders rejected any approach that filters findings before developers see them. What change best addresses investigation time?

What change best addresses investigation time?

• A) Require Claude to include its rationale and confidence estimate directly in each finding. [CORRECT]
• B) Add a post-processor that analyzes finding patterns and automatically suppresses those that match historical false-positive signatures.
• C) Categorize findings as “blocking issues” vs “suggestions,” with different review requirements by level.
• D) Configure Claude to show only high-confidence findings, filtering uncertain flags before developers see them.

Why A: Including rationale and confidence directly in each finding reduces investigation time by letting developers quickly triage without opening each finding. It satisfies the “no filtering” constraint because all findings remain visible while accelerating developer decision-making.

Question 29 (Scenario: Claude Code for Continuous Integration)

Situation: Analysis of your automated code review shows large differences in false-positive rates by finding category: security/correctness findings have 8% false positives, performance findings 18%, style/naming findings 52%, and documentation findings 48%. Developer surveys show growing distrust—many start dismissing findings without reading because “half are wrong.” High-false-positive categories erode trust in accurate categories. Which approach best restores developer trust while improving the system?

Which approach best restores developer trust?

• A) Temporarily disable high-false-positive categories (style, naming, documentation) and keep only high-precision categories while improving prompts. [CORRECT]
• B) Keep all categories enabled but display confidence scores with each finding so developers can decide what to investigate.
• C) Keep all categories enabled and add few-shot examples to improve accuracy for each category over the next few weeks.
• D) Apply a uniform strictness reduction across all categories to bring the overall false-positive rate down.

Why A: Temporarily disabling high-false-positive categories immediately stops trust erosion by removing noisy findings that cause developers to dismiss everything, while preserving value from high-precision categories like security and correctness. It also creates space to improve prompts for problematic categories before re-enabling them.

Question 30 (Scenario: Claude Code for Continuous Integration)

Situation: Your team wants to reduce API costs for automated analysis. Currently, synchronous Claude calls support two workflows: (1) a blocking pre-merge check that must complete before developers can merge, and (2) a technical debt report generated overnight for review the next morning. Your manager proposes moving both to the Message Batches API to save 50%. How should you evaluate this proposal?

How should you evaluate this proposal?

• A) Move both to batch processing with fallback to synchronous calls if batches take too long.
• B) Move both workflows to batch processing with status polling to verify completion.
• C) Use batch processing only for technical debt reports; keep synchronous calls for pre-merge checks. [CORRECT]
• D) Keep synchronous calls for both workflows to avoid issues with batch result ordering.

Why C: Message Batches API processing can take up to 24 hours with no latency SLA, which is acceptable for overnight technical debt reports but unacceptable for blocking pre-merge checks where developers wait. This matches each workflow to the right API based on latency requirements.

Scenario: Code Generation with Claude Code

Question 31 (Scenario: Code Generation with Claude Code)

Situation: You asked Claude Code to implement a function that transforms API responses into an internal normalized format. After two iterations, the output structure still doesn’t match expectations—some fields are nested differently and timestamps are formatted incorrectly. You described requirements in prose, but Claude interprets them differently each time.

Which approach is most effective for the next iteration?

• A) Write a JSON schema describing the expected output structure and validate Claude’s output against it after each iteration.
• B) Provide 2–3 concrete input-output examples showing the expected transformation for representative API responses. [CORRECT]
• C) Rewrite requirements with more technical precision, specifying exact field mappings, nesting rules, and timestamp format strings.
• D) Ask Claude to explain its current understanding of the requirements to identify where interpretations diverge.

Why B: Concrete input-output examples remove ambiguity inherent in prose descriptions by showing Claude the exact expected transformation results. This directly addresses the root cause—misinterpretation of textual requirements—by providing unambiguous patterns for field nesting and timestamp formatting.

Question 32 (Scenario: Code Generation with Claude Code)

Situation: You need to add Slack as a new notification channel. The existing codebase has clear, established patterns for email, SMS, and push channels. However, Slack’s API offers fundamentally different integration approaches—incoming webhooks (simple, one-way), bot tokens (support delivery confirmation and programmatic control), or Slack Apps (two-way events, requires workspace approval). Your task says “add Slack support” without specifying integration method or requiring advanced features like delivery tracking.

How should you approach this task?

• A) Start in direct execution mode using incoming webhooks to match the existing one-way notification pattern.
• B) Switch to planning mode to explore integration options and architectural implications, then present a recommendation before implementation. [CORRECT]
• C) Start in direct execution mode by scaffolding a Slack channel class using existing patterns, deferring the integration method decision.
• D) Start in direct execution mode using a bot-token approach to ensure delivery confirmation is possible.

