Every question below is written for this workspace, grounded in the task statements and worked examples from the official exam guide PDF — none of it reproduces the guide's own 12 sample questions verbatim (those are already woven into Lessons 1–14). Treat this as mixed-domain retrieval practice, not a leaked copy of the real exam. The guide mentions an official practice exam with a separately-provided link — see the note in RESOURCES.md; if you locate it, run that too before sitting the real thing.
How to run this drill
The real exam gives you 120 minutes for 60 questions across 4 of these 6 scenarios — roughly 2 minutes a question. Pick 4 of the 6 sections below at random, set a 24-minute timer (18 questions × ~1.3 min, scaled down since this is a fraction of the real thing), and go through without pausing to check answers until you're done with all four sections.
Q1. A customer explicitly says "I want to speak to a person" about a routine, easily resolvable billing question. What should the agent do?
Q2. process_refund returns { isError: true, errorCategory: "transient", isRetryable: true }. What should the agent do?
Q3. Your team decides refunds under $50 never require identity verification, but refunds over $50 always must. What's the correct way to implement this?
Scenario 2: Code Generation with Claude Code
Primary domains: Claude Code Configuration & Workflows, Context Management & Reliability. Custom slash commands, CLAUDE.md configuration, plan mode vs. direct execution.
Q1. You're fixing a single failing test with a clear stack trace pointing to one function. Should you use plan mode?
Q2. Hours into a resumed refactoring session, the agent starts giving generic answers about "typical patterns" instead of the specifics it found earlier. What's the most direct fix?
Q3. You ask Claude Code to "clean up this module" and get inconsistent results on each attempt. What's the most effective next step?
Q1. A coordinator spawns its synthesis subagent before both the web-search and document-analysis subagents have returned. What's wrong with this?
Q2. A document-analysis subagent is given all 12 of the system's tools, including web search and unrelated refund tools. What's the likely consequence?
Q3. A document-analysis subagent finds two paragraphs in the same report stating contradictory figures for the same metric. What should it do?
Scenario 4: Developer Productivity with Claude
Primary domains: Tool Design & MCP Integration, Claude Code Configuration & Workflows, Agentic Architecture & Orchestration. Built-in tools plus MCP servers for codebase exploration.
Q1. You need to find every place in a codebase that calls a specific function by name. Which built-in tool fits best?
Q2. A productivity agent needs to explore a large, unfamiliar codebase before changing anything, and full exploration risks filling the context window. What's the best approach?
Q3. This agent has both built-in tools and several MCP server tools connected. Should you expect all of them to be available to the model at once?
Scenario 5: Claude Code for Continuous Integration
Primary domains: Claude Code Configuration & Workflows, Prompt Engineering & Structured Output. Automated review and test generation in CI/CD, minimizing false positives.
Q1. Your CI script runs Claude Code with a review prompt, but the job hangs waiting for input. Which flag fixes this?
Q2. Your team wants to cut API costs on two workflows: a blocking pre-merge check, and an overnight technical-debt report. Which change is appropriate?
Q3. An automated review over an 18-file pull request produces contradictory feedback — flagging a pattern as an issue in one file while approving the identical pattern elsewhere in the same PR. What's the most effective restructuring?
Q1. You're extracting invoice data, and roughly 15% of source PDFs simply don't list a purchase-order number. What schema design avoids fabricated PO numbers?
Q2. Your extraction pipeline reports 98% field-level accuracy overall, but you suspect handwritten invoices perform worse than typed ones. How do you verify this?
Q3. You have three different extraction tools for three document types, and don't know in advance which type a given document is. You need to guarantee some extraction tool is called, but let the model pick which. Which tool_choice value fits?