Skip to main content

Exam Scenarios

Every item sits inside a scenario. At delivery, 4 of the following 6 scenarios are selected at random and each frames a block of questions. Knowing the scenario tells you which domain's judgment rules the item is probing.

Preparation strategy

You cannot know which 4 scenarios will appear. Study all 6, but pay extra attention to Scenarios 1, 3, and 4 — they cover Domains 1 and 2, which together account for 45% of the exam.

Scenario 1 — Customer Support Resolution Agent​

A customer support resolution agent built on the Claude Agent SDK. It handles high-ambiguity requests (returns, billing disputes, account issues) and reaches backend systems through custom MCP tools: get_customer, lookup_order, process_refund, escalate_to_human. Target: 80%+ first-contact resolution while escalating correctly.

Primary domains: 1 (Agentic Architecture & Orchestration), 2 (Tool Design & MCP Integration), 5 (Context Management & Reliability)

Recurring item themes

  • Enforcing get_customer before lookup_order / process_refund — programmatic prerequisite, not a prompt
  • Disambiguating get_customer vs lookup_order through tool descriptions
  • Blocking refunds above a threshold (e.g. $500) with a tool-call interception hook
  • Escalation calibration: policy gaps, explicit customer requests, inability to progress
  • Trimming 40-field order lookups; persisting a "case facts" block across summarization
  • Structured handoff summaries for human agents with no transcript access

Scenario 2 — Code Generation with Claude Code​

Claude Code accelerating day-to-day development: generation, refactoring, debugging, documentation. Integration via custom slash commands and CLAUDE.md configuration, plus judgment about plan mode vs direct execution.

Primary domains: 3 (Claude Code Configuration & Workflows), 5 (Context Management & Reliability)

Recurring item themes

  • .claude/commands/ (project, version-controlled) vs ~/.claude/commands/ (personal)
  • CLAUDE.md hierarchy — why a teammate is not receiving your instructions
  • .claude/rules/ with paths: globs vs subdirectory CLAUDE.md files
  • Plan mode for architectural, multi-file, multiple-valid-approach work; direct execution for scoped fixes
  • Iterative refinement: input/output examples, test-driven iteration, the interview pattern
  • /memory, /compact, the Explore subagent

Scenario 3 — Multi-Agent Research System​

A coordinator agent delegating to specialized subagents: web search, document analysis, synthesis, report generation — the architecture described in how Anthropic built its multi-agent research system. The system produces comprehensive, cited reports.

Primary domains: 1 (Agentic Architecture & Orchestration), 2 (Tool Design & MCP Integration), 5 (Context Management & Reliability)

Recurring item themes

  • Hub-and-spoke: all subagent communication routed through the coordinator
  • Overly narrow task decomposition as the root cause of coverage gaps
  • Subagents do not inherit coordinator context — pass findings explicitly in the prompt
  • Parallel spawning: multiple Task calls in a single coordinator response
  • Scoped cross-role tools (e.g. verify_fact for the synthesis agent) vs over-provisioning
  • Structured error propagation on subagent timeout; partial results plus coverage annotations
  • Claim–source mappings, publication dates, conflict annotation through synthesis

Scenario 4 — Developer Productivity with Claude​

Agent SDK tooling that helps engineers explore unfamiliar codebases, understand legacy systems, generate boilerplate, and automate repetitive work. Uses the built-in tools (Read, Write, Bash, Grep, Glob) alongside MCP servers.

Primary domains: 2 (Tool Design & MCP Integration), 3 (Claude Code Configuration & Workflows), 1 (Agentic Architecture & Orchestration)

Recurring item themes

  • Grep (content) vs Glob (paths) vs Read/Write (whole file) vs Edit (unique-anchor edit)
  • Edit failing on non-unique text → Read + Write fallback
  • Incremental codebase understanding instead of reading everything upfront
  • MCP tool descriptions detailed enough that the agent stops preferring built-in Grep
  • Context degradation over long exploration sessions; scratchpad files; /compact
  • Session resumption with --resume vs starting fresh with an injected summary; fork_session

Scenario 5 — Claude Code for Continuous Integration​

Claude Code inside a CI/CD pipeline running automated code review, generating test cases, and posting PR feedback (non-interactive CLI). Prompts must be actionable and minimize false positives.

Primary domains: 3 (Claude Code Configuration & Workflows), 4 (Prompt Engineering & Structured Output)

Recurring item themes

  • -p / --print for non-interactive execution; the job hanging on interactive input
  • --output-format json with --json-schema for machine-parseable findings
  • Explicit categorical review criteria instead of "be conservative"
  • Per-file passes plus a cross-file integration pass for large PRs
  • Independent review instance instead of self-review by the generating session
  • Message Batches API for overnight debt reports; synchronous API for blocking pre-merge checks
  • CLAUDE.md documenting testing standards, fixtures, and review criteria

Scenario 6 — Structured Data Extraction​

An extraction system pulling information from unstructured documents, validating output against JSON schemas, and feeding downstream systems (structured outputs). It must degrade gracefully on edge cases.

Primary domains: 4 (Prompt Engineering & Structured Output), 5 (Context Management & Reliability)

Recurring item themes

  • tool_use with JSON schemas as the reliable structured-output mechanism
  • tool_choice — "auto" vs "any" vs forced {"type": "tool", "name": "..."}
  • Nullable/optional fields so the model returns null instead of fabricating
  • enum with "unclear" / "other" + detail string for extensible categories
  • Validation-retry with the specific error appended; when retry cannot help
  • Semantic validation: calculated_total vs stated_total, conflict_detected booleans
  • Batch submission cadence against an SLA; resubmitting failures by custom_id
  • Field-level confidence, stratified sampling, accuracy by document type and field