Why B: Slack integration has multiple valid approaches with significantly different architectural implications, and requirements are ambiguous. Planning mode lets you evaluate trade-offs among webhooks, bot tokens, and Slack Apps and align on an approach before implementation.

Question 33 (Scenario: Code Generation with Claude Code)

Situation: Your CLAUDE.md file has grown to 400+ lines containing coding standards, testing conventions, a detailed PR review checklist, deployment instructions, and database migration procedures. You want Claude to always follow coding standards and testing conventions, but apply PR review, deploy, and migration guidance only when doing those tasks.

Which restructuring approach is most effective?

• A) Move all guidance into separate Skills files organized by workflow type, leaving only a brief project description in CLAUDE.md.
• B) Keep everything in CLAUDE.md but use @import syntax to organize into separately maintained files by category.
• C) Split CLAUDE.md into files under .claude/rules/ with path-bound glob patterns so each rule loads only for the relevant file types.
• D) Keep universal standards in CLAUDE.md and create Skills for workflow-specific guidance (PR review, deploy, migrations) with trigger keywords. [CORRECT]

Why D: CLAUDE.md content loads in every session, ensuring coding standards and testing conventions always apply, while Skills are invoked on demand when Claude detects trigger keywords—ideal for workflow-specific guidance like PR review, deployment, and migrations.

Question 34 (Scenario: Code Generation with Claude Code)

Situation: You’re tasked with restructuring your team’s monolithic application into microservices. This impacts changes across dozens of files and requires decisions about service boundaries and module dependencies.

Which approach should you choose?

• A) Switch to planning mode to explore the codebase, understand dependencies, and design the implementation approach before making changes. [CORRECT]
• B) Start in direct execution mode and switch to planning only after encountering unexpected complexity during implementation.
• C) Start in direct execution mode and make incremental changes, letting implementation reveal natural service boundaries.
• D) Use direct execution with detailed upfront instructions that specify each service structure.

Why A: Planning mode is the right strategy for complex architectural restructuring like splitting a monolith: it allows safe exploration and informed decisions about boundaries before committing to potentially expensive changes across many files.

Question 35 (Scenario: Code Generation with Claude Code)

Situation: Your team created a /analyze-codebase skill that performs deep code analysis—dependency scanning, test coverage counts, and code quality metrics. After running the command, team members report Claude becomes less responsive in the session and loses the context of the original task.

How do you most effectively fix this while keeping full analysis capabilities?

• A) Add context: fork in the skill frontmatter to run the analysis in an isolated subagent context. [CORRECT]
• B) Add model: haiku in frontmatter to use a faster, cheaper model for analysis.
• C) Split the skill into three smaller skills, each producing less output.
• D) Add instructions to the skill to compress all results into a short summary before displaying them.

Why A: context: fork runs the analysis in an isolated subagent context so the large output does not pollute the main session’s context window and Claude does not lose track of the original task. It preserves full analysis capability while keeping the main session responsive.

Question 36 (Scenario: Code Generation with Claude Code)

Situation: Your team uses a /commit skill in .claude/skills/commit/SKILL.md. A developer wants to customize it for their personal workflow (different commit message format, extra checks) without affecting teammates.

What do you recommend?

• A) Create a personal version under ~/.claude/skills/ with a different name, e.g., /my-commit.
• B) Add conditional logic based on username in the project skill frontmatter.
• C) Create a personal version at ~/.claude/skills/commit/SKILL.md with the same name. [CORRECT]
• D) Set override: true in the personal skill frontmatter to prioritize it over the project version.

Why C: Personal skills take precedence over project skills with the same name. A personal skill at ~/.claude/skills/commit/SKILL.md will override the team’s project skill, allowing the developer to customize their workflow while maintaining the familiar /commit command name for their personal use. This approach is better than option A because it preserves the original command name, improving the developer’s workflow without affecting teammates.

Question 37 (Scenario: Code Generation with Claude Code)

Situation: Your team has used Claude Code for months. Recently, three developers report Claude follows the guidance “always include comprehensive error handling,” but a fourth developer who just joined says Claude does not follow it. All four work in the same repo and have up-to-date code.

What is the most likely cause and fix?

• A) The guidance lives in the original developers’ user-level ~/.claude/CLAUDE.md files, not in the project .claude/CLAUDE.md. Move the instruction to the project-level file so all team members receive it. [CORRECT]
• B) The new developer’s ~/.claude/CLAUDE.md contains conflicting instructions overriding project settings; they should delete the conflicting section.
• C) Claude Code learns per-user preferences over time; the new developer must repeat the requirement until Claude “remembers” it.
• D) Claude Code caches CLAUDE.md after first read; original developers use cached versions. Everyone should clear the Claude Code cache.

