CCAR-F Practice Test 13 — 15 Questions

Question 1

An architect is configuring a Claude agent that must decide when to search an internal knowledge base versus when to answer directly from its own knowledge. The team wants the agent to retrieve only when the question references internal policies, products, or procedures. Which configuration most directly supports this decision?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Set tool_choice to auto and provide a well-described search tool so Claude decides per turn whether retrieval is needed.. Selective retrieval is achieved by letting the model decide per turn using tool_choice auto plus a precise tool description. The description tells Claude what content the tool covers, so it invokes search for internal policy, product, or procedure questions and answers directly otherwise. Forcing tool use, disabling tools, or gating on a keyword all remove that contextual judgment and produce either over-retrieval or missed retrieval.

Answer

Register the search tool but instruct Claude to call it only when the user explicitly types the word search.

Answer

Disable tools entirely and rely on a large system prompt containing the most frequently asked policy excerpts.

Answer

Set tool_choice to any and register only the knowledge base search tool so the model must always retrieve before answering.

Question 2

An architect is designing a multi-turn contract-review assistant. During long sessions, earlier redline decisions get contradicted in later turns because the model loses track of prior conclusions. The team cannot shorten sessions. Which approach best preserves decision consistency across the full session?

Accepted Answer

Maintain a structured running summary of decisions and inject it into each new turn alongside the most recent exchange.. The contradiction pattern stems from decision state being implicit in a long transcript. A structured running summary makes prior redline conclusions an explicit, compact artifact that is re-injected every turn, so consistency no longer depends on the model surfacing distant text. Window size, higher randomness, and user restatement all fail to externalize and enforce that state.

Answer

Raise the temperature so the model explores alternative interpretations and avoids repeating itself.

Answer

Rely on the model's extended context window to retain all prior turns without additional structure.

Answer

Ask the user to restate all prior decisions at the start of every turn.

Question 3

A developer wants Claude to summarize a long support thread and return the result as a JSON object with a single field "summary". They want to minimize the chance that Claude adds commentary outside the JSON. Which is the most direct Anthropic-recommended technique to apply?

Accepted Answer

Prefill the assistant turn with `{"summary": "` so Claude continues directly with the summary text and closing quote.. Prefilling the assistant turn with `{"summary": "` constrains Claude to continue inside the JSON string value, so the response cannot contain commentary before the object or wrap it in markdown. Instructions, stop sequences, and sampling parameters do not structurally enforce the format. The prefill is the most direct Anthropic-recommended technique for this single-field JSON output.

Answer

Use the `stop_sequences` parameter set to '}' to end generation after the JSON object.

Answer

Set `temperature` to 0 and `top_p` to 1 to make the output more deterministic.

Answer

Add the phrase 'Return only JSON' to the end of the user message.

Question 4

A legal team uses Claude to summarize case files. They require that the summary never includes personally identifiable information (PII) such as names, addresses, or phone numbers. The case files are lengthy and contain PII in various formats. Which approach is most reliable to ensure PII is not included in the summary?

Accepted Answer

Use a named entity recognition (NER) tool to detect and redact PII from the case files before sending them to Claude.. The most reliable way to prevent PII in summaries is to remove it before Claude processes the text. A dedicated NER tool can be configured for high recall and deterministic redaction, ensuring that no PII reaches the model. System prompt instructions and fine-tuning are probabilistic and may miss instances, while manual review is impractical at scale.

Answer

Fine-tune Claude on a dataset of legal documents with PII already redacted.

Answer

Ask Claude to output the summary in a structured format with a separate field for PII, then manually review it.

Answer

Instruct Claude in the system prompt to redact all PII from the summary.

Question 5

An architect is designing an MCP server that will be used by multiple client applications, some of which support only a subset of MCP capabilities. The architect wants tools to remain usable across clients without requiring per-client server forks. Which design principle should guide the server implementation?

Accepted Answer

Declare server capabilities during initialization and implement graceful degradation so tools work with the features each client actually supports.. MCP is built around capability negotiation during initialization, so a server should declare what it supports and design tools to degrade gracefully when a client lacks optional features. This allows one server to serve heterogeneous clients without forks or identity-based branching. Lowest-common-denominator design and mandatory capability requirements both undermine that goal.

Answer

Advertise every capability the server implements and require clients to support all of them before connecting.

Answer

Detect the client's product name and branch behavior with client-specific code paths for each known application.

Answer

Expose only the lowest common denominator of features so every client behaves identically regardless of its capabilities.

Question 6

An MCP server exposes an `execute_task` tool that can run arbitrary shell commands in a sandbox. During evaluation, the architect observes that the model sometimes calls the tool with a command that succeeds but produces no useful output, then calls it again with a nearly identical command, repeating several times. Which design change most directly addresses this loop behavior?

Accepted Answer

Return a structured result that distinguishes success with empty output from failure, and include guidance on the next valid action.. Loops typically arise when the model cannot tell the difference between an uninformative success and a failure, so it retries in the hope of a better result. A structured result that flags empty output as a distinct outcome, together with guidance about acceptable next actions, breaks that cycle. Input schema changes and tool splitting do not supply the missing feedback.

