Exam Scenarios

MCP Tools

What the Exam Tests

• When to enforce tool ordering programmatically (prerequisite gates) vs relying on prompt instructions — prompt instructions have non-zero failure rates for safety-critical sequences
• How expanding tool descriptions with input formats, example queries, and boundary explanations reduces misrouting between similar tools (get_customer vs lookup_order)
• Escalation calibration: adding explicit criteria + few-shot examples to calibrate when to resolve autonomously vs escalate
• Structured handoff summaries: what to include when calling escalate_to_human (customer ID, root cause, refund amount, recommended action)
• PostToolUse hooks for normalizing heterogeneous data formats (Unix timestamps, ISO 8601, numeric status codes) before the agent processes them
• Task decomposition for multi-concern requests: splitting multi-issue cases into distinct items, investigating in parallel, then synthesizing a unified resolution

What the Exam Tests

• CLAUDE.md hierarchy: which level (user / project / directory) controls what, and why user-level settings aren't shared with teammates
• When to use plan mode vs direct execution — the triggering criteria (architectural decisions, multiple valid approaches, large-scale changes)
• Custom slash commands and skills: .claude/commands/ for team workflows, context: fork frontmatter to isolate verbose skill output from the main conversation
• Path-specific rules via .claude/rules/ with YAML paths frontmatter — load only when editing matching files
• Session management: --resume <session-name> for resumption, fork_session for parallel exploration branches
• When to start fresh with an injected summary vs resume (prior tool results stale after file modifications)

Subagent Roles

What the Exam Tests

• Hub-and-spoke coordination: all inter-subagent communication routes through the coordinator for consistent error handling and observability
• Context isolation: subagents do not inherit coordinator conversation history — context must be explicitly included in each subagent's prompt
• Parallel vs sequential spawning: emitting multiple Task tool calls in a single coordinator response to run subagents in parallel
• Scope partitioning: assigning distinct subtopics or source types to each subagent to minimize duplication
• Citation provenance: using structured data formats to separate content from metadata (source URLs, document names, page numbers) when passing context between agents
• Iterative refinement: coordinator evaluates synthesis gaps, re-delegates targeted queries, re-invokes synthesis until coverage is sufficient
• Tool scoping: restricting subagent tool sets to their role (synthesis agent should not have web search tools)

Built-in Tools

What the Exam Tests

• Selecting the right built-in tool: Grep for content search, Glob for path patterns, Read for full file loads, Edit for targeted modifications
• Incremental codebase understanding: start with Grep to find entry points, then Read to follow imports and trace flows — not reading all files upfront
• Tool set size management: giving an agent 18 tools instead of 4–5 degrades selection reliability; agents with out-of-scope tools tend to misuse them
• MCP server scoping: project-level .mcp.json for tools shared via version control; user-level ~/.claude.json for personal/experimental servers
• Enhancing MCP tool descriptions to prevent the agent from preferring built-in tools (like Grep) over more capable MCP-provided alternatives
• Choosing community MCP servers over custom implementations for standard integrations (e.g., Jira)

What the Exam Tests

• The -p / --print flag: required for running Claude Code non-interactively in CI — prevents hangs waiting for user input
• --output-format json with --json-schema: produces structured findings that can be automatically posted as inline PR comments
• Why the session that generated code is less effective at reviewing it — a model retains its generation reasoning context and is less likely to question its own decisions
• Including prior review findings on re-runs to instruct Claude to report only new or still-unaddressed issues, avoiding duplicate comments
• Providing existing test files in context so test generation avoids duplicating already-covered scenarios
• Explicit criteria for what to flag: "flag only issues where code behavior contradicts documented requirements" beats "be conservative"
• Temporarily disabling high false-positive categories to restore developer trust while improving prompts for those categories

What the Exam Tests

• tool_use with JSON schema is the most reliable approach for structured output — it eliminates syntax errors but does not prevent semantic errors (values in wrong fields, line items that don't sum to total)
• tool_choice: "any" guarantees the model calls a tool rather than returning conversational text — use when document type is unknown and multiple schemas exist
• Designing optional (nullable) schema fields when source documents may not contain the information — prevents model from fabricating values to satisfy required fields
• Retry-with-error-feedback: including the original document, failed extraction, and specific validation errors in the retry prompt for model self-correction
• When retries are ineffective: if information is simply absent from the source document, retrying cannot produce it — address the schema or accept null
• Few-shot examples with varied document structures reduce extraction failures from format diversity (inline citations vs bibliographies, measurement formats)
• Batch API matching: use synchronous API for pre-merge checks (blocking); batch API for overnight document processing (non-blocking, latency-tolerant)