Why A: If the guidance was added only to the original developers’ user-level configs and not to the project-level .claude/CLAUDE.md, new team members won’t receive it. Moving it to the project-level configuration ensures all current and future team members automatically get the guidance.

Question 38 (Scenario: Code Generation with Claude Code)

Situation: You find that including 2–3 full endpoint implementation examples as context significantly improves consistency when generating new API endpoints. However, this context is useful only when creating new endpoints—not when debugging, reviewing code, or other work in the API directory.

Which configuration approach is most effective?

• A) Add endpoint examples and pattern documentation to the project CLAUDE.md so they are always available.
• B) Manually reference endpoint examples in every generation request by copying code into the prompt.
• C) Configure path-specific rules in .claude/rules/api/ that include endpoint examples and activate when working in the API directory.
• D) Create a skill that references the endpoint examples and contains pattern-following instructions, invoked on demand via a slash command. [CORRECT]

Why D: A skill invoked on demand loads the example context only when generating new endpoints, not during unrelated tasks like debugging or review. This keeps the main context clean while preserving high-quality generation when needed.

Question 39 (Scenario: Code Generation with Claude Code)

Situation: Your team created a /migration skill that generates database migration files. It takes the migration name via $ARGUMENTS. In production you observe three issues: (1) developers often run the skill without arguments, causing poorly named files, (2) the skill sometimes uses database schema details from unrelated prior conversations, and (3) a developer accidentally ran destructive test cleanup when the skill had broad tool access.

Which configuration approach fixes all three problems?

• A) Use positional parameters $1 and $2 instead of $ARGUMENTS to enforce specific inputs, include explicit schema file references via @ syntax for context control, and add a frontmatter description warning about destructive operations.
• B) Add argument-hint in frontmatter to request required parameters, use context: fork to isolate execution, and restrict allowed-tools to file-write operations. [CORRECT]
• C) Split into /migration-create and /migration-apply skills, add validation instructions to request migration name if missing, and use different allowed-tools scopes for each.
• D) Add validation instructions in the skill SKILL.md to ensure $ARGUMENTS is a valid name, add prompts to ignore prior conversation context, and list prohibited operations to avoid.

Why B: This uses three separate configuration features to address each problem: argument-hint improves argument entry and reduces missing arguments, context: fork prevents context leakage from prior conversations, and allowed-tools constrains the skill to safe file-writing operations, preventing destructive actions.

Question 40 (Scenario: Code Generation with Claude Code)

Situation: Your codebase contains areas with different coding conventions: React components use functional style with hooks, API handlers use async/await with specific error handling, and database models follow the repository pattern. Test files are distributed across the codebase next to the code under test (e.g., Button.test.tsx next to Button.tsx), and you want all tests to follow the same conventions regardless of location.

What is the most supported way to ensure Claude automatically applies the correct conventions when generating code?

• A) Put all conventions in the root CLAUDE.md under headings for each area and rely on Claude to infer which section applies.
• B) Create skills in .claude/skills/ for each code type, embedding conventions in each SKILL.md.
• C) Place a separate CLAUDE.md file in each subdirectory containing conventions for that area.
• D) Create rule files under .claude/rules/ with YAML frontmatter specifying glob patterns to conditionally apply conventions based on file paths. [CORRECT]

Why D: .claude/rules/ files with YAML frontmatter and glob patterns (e.g., **/*.test.tsx, src/api/**/*.ts) enable deterministic, path-based convention application regardless of directory structure. This is the most supported approach for cross-cutting patterns like distributed test files.

Question 41 (Scenario: Code Generation with Claude Code)

Situation: You want to create a custom slash command /review that runs your team’s standard code review checklist. It should be available to every developer when they clone or update the repository.

Where should you create the command file?

• A) In ~/.claude/commands/ in each developer’s home directory.
• B) In the project repository under .claude/commands/. [CORRECT]
• C) In .claude/config.json as an array of commands.
• D) In the root project CLAUDE.md.

Why B: Putting custom slash commands under .claude/commands/ inside the project repository ensures they are version-controlled and automatically available to every developer who clones or updates the repo. This is the intended location for project-level custom commands in Claude Code.

Question 42 (Scenario: Code Generation with Claude Code)

Situation: Your team’s CLAUDE.md grew beyond 500 lines mixing TypeScript conventions, testing guidance, API patterns, and deployment procedures. Developers find it hard to locate and update the right sections.

What approach does Claude Code support to organize project-level instructions into focused topical modules?

