Claude Certified Architect Foundations Practice Exam Questions and Answers – Part 24/24

Answer: C. Append the assistant tool use turn, then add an immediate user turn with the matching tool result block.

The application must preserve the assistant message that requested the tool, then send the result back as a user message. The tool result must reference the original tool use ID so the model can connect the output to the requested action.

Answer: B. Use a maintained Jira MCP server for standard issue operations, and build a focused custom MCP server for the proprietary workflow.

This approach reuses an existing integration for common capabilities while limiting custom code to the team-specific behavior outside servers cannot know. Building one monolithic custom server increases your API maintenance burden without any real integration benefits.

Answer: C. Define the investigation goal, evidence requirements, output format, tool/source guidance, task boundaries, and effort budget heuristics.

This design gives the coordinator enough structure to delegate useful work without freezing the investigation into a fixed path. Assigning hardcoded subagents or rigid file sequences destroys the dynamic orchestration needed for unfamiliar codebase questions.

Answer: D. Enter plan mode, give one detailed request covering both defects, affected areas, constraints, and verification checks before approving edits.

A single detailed prompt gives Claude Code the shared context needed to reason about the coupled change. Splitting the defects into separate sessions creates inconsistent edits because each session makes different assumptions about shared interfaces.

Answer: C. Return the validator's field paths and citation failures with the original note, then retry and route unresolved cases to review.

This directly addresses validation and retry quality by giving Claude specific, actionable feedback about what failed. Generic retry prompts without field paths or validator context fail to correct citation formatting issues.

Answer: A. Split the concerns into parallel extraction tasks, pass shared packet context to each task, then merge and validate the unified JSON.

This matches the separability of the work, allowing each extraction concern to run independently using shared source context. Processing sequentially or increasing the output budget fails to resolve the core latency and coverage degradation.

Answer: A. Resume the baseline session ID with fork_session enabled for each strategy and run each investigation in a new session.

Resuming with fork_session creates separate sessions that start with a copy of the original session history while leaving the original unchanged. Manual duplication or continuing in the same thread fails to isolate the reasoning history properly.

Answer: D. Assign each subagent a distinct document group or extraction concern, provide needed context, and require concise findings for synthesis.

Using subagents for independent investigations requires clear scope boundaries to reduce duplicate work. Requiring concise return formats limits how much subagent output flows back into the main conversation, preventing synthesis bloat.

Answer: A. Create .claude/skills/refund-audit/SKILL.md with supporting files referenced from it, and set disallowed-tools for Write, Edit, MultiEdit, NotebookEdit, and Bash.

A project skill placed in the .claude/skills directory is version-controlled with the repository and automatically generates the corresponding slash command. Using disallowed-tools explicitly revokes specific capabilities like file editing and shell execution, which is safer than relying solely on an allowed-tools whitelist.

Answer: C. Read the current file contents, assemble the full intended file, and use Write to overwrite that file.

Using the Write tool to overwrite the entire file bypasses the strict unique substring matching required by the Edit tool. After reading the file to ensure the latest contents are in context, Write provides a reliable mechanism to commit the intended changes without risking the wrong replacement.

Answer: D. Ask Claude Code to make the small change directly and run the specified test.

Direct execution is optimal for small, well-scoped changes where the expected diff is clearly understood and easily described. Plan mode adds unnecessary overhead here, and including the specific unit test gives Claude Code a concrete verification signal to confirm the task succeeded.

Answer: C. Replace analyze_document with a few non-overlapping tools for distinct workflows, each with clear descriptions, JSON Schema inputs, and minimal stable result fields.

Breaking the monolithic tool into non-overlapping, distinct workflows enforces a clear input and output contract for each task. Avoid creating highly overlapping tools, as tool sprawl degrades the model's selection accuracy and makes the agent harder to orchestrate.

Answer: B. Have the coordinator scout initial findings first, then create focused subagent assignments from the discovered work list.

Having the coordinator scout initial findings first allows the system to adapt its investigation plan. The coordinator can then discover relevant branches before fanning out work, which provides better coverage than prewriting a larger fixed list of subtasks.

Answer: D. Include the relevant findings and source metadata directly in the synthesis subagent's Agent prompt.

A subagent operates with a fresh context window and cannot see previous tool results or parent history. The coordinator must explicitly pass all required source metadata and findings directly inside the synthesis subagent prompt to preserve citation provenance.

Answer: A. Sample high-confidence records within each document type and critical field, then track reviewed error patterns by segment.

Segmenting your manual review sample across different document types and fields exposes hidden errors in high-confidence records. Relying strictly on aggregate random sampling or ignoring high-confidence outputs entirely masks rare failure patterns.