Exam Scenarios (4 of 6 drawn at random)

Scenario 1: Customer Support Resolution Agent

Agent SDK-based agent for high-ambiguity requests (returns, billing disputes, account issues). Custom MCP tools: get_customer, lookup_order, process_refund, escalate_to_human. Target: 80%+ first-contact resolution while knowing when to escalate.

Primary domains: 1 (Agentic Architecture), 2 (Tool/MCP), 5 (Context/Reliability)

Key ideas to have ready: programmatic prerequisite gating before process_refund (identity verification); explicit escalation criteria with few-shot examples; structured error responses per tool; hooks to block refunds above a $ threshold.

Scenario 2: Code Generation with Claude Code

Team uses Claude Code for generation, refactoring, debugging, docs. Needs custom slash commands, CLAUDE.md configs, plan mode vs. direct execution judgment.

Primary domains: 3 (Claude Code Config), 5 (Context/Reliability)

Key ideas: CLAUDE.md hierarchy pitfalls (user vs project scope); .claude/commands/ vs ~/.claude/commands/; .claude/rules/ glob-scoped conventions; plan mode for architecturally complex/multi-file work, direct execution for well-scoped single-file changes; Explore subagent for verbose discovery.

Scenario 3: Multi-Agent Research System

Coordinator delegates to: web search, document analysis, synthesis, report-generation subagents. Produces comprehensive, cited reports.

Primary domains: 1 (Agentic Architecture), 2 (Tool/MCP), 5 (Context/Reliability)

Key ideas: narrow task decomposition as a root cause of coverage gaps; structured error propagation on subagent timeout; claim-source mapping preservation through synthesis; scoped cross-role tools (e.g., verify_fact on synthesis agent) vs. full tool access; parallel Task calls in one turn.

Scenario 4: Developer Productivity with Claude

Agent SDK-based tool to explore unfamiliar codebases, understand legacy systems, generate boilerplate, automate repetitive tasks. Uses built-ins (Read, Write, Bash, Grep, Glob) + MCP servers.

Primary domains: 2 (Tool/MCP), 3 (Claude Code Config), 1 (Agentic Architecture)

Key ideas: Grep vs Glob vs Edit/Write selection; incremental codebase understanding (Grep → Read, not read-everything); MCP server scoping (project vs user); scratchpad files + subagent delegation for context management during exploration.

Scenario 5: Claude Code for Continuous Integration

Automated code review, test case generation, PR feedback. Needs actionable feedback and minimal false positives.

Primary domains: 3 (Claude Code Config), 4 (Prompt Engineering)

Key ideas: -p flag for non-interactive CI runs; --output-format json + --json-schema; explicit review criteria over vague "be conservative" instructions; multi-pass (per-file + integration) review to avoid attention dilution; independent review instance vs. self-review; avoiding duplicate comments across re-runs.

Scenario 6: Structured Data Extraction

Extracts info from unstructured docs, validates via JSON schemas, must handle edge cases and integrate downstream.

Primary domains: 4 (Prompt Engineering), 5 (Context/Reliability)

Key ideas: tool_use + JSON schema for guaranteed syntax compliance (not semantic); nullable/optional fields to prevent fabrication; retry-with-error-feedback vs. retry-is-futile (missing source data); few-shot for varied document structures; stratified sampling + field-level confidence for human review routing; Batches API for non-blocking bulk extraction only.

Cross-Scenario Pattern Recognition

Most scenarios blend 2-3 domains. When reading a question, first identify: (a) which scenario frame it's in, (b) which domain task-statement it's really testing, (c) whether the "fix" needed is prompt-level, hook/programmatic-level, architecture-level, or schema-level. The exam consistently rewards picking the minimal sufficient fix for the stated root cause — not the most advanced-sounding option.