• A) Define a .claude/config.yaml mapping file patterns to specific sections inside CLAUDE.md.
• B) Create separate Markdown files in .claude/rules/, each covering one topic (e.g., testing.md, api-conventions.md). [CORRECT]
• C) Split instructions into README.md files in relevant subdirectories that Claude automatically loads as instructions.
• D) Create multiple files named CLAUDE.md at different levels of the directory tree, each overriding parent instructions.

Why B: Claude Code supports a .claude/rules/ directory where you can create separate Markdown files for topical guidance (e.g., testing.md, api-conventions.md), allowing teams to organize large instruction sets into focused, maintainable modules.

Question 43 (Scenario: Code Generation with Claude Code)

Situation: You create a custom skill /explore-alternatives that your team uses to brainstorm and evaluate implementation approaches before choosing one. Developers report that after running the skill, subsequent Claude responses are influenced by the alternatives discussion—sometimes referencing rejected approaches or retaining exploration context that interferes with actual implementation.

How should you most effectively configure this skill?

• A) Use the ! prefix in the skill to run exploration logic as a bash subprocess.
• B) Add context: fork in the skill frontmatter. [CORRECT]
• C) Split into two skills—/explore-start and /explore-end—to mark boundaries when exploration context should be discarded.
• D) Create the skill in ~/.claude/skills/ instead of .claude/skills/.

Why B: context: fork runs the skill in an isolated subagent context so exploration discussions do not pollute the main conversation history. This prevents rejected approaches and brainstorming context from influencing subsequent implementation work.

Question 44 (Scenario: Code Generation with Claude Code)

Situation: Your team wants to add a GitHub MCP server for searching PRs and checking CI status via Claude Code. Each of six developers has their own personal GitHub access token. You want consistent tooling across the team without committing credentials to version control.

Which configuration approach is most effective?

• A) Have each developer add the server in user scope via claude mcp add --scope user.
• B) Create an MCP server wrapper that reads tokens from a .env file and proxies GitHub API calls, then add the wrapper to the project .mcp.json.
• C) Add the server to the project .mcp.json using environment variable substitution (${GITHUB_TOKEN}) for auth and document the required environment variable in the project README. [CORRECT]
• D) Configure the server in project scope with a placeholder token, then tell developers to override it in their local config.

Why C: A project .mcp.json with environment variable substitution is idiomatic: it provides a single version-controlled source of truth for MCP configuration while letting each developer supply credentials via environment variables. Documenting the variable makes onboarding easy without committing secrets.

Question 45 (Scenario: Code Generation with Claude Code)

Situation: You’re adding error-handling wrappers around external API calls across a 120-file codebase. The work has three phases: (1) discover all call sites and patterns, (2) collaboratively design the error-handling approach, and (3) implement wrappers consistently. In Phase 1, Claude generates large output listing hundreds of call sites with context, quickly filling the context window before discovery finishes.

Which approach is most effective to complete the task while maintaining implementation consistency?

• A) Use an Explore subagent for Phase 1 to isolate verbose discovery output and return a summary, then continue Phases 2–3 in the main conversation. [CORRECT]
• B) Do all phases in the main conversation, periodically using /compact to reduce context usage while moving through files.
• C) Switch to headless mode with --continue, passing explicit context summaries between batch calls to maintain continuity.
• D) Define the error-handling pattern in CLAUDE.md, then process files in batches across multiple sessions relying on the shared memory file for consistency.

Why A: An Explore subagent isolates the verbose discovery output in a separate context and returns only a concise summary to the main conversation. This preserves the main context window for the collaborative design and consistent implementation phases where retained context is most valuable.

Scenario: Customer Support Agent

Question 46 (Scenario: Customer Support Agent)

Situation: While testing, you notice the agent often calls get_customer when users ask about order status, even though lookup_order would be more appropriate. What should you check first to address this problem?

What should you check first?

• A) Implement a preprocessing classifier to detect order-related requests and route them directly to lookup_order.
• B) Reduce the number of tools available to the agent to simplify choice.
• C) Add few-shot examples to the system prompt covering all possible order request patterns to improve tool selection.
• D) Check the tool descriptions to ensure they clearly differentiate each tool’s purpose. [CORRECT]

Why D: Tool descriptions are the primary input the model uses to decide which tool to call. When an agent consistently picks the wrong tool, the first diagnostic step is to verify that tool descriptions clearly separate each tool’s purpose and usage boundaries.

Question 47 (Scenario: Customer Support Agent)

Situation: Your agent handles single-issue requests with 94% accuracy (e.g., “I need a refund for order #1234”). But when customers include multiple issues in one message (e.g., “I need a refund for order #1234 and also want to update the shipping address for order #5678”), tool selection accuracy drops to 58%. The agent usually solves only one issue or mixes parameters across requests. What approach most effectively improves reliability for multi-issue requests?

What approach is most effective?

