CClaudeMentor

Architect practice questions — set 1

10 free questions for the Claude Certified Architect – Foundations exam (CCAR-F). The correct option and its reasoning appear under each question — read them through, or cover the answer block and treat it as a timed self-check.

Question 1
Claude Code Configuration & Workflows

A developer includes prior review findings in context when re-running reviews after new commits. How should a developer modularize a large CLAUDE.md file?

  1. A.Keep all instructions in a single monolithic file for simplicity and centralization
  2. B.Duplicate the entire file into every subdirectory for local access
  3. C.Use @import syntax to reference external standards files and split into topic-specific rule files
  4. D.Remove all configuration and rely on default Claude behavior instead

Correct answer: C

@import and topic-specific rule files keep CLAUDE.md modular and maintainable.

Question 2
Agentic Architecture & Orchestration

A system requires subagents to include metadata like dates and source locations in structured outputs. Which sampling method best measures error rates in high-confidence extractions?

  1. A.Stratified random sampling to detect novel error patterns across document types and fields
  2. B.Reviewing only extractions where the model reported lowest confidence scores
  3. C.Sampling only the most common document types to maximize review efficiency
  4. D.Waiting for users to report errors before measuring any error rates

Correct answer: A

Stratified sampling detects errors that aggregate metrics might mask in specific segments.

Question 3
Claude Code Configuration & Workflows

A project uses .claude/rules/ with paths fields containing glob patterns for conditional activation. Which approach best reduces duplicate test suggestions from Claude Code?

  1. A.Provide existing test files in context so generation avoids already-covered scenarios
  2. B.Limit Claude to generating only one test per file regardless of coverage needs
  3. C.Disable test generation entirely and require manual test writing by developers
  4. D.Remove all existing tests and let Claude generate a completely new test suite

Correct answer: A

Providing existing tests in context prevents duplicate scenario suggestions.

Question 4
Claude Code Configuration & Workflows

A developer manages context effectively during large codebase exploration with extended sessions. Why are self-reported confidence scores unreliable for escalation decisions?

  1. A.Confidence scores only work for simple queries and fail on multi-concern requests
  2. B.Confidence scores are always perfectly calibrated and reliable for all decision types
  3. C.The model cannot generate confidence scores for any output it produces
  4. D.The agent is already incorrectly confident on hard cases where escalation is actually needed

Correct answer: D

LLM self-reported confidence is poorly calibrated, especially on complex cases requiring escalation.

Question 5
Prompt Engineering & Structured Output

A team analyzes accuracy by document type and field segment before automating high-confidence extractions. How should coverage gaps be communicated in synthesis output?

  1. A.Mark all findings as tentative when any single source was unavailable during research
  2. B.Omit any mention of gaps and present findings as comprehensive regardless of coverage
  3. C.Randomly flag some findings as unsupported to encourage additional verification
  4. D.Include coverage annotations indicating which findings are well-supported versus which areas have gaps

Correct answer: D

Explicit coverage annotations enable informed decisions about output reliability.

Question 6
Claude Code Configuration & Workflows

A system manages conversation context to preserve critical information across long customer support interactions. Why are self-reported confidence scores unreliable for escalation decisions?

  1. A.The agent is already incorrectly confident on hard cases where escalation is actually needed
  2. B.The model cannot generate confidence scores for any output it produces
  3. C.Confidence scores only work for simple queries and fail on multi-concern requests
  4. D.Confidence scores are always perfectly calibrated and reliable for all decision types

Correct answer: A

LLM self-reported confidence is poorly calibrated, especially on complex cases requiring escalation.

Question 7
Tool Design & MCP Integration

A system requires tools to support cancellation of in-flight operations to free resources. Which pattern best handles both synchronous and asynchronous tool calls?

  1. A.Unified interface with async-first design and synchronous blocking wrappers
  2. B.Completely separate codebases with no shared logic between sync and async paths
  3. C.Async-only design rejecting all synchronous consumers without exception
  4. D.Synchronous-only design with manual thread spawning for async-like behavior

Correct answer: A

Async-first with sync wrappers provides flexibility while maintaining clean internals.

Question 8
Prompt Engineering & Structured Output

An architect designs caching for tool outputs with varying freshness requirements. What validation strategy is most robust for complex nested JSON payloads?

  1. A.Trusting the client to send valid JSON and skipping all server-side checks
  2. B.Recursive schema validation with type coercion, custom format checkers, and strict mode
  3. C.Basic string matching on the raw JSON text without any structural parsing
  4. D.Validating only the top-level keys and ignoring all nested structures entirely

Correct answer: B

Recursive validation with type coercion and format checkers handles complex nested structures safely.

Question 9
Context Management & Reliability

A team handles conflicting statistics from credible sources by annotating conflicts with source attribution. What is the lost in the middle effect in long context processing?

  1. A.Models reliably process information at the beginning and end but may omit middle section findings
  2. B.Models lose all context after exactly 50 percent of the input length is reached
  3. C.Middle sections are always processed more accurately than beginning or end sections
  4. D.The effect only occurs with code context and not with natural language documents

Correct answer: A

The lost in the middle effect means key information in middle sections may be overlooked.

Question 10
Tool Design & MCP Integration

A tool supports streaming responses for long-running operations producing incremental results. Which rate limiting strategy respects upstream API quotas most effectively?

  1. A.Token bucket with adaptive refill rates based on quota headers and burst allowance
  2. B.Randomly throttling requests without any predictable or fair allocation method
  3. C.Fixed request count per minute with no consideration of upstream feedback signals
  4. D.Allowing unlimited requests and relying on upstream to reject excess traffic

Correct answer: A

Token buckets with adaptive refill respect upstream quotas while allowing bursts.

Keep going

Try the interactive Flash sets free, or open the full Architect track for Focused, Foundation, Finals and timed exams with a readiness score.

Practice interactivelySee pricing