# Full Practice Exam 1 (Week 11)
50 Questions -- All 6 Scenarios -- Simulates Actual Exam Format
Target time: 90 minutes | Passing score: 36/50 (720/1000)
---
## Scenario A: Customer Support Resolution Agent
You are building a customer support resolution agent using the Claude Agent SDK. The agent handles returns, billing disputes, and account issues via MCP tools (get_customer, lookup_order, process_refund, escalate_to_human). Target: 80%+ first-contact resolution.
### Q1
The agent occasionally processes refunds for the wrong customer because it uses the customer's stated name to call `lookup_order` instead of first verifying identity through `get_customer`. The system prompt clearly instructs it to verify identity first. What is the most reliable fix?
A) Rephrase the system prompt with stronger language about verification requirements.
B) Add 5 few-shot examples showing correct verification-first sequences.
C) Implement a programmatic prerequisite that blocks `lookup_order` and `process_refund` until `get_customer` returns a verified customer ID.
D) Add a pre-processing classifier that detects when verification is needed.
Answer
**C)** Programmatic enforcement provides deterministic guarantees. Prompt-based approaches (A, B) have non-zero failure rates. A classifier (D) adds complexity without guaranteed compliance.Answer
**B)** Tool descriptions are the primary mechanism for tool selection. Expanding them with clear differentiation is the highest-leverage, lowest-effort fix.Answer
**B)** Explicit decomposition into distinct items with parallel investigation and unified synthesis ensures complete coverage.Answer
**B)** Trimming verbose tool outputs to relevant fields prevents disproportionate context consumption.Answer
**B)** PostToolUse hooks should normalize heterogeneous data formats to a consistent format before model processing.Answer
**B)** Tool call interception hooks provide deterministic enforcement of business rules.Answer
**B)** Explicit customer requests for human agents must be honored immediately.Answer
**B)** Sentiment != complexity. Explicit criteria (policy gaps, customer explicit requests, inability to progress) are more reliable than sentiment-based triggers.Answer
**B)** User-level config is not shared via version control. Project-level config (.claude/CLAUDE.md or root CLAUDE.md) is needed for team-wide standards.Answer
**A)** Project-scoped commands in `.claude/commands/` are version-controlled and available to all developers automatically.Answer
**B)** Path-specific rules with glob patterns apply conventions to files by type regardless of directory location, which is ideal for test files spread throughout the codebase.Answer
**B)** Plan mode is designed for complex tasks with large-scale changes, multiple valid approaches, and architectural decisions.Answer
**B)** Concrete examples are the most effective way to communicate transformations when prose produces inconsistent results.Answer
**B)** `.claude/rules/` is designed for organizing topic-specific rule files as an alternative to a monolithic CLAUDE.md.Answer
**B)** `context: fork` runs the skill in isolation, preventing verbose output from consuming main conversation context.Answer
**B)** `@import` enables selective inclusion of relevant standards files per package while avoiding duplication.Answer
**B)** The coordinator's decomposition directly caused the coverage gap by only assigning visual arts subtopics to subagents.Answer
**B)** Parallel execution via multiple Task calls in a single response, followed by synthesis after both complete.Answer
**B)** Structured error context enables intelligent coordinator recovery decisions.Answer
**B)** A scoped verification tool handles the 85% common case while preserving the coordination pattern for complex cases. This follows the principle of least privilege.Answer
**B)** Structured claim-source mappings are the architectural fix. Prose-based approaches lose information through summarization.Answer
**C)** Preserve both values, annotate the conflict, and let readers evaluate the sources.Answer
**B)** `allowedTools` must explicitly include "Task" for a coordinator to invoke subagents.Answer
**B)** Partitioning scope at assignment time is more efficient than post-hoc deduplication.Answer
**B)** Grep is the built-in tool for content search -- searching file contents for patterns like function names.Answer
**C)** Glob is for file path pattern matching -- finding files by name/extension patterns.Answer
**B)** When Edit cannot find unique anchor text, Read + Write is the documented fallback for reliable file modification.Answer
**B)** Vague MCP tool descriptions cause the agent to default to familiar built-in alternatives. Enhanced descriptions fix this.Answer
**B)** Project-scoped `.mcp.json` with environment variable expansion enables shared config with individual credentials.Answer
**B)** User-scoped `~/.claude.json` for personal/experimental servers that should not affect teammates.Answer
**B)** The `-p` (or `--print`) flag is the documented way to run Claude Code in non-interactive mode for CI/CD.Answer
**B)** `--output-format json` with `--json-schema` produces machine-parseable structured output.Answer
**B)** Batch API is for latency-tolerant workloads only. Pre-merge checks are blocking and need real-time responses.Answer
**B)** Multi-pass review (per-file + integration) addresses attention dilution in large reviews.Answer
**B)** Self-review in the same session is limited by retained reasoning context bias.Answer
**B)** Including prior findings prevents duplicate reporting while maintaining awareness of all issues.Answer
**B)** Context awareness of existing tests prevents duplication at generation time.Answer
**B)** CLAUDE.md is the mechanism for providing project context to CI-invoked Claude Code.Answer
**B)** Tool use with JSON schemas eliminates syntax errors but cannot prevent semantic errors (values in wrong fields, logical inconsistencies).Answer
**B)** Nullable fields allow the model to signal "not found" rather than fabricating values to satisfy required constraints.Answer
**B)** Retry-with-error-feedback gives the model specific information about what went wrong.Answer
**B)** Retries cannot conjure information that does not exist in the source document.Answer
**B)** `"any"` guarantees a tool call while letting the model choose the correct extraction schema for the document type.Answer
**B)** Accuracy must be validated per document type and per field before automating. Aggregate metrics mask segment-specific failures.Answer
**B)** Stratified sampling provides continuous monitoring for quality degradation and novel error patterns.Answer
**B)** The "other" + detail string pattern provides extensible categorization.Answer
**B)** Prompt refinement on a sample before scaling reduces iterative resubmission costs.Answer
**B)** Targeted resubmission with appropriate modifications (chunking for context issues, simple retry for transient errors).Answer
**B)** Self-correction validation that extracts both values and flags discrepancies preserves accuracy while alerting to document-level issues.Answer
**B)** Few-shot examples showing correct handling of varied document structures (the specific challenge) enable the model to generalize to informal formats.