• A) Implement a preprocessing layer that uses a separate model call to decompose multi-issue messages into separate requests, handle each independently, and merge results.
• B) Combine related tools into fewer universal tools.
• C) Add few-shot examples to the prompt demonstrating correct reasoning and tool sequencing for multi-issue requests. [CORRECT]
• D) Implement response validation that detects incomplete answers and automatically reprompts the agent to resolve missed issues.

Why C: Few-shot examples that demonstrate correct reasoning and tool sequencing for multi-issue requests are most effective because the agent already performs well on single issues—what it needs is guidance on the pattern for decomposing and routing multiple issues and keeping parameters separated.

Question 48 (Scenario: Customer Support Agent)

Situation: Production logs show that for simple requests like “refund for order #1234,” your agent resolves the issue in 3–4 tool calls with 91% success. But for complex requests like “I was billed twice, my discount didn’t apply, and I want to cancel,” the agent averages 12+ tool calls with only 54% success—often investigating issues sequentially and fetching redundant customer data for each. What change most effectively improves handling of complex requests?

What change is most effective?

• A) Add explicit verification checkpoints between stages, requiring the agent to record progress after resolving each issue before moving to the next.
• B) Reduce the number of tools by combining get_customer, lookup_order, and billing-related tools into a single investigate_issue tool.
• C) Decompose the request into separate issues, then investigate each in parallel using shared customer context before synthesizing a final resolution. [CORRECT]
• D) Add few-shot examples to the system prompt demonstrating ideal tool-call sequences for various multi-faceted billing scenarios.

Why C: Decomposing into separate issues and investigating in parallel with shared customer context fixes both key problems: it eliminates redundant data retrieval by reusing shared context across issues and reduces total tool-call loops by parallelizing investigation before synthesizing a single resolution.

Question 49 (Scenario: Customer Support Agent)

Situation: Your agent achieves 55% first-contact resolution, well below the 80% target. Logs show it escalates simple cases (standard replacements for damaged goods with photo proof) while trying to handle complex situations requiring policy exceptions autonomously. What is the most effective way to improve escalation calibration?

What is the most effective way to improve escalation calibration?

• A) Require the agent to self-rate confidence on a 1–10 scale before each response and automatically route to humans when confidence drops below a threshold.

• B) Deploy a separate classifier model trained on historical tickets to predict which requests need escalation before the main agent starts processing.

• C) Add explicit escalation criteria to the system prompt with few-shot examples showing when to escalate versus resolve autonomously. [CORRECT]

• D) Implement sentiment analysis to determine customer frustration level and automatically escalate past a negative sentiment threshold.

Why C: Explicit escalation criteria with few-shot examples directly address the root cause—unclear decision boundaries between simple and complex cases. It’s the most proportional, effective first intervention that teaches the agent when to escalate and when to resolve autonomously without extra infrastructure.

Question 50 (Scenario: Customer Support Agent)

Situation: After calling get_customer and lookup_order, the agent has all available system data but still faces uncertainty. Which situation is the most justified trigger for calling escalate_to_human?

Which situation is most justified for escalation?

• A) A customer wants to cancel an order shipped yesterday and arriving tomorrow. The agent should escalate because the customer might change their mind after receiving the package.
• B) A customer claims they didn’t receive an order, but tracking shows it was delivered and signed for at their address three days ago. The agent should escalate because presenting contradictory evidence could harm the customer relationship.
• C) A customer requests competitor price matching. Your policies allow price adjustments for price drops on your own site within 14 days, but say nothing about competitor prices. The agent should escalate for policy interpretation. [CORRECT]
• D) A customer message contains both a billing question and a product return. The agent should escalate so a human can coordinate both issues in one interaction.

Why C: This is a genuine policy gap: company rules cover price drops on your own site but do not address competitor price matching. The agent must not invent policy and should escalate for human judgment on how to interpret or extend existing rules.

Question 51 (Scenario: Customer Support Agent)

Situation: Production logs show that in 12% of cases your agent skips get_customer and calls lookup_order directly using only the customer-provided name, sometimes leading to misidentified accounts and incorrect refunds. What change most effectively fixes this reliability problem?

What change is most effective?

• A) Add few-shot examples showing that the agent always calls get_customer first, even when customers voluntarily provide order details.
• B) Implement a routing classifier that analyzes each request and enables only a subset of tools appropriate for that request type.
• C) Add a programmatic precondition that blocks lookup_order and process_refund until get_customer returns a verified customer identifier. [CORRECT]
• D) Strengthen the system prompt stating that customer verification via get_customer is mandatory before any order operations.

Why C: A programmatic precondition provides a deterministic guarantee that required sequencing is followed. It’s the most effective approach because it eliminates the possibility of skipping verification, regardless of LLM behavior.

Question 52 (Scenario: Customer Support Agent)

Situation: Production metrics show that when resolving complex billing disputes or multi-order returns, customer satisfaction scores are 15% lower than for simple cases—even when the resolution is technically correct. Root-cause analysis shows the agent provides accurate solutions but inconsistently explains rationale: sometimes omitting relevant policy details, sometimes missing timeline info or next steps. The specific context gaps vary case by case. You want to improve solution quality without adding human oversight. What approach is most effective?

What approach is most effective?

• A) Add a self-critique stage where the agent evaluates a draft response for completeness—ensuring it resolves the customer’s issue, includes relevant context, and anticipates follow-up questions. [CORRECT]
• B) Add a confirmation stage where the agent asks “Does this fully resolve your issue?” before closing, allowing customers to request additional information if needed.
• C) Upgrade the model from Haiku to Sonnet for complex cases, routing based on a defined complexity metric.
• D) Implement few-shot examples in the system prompt showing complete explanations for five common complex case types, demonstrating how to include policy context, timelines, and next steps.

Why A: A self-critique stage (the evaluator-optimizer pattern) directly addresses inconsistent explanation completeness by forcing the agent to assess its own draft against concrete criteria—such as policy context, timelines, and next steps—before presenting it. This catches case-specific gaps without human oversight.

Question 53 (Scenario: Customer Support Agent)

Situation: Production metrics show your agent averages 4+ API loops per resolution. Analysis reveals Claude often requests get_customer and lookup_order in separate sequential turns even when both are needed initially. What is the most effective way to reduce the number of loops?

What is the most effective way to reduce loops?

• A) Implement speculative execution that automatically calls likely-needed tools in parallel with any requested tool and returns all results regardless of what was requested.
• B) Increase max_tokens to give Claude more room to plan and naturally combine tool requests.
• C) Create composite tools like get_customer_with_orders that bundle common lookup combinations into single calls.
• D) Instruct Claude in the prompt to bundle tool requests into one turn and return all results together before the next API call. [CORRECT]

Why D: Prompting Claude to bundle related tool requests into a single turn leverages its native ability to request multiple tools at once. It directly fixes the sequential-call pattern with minimal architectural change.

Question 54 (Scenario: Customer Support Agent)

Situation: Production logs show a pattern: customers reference specific amounts (e.g., “the 15% discount I mentioned”), but the agent responds with incorrect values. Investigation shows these details were mentioned 20+ turns ago and condensed into vague summaries like “promotional pricing was discussed.” What fix is most effective?

What fix is most effective?

• A) Increase the summarization threshold from 70% to 85% so conversations have more room before summarization triggers.
• B) Store full conversation history in external storage and implement retrieval when the agent detects references like “as I mentioned.”
• C) Extract transactional facts (amounts, dates, order numbers) into a persistent “case facts” block included in every prompt outside the summarized history. [CORRECT]
• D) Revise the summarization prompt to explicitly preserve all numbers, percentages, dates, and customer-stated expectations verbatim.

Why C: Summarization inherently loses precise details. Extracting transactional facts into a structured “case facts” block outside the summarized history preserves critical information so it’s reliably available in every prompt regardless of how many turns have been summarized.

Question 55 (Scenario: Customer Support Agent)

Situation: Your get_customer tool returns all matches when searching by name. Currently, when there are multiple results, Claude picks the customer with the most recent order, but production data shows this selects the wrong account 15% of the time for ambiguous matches. How should you address this?

How should you address this?

• A) Implement a confidence scoring system that acts autonomously above 85% confidence and requests clarification below the threshold.
• B) Instruct Claude to request an additional identifier (email, phone, or order number) when get_customer returns multiple matches before taking any customer-specific action. [CORRECT]
• C) Modify get_customer to return only a single most-likely match based on a ranking algorithm, eliminating ambiguity.
• D) Add few-shot examples to the prompt demonstrating correct reasoning and tool sequencing for ambiguous matches.

Why B: Asking the user for an additional identifier is the most reliable way to resolve ambiguity because the user has definitive knowledge of their identity. One extra conversational turn is a small price to pay to eliminate a 15% error rate caused by choosing the wrong account.

Question 56 (Scenario: Customer Support Agent)

Situation: Production logs show a consistent pattern: when customers include the word “account” in their message (e.g., “I want to check my account for an order I made yesterday”), the agent calls get_customer first 78% of the time. When customers phrase similar requests without “account” (e.g., “I want to check an order I made yesterday”), it calls lookup_order first 93% of the time. Tool descriptions are clear and unambiguous. What is the most likely root cause of this discrepancy?

What is the most likely root cause?

• A) The system prompt contains keyword-sensitive instructions that steer behavior based on terms like “account,” creating unintended tool-selection patterns. [CORRECT]
• B) The model’s base training creates associations between “account” terminology and customer-related operations that override tool descriptions.
• C) The model needs more training data on multi-concept messages and should be fine-tuned on examples containing both account and order terminology.
• D) Tool descriptions need additional negative examples specifying when NOT to use each tool to prevent this keyword-induced confusion.

Why A: The systematic keyword-driven pattern (78% vs 93%) strongly indicates explicit routing logic in the system prompt reacting to the word “account” and steering the agent toward customer-related tools. Since tool descriptions are already clear, the discrepancy points to prompt-level instructions creating unintended behavioral steering.

Question 57 (Scenario: Customer Support Agent)

Situation: Production logs show the agent often calls get_customer when users ask about orders (e.g., “check my order #12345”) instead of calling lookup_order. Both tools have minimal descriptions (“Gets customer information” / “Gets order details”) and accept similar-looking identifier formats. What is the most effective first step to improve tool selection reliability?

What is the most effective first step?

• A) Implement a routing layer that analyzes user input before each turn and preselects the correct tool based on detected keywords and ID patterns.
• B) Combine both tools into a single lookup_entity that accepts any identifier and internally decides which backend to query.
• C) Add few-shot examples to the system prompt demonstrating correct tool selection patterns, with 5–8 examples routing order-related queries to lookup_order.
• D) Expand each tool’s description to include input formats, example queries, edge cases, and boundaries explaining when to use it versus similar tools. [CORRECT]

Why D: Expanding tool descriptions with input formats, example queries, edge cases, and clear boundaries directly fixes the root cause—minimal descriptions that don’t give the LLM enough information to distinguish similar tools. It’s a low-effort, high-impact first step that improves the primary mechanism the LLM uses for tool selection.

Question 58 (Scenario: Customer Support Agent)

Situation: You are implementing the agent loop for your support agent. After each Claude API call, you must decide whether to continue the loop (run requested tools and call Claude again) or stop (present the final answer to the customer). What determines this decision?

What determines this decision?

• A) Check the stop_reason field in Claude’s response—continue if it is tool_use and stop if it is end_turn. [CORRECT]
• B) Parse Claude’s text for phrases like “I’m done” or “Can I help with anything else?”—natural language signals indicate task completion.
• C) Set a maximum iteration count (e.g., 10 calls) and stop when reached, regardless of whether Claude indicates more work is needed.
• D) Check whether the response contains assistant text content—if Claude generated explanatory text, the loop should terminate.

Why A: stop_reason is Claude’s explicit structured signal for loop control: tool_use indicates Claude wants to run a tool and receive results back, while end_turn indicates Claude has completed its response and the loop should end.

Question 59 (Scenario: Customer Support Agent)

Situation: Production logs show the agent misinterprets outputs from your MCP tools: Unix timestamps from get_customer, ISO 8601 dates from lookup_order, and numeric status codes (1=pending, 2=shipped). Some tools are third-party MCP servers you cannot modify. Which approach to data format normalization is most maintainable?

Which approach is most maintainable?

• A) Use a PostToolUse hook to intercept tool outputs and apply formatting transformations before the agent processes them. [CORRECT]
• B) Modify tools you control to return human-readable formats and create wrappers for third-party tools.
• C) Create a normalize_data tool that the agent calls after every data retrieval to transform values.
• D) Add detailed format documentation to the system prompt explaining each tool’s data conventions.

Why A: A PostToolUse hook provides a centralized, deterministic point to intercept and normalize all tool outputs—including third-party MCP server data—before the agent processes them. It’s more maintainable because transformations live in code and apply uniformly, rather than relying on LLM interpretation.

Question 60 (Scenario: Customer Support Agent)

Situation: Production logs show the agent sometimes chooses get_customer when lookup_order would be more appropriate, especially for ambiguous queries like “I need help with my recent purchase.” You decide to add few-shot examples to the system prompt to improve tool selection. Which approach most effectively addresses the problem?

Which approach is most effective?

• A) Add explicit “use when” and “don’t use when” guidance in each tool description covering ambiguous cases.
• B) Add examples grouped by tool—all get_customer scenarios together, then all lookup_order scenarios.
• C) Add 4–6 examples targeted at ambiguous scenarios, each with rationale for why one tool was chosen over plausible alternatives. [CORRECT]
• D) Add 10–15 examples of clear, unambiguous requests demonstrating correct tool choice for typical scenarios for each tool.

Why C: Targeting few-shot examples at the specific ambiguous scenarios where errors occur, with explicit rationale for why one tool is preferable to alternatives, teaches the model the comparative decision process needed for edge cases. This is more effective than generic examples or declarative rules.

Practical Exercises

Exercise 1: Multi-tool Agent with Escalation Logic

Goal: Design an agent loop with tool integration, structured error handling, and escalation.

Steps:

1. Define 3–4 MCP tools with detailed descriptions (include two similar tools to test tool selection)
2. Implement an agent loop checking stop_reason ("tool_use" / "end_turn")
3. Add structured error responses: errorCategory, isRetryable, description
4. Implement an interceptor hook that blocks operations above a threshold and routes to escalation
5. Test with multi-aspect requests

Domains: 1 (Agent architecture), 2 (Tools and MCP), 5 (Context and reliability)

Exercise 2: Configuring Claude Code for Team Development

Goal: Configure CLAUDE.md, custom commands, path-specific rules, and MCP servers.

Steps:

1. Create a project-level CLAUDE.md with universal standards
2. Create .claude/rules/ files with YAML frontmatter for different code areas (paths: ["src/api/**/*"], paths: ["**/*.test.*"])
3. Create a project skill under .claude/skills/ with context: fork and allowed-tools
4. Configure an MCP server in .mcp.json with environment variables + a personal override in ~/.claude.json
5. Test planning mode vs direct execution on tasks of different complexity

Domains: 3 (Claude Code configuration), 2 (Tools and MCP)

Exercise 3: Structured Data Extraction Pipeline

Goal: JSON schemas, tool_use for structured output, validation/retry loops, batch processing.

Steps:

1. Define an extraction tool with a JSON schema (required/optional fields, enums with "other", nullable fields)
2. Build a validation loop: on error, retry with the document, the incorrect extraction, and the specific validation error
3. Add few-shot examples for documents with different structures
4. Use batch processing via the Message Batches API: 100 documents, handle failures via custom_id
5. Route to humans: field-level confidence scores, document-type analysis

Domains: 4 (Prompt engineering), 5 (Context and reliability)

Exercise 4: Designing and Debugging a Multi-agent Research Pipeline

Goal: Subagent orchestration, context passing, error propagation, synthesis with source tracking.

Steps:

1. A coordinator with 2+ subagents (allowedTools includes "Task", context is passed explicitly in prompts)
2. Run subagents in parallel via multiple Task calls in a single response
3. Require structured subagent output: claim, quote, source URL, publication date
4. Simulate a subagent timeout: return structured error context to the coordinator and continue with partial results
5. Test with conflicting data: preserve both values with attribution; separate confirmed vs disputed findings

Domains: 1 (Agent architecture), 2 (Tools and MCP), 5 (Context and reliability)

Appendix: Technologies and Concepts

Out-of-Scope Topics

The following adjacent topics will NOT be on the exam:

• Fine-tuning Claude models or training custom models
• Claude API authentication, billing, or account management
• Detailed implementation in specific programming languages or frameworks (beyond what’s needed for tool/schema configuration)
• Deploying or hosting MCP servers (infrastructure, networking, container orchestration)
• Claude’s internal architecture, training process, or model weights
• Constitutional AI, RLHF, or safety training methodologies
• Embedding models or vector database implementation details
• Computer use (browser automation, desktop interaction)
• Image analysis capabilities (Vision)
• Streaming API or server-sent events
• Rate limiting, quotas, or detailed API cost calculations
• OAuth, API key rotation, or authentication protocol details
• Cloud-provider-specific configurations (AWS, GCP, Azure)
• Performance benchmarks or model comparison metrics
• Prompt caching implementation details (beyond knowing it exists)
• Token counting algorithms or tokenization specifics

Preparation Recommendations

1. Build an agent with the Claude Agent SDK — implement a full agent loop with tool calling, error handling, and session management. Practice subagents and explicit context passing.

2. Configure Claude Code for a real project — use CLAUDE.md hierarchy, path-specific rules in .claude/rules/, skills with context: fork and allowed-tools, and MCP server integration.

3. Design and test MCP tools — write descriptions that differentiate similar tools, return structured errors with categories and retry flags, and test against ambiguous user requests.

4. Build a data extraction pipeline — use tool_use with JSON schemas, validation/retry loops, optional/nullable fields, and batch processing via the Message Batches API.

5. Practice prompt engineering — add few-shot examples for ambiguous scenarios, explicit review criteria, and multi-pass architectures for large code reviews.

6. Study context management patterns — extract facts from verbose outputs, use scratchpad files, and delegate discovery to subagents to handle context limits.

7. Understand escalation and human-in-the-loop — when to escalate (policy gaps, explicit user request, inability to make progress) and confidence-based routing workflows.

8. Take a practice exam before the real one. It uses the same scenarios and format.