Answer

Add a required `command` parameter to the schema so the model must always supply a command string.

Answer

Split `execute_task` into separate tools for read-only and mutating commands to reduce accidental side effects.

Answer

Increase the tool's timeout so slow commands have more time to return useful output.

Question 7

A research agent must query three independent sources (a database, a web search API, and a file store) to assemble a report. Latency matters, and the sources do not depend on each other. Which orchestration pattern should you use with Claude's tool use to minimize wall-clock time?

Accepted Answer

A single completion in which the model emits three `tool_use` blocks, followed by one user turn containing three `tool_result` blocks.. Parallel tool use lets Claude emit several `tool_use` blocks in one response. The application runs the independent calls concurrently and returns all results in a single user turn, each tied to its `tool_use_id`. For a report assembled from unrelated sources, this overlaps latency and avoids serial model round trips.

Answer

Three separate agents, each with one tool, coordinated by a fourth router agent.

Answer

Sequential single-tool turns, where each result is returned before the next tool is requested.

Answer

Set `tool_choice` to `{"type": "tool", "name": "web_search"}` so the fastest tool runs first.

Question 8

A developer is working on a legacy monorepo where the Claude Code CLI's auto-generated commit messages keep mixing unrelated changes from multiple microservices. To improve this, the developer wants Claude Code to read a project-specific file that describes the commit message format and scope conventions. Which file should the developer create at the repository root to provide this persistent, team-shared guidance?

Accepted Answer

CLAUDE.md. CLAUDE.md is the documented project memory mechanism that Claude Code reads automatically at the start of every session, making it the correct place for repository-wide conventions such as commit message formats and scope naming. Tool-specific config files and Git templates are not consumed as model context.

Answer

commitlint.config.js

Answer

settings.json

Answer

.gitmessage

Question 9

A developer is building a chatbot that must never reveal internal system instructions. During testing, a user asks, 'What is your system prompt?' and Claude begins to repeat the instructions verbatim. Which approach is most effective to prevent this leakage?

Accepted Answer

Add a rule in the system prompt instructing Claude not to reveal the system prompt.. The most direct and effective way to prevent system prompt leakage is to include an explicit instruction in the system prompt itself, telling Claude not to reveal it. This leverages the model's training to follow system-level directives. While additional layers like classifiers can help, they are secondary. Temperature and obfuscation do not address the model's willingness to disclose.

Answer

Obfuscate the system prompt by encoding it in base64.

Answer

Use a separate model to classify user queries and block any that ask about system prompts.

Answer

Set the temperature to 0 to make Claude's responses deterministic and less likely to deviate.

Question 10

An architect is integrating an MCP server with a legacy inventory system that exposes a SOAP API. The MCP server must translate tool calls into SOAP requests. The SOAP API is slow and occasionally returns faults. Which design choice best ensures that the model can recover gracefully from SOAP faults without hallucinating inventory data?

Accepted Answer

Catch SOAP faults in the MCP server and return a structured error object with a clear message type, such as 'inventory_unavailable', and no data fields.. Catching SOAP faults and returning a structured error object with no data fields prevents the model from fabricating inventory data. The clear error type allows the model to communicate the failure or trigger a retry. This design keeps the tool's success and error paths distinct, which is critical for trustworthy behavior.

Answer

Configure the tool to return an empty array on SOAP faults so the model sees an empty inventory.

Answer

Set a very long timeout on the SOAP client so the model waits until the API eventually responds, and if it times out, return a generic 'error' string.

Answer

Return the raw SOAP fault XML as the tool result and let the model interpret it.

Question 11

An architect is designing an MCP server that wraps a ticketing system. The design must let Claude call tools safely and must keep tool definitions maintainable as the API grows. Which two practices should the architect apply? (Choose two.)

Accepted Answer

Return structured, machine-readable errors from tool handlers so the model can correct invalid arguments on retry.. Safe, maintainable MCP tooling rests on precise, versioned schemas that tell the model exactly what each tool accepts, plus structured error results that let the model recover from invalid arguments. Together these raise call accuracy without widening privileges. Generic passthrough tools, over-broad service accounts, and prompt-embedded descriptions each undermine either safety or maintainability, and none of them scale cleanly as the ticketing API grows.

Answer

Expose one generic `call_ticket_api` tool that accepts an arbitrary endpoint path and HTTP method as string arguments.

Answer

Inline all tool descriptions as long prose in the system prompt instead of in the tool definitions themselves.

Answer

Grant the server's service account full administrative scope on the ticketing system so any tool call will succeed.

Question 12

You are architecting a Claude-based system that answers questions about internal policies. The knowledge base is updated frequently, and you need to ensure answers are accurate and traceable to source documents. Which TWO practices should you implement to maximize reliability and auditability? (Choose two.)

Accepted Answer

Instruct Claude to include the source document title and section in its answer.. RAG grounds answers in current policy documents, and instructing Claude to cite sources makes answers traceable. Together, they ensure accuracy and auditability. Fine-tuning is inflexible for frequent updates, low temperature does not guarantee correctness, and caching can serve outdated information.

Answer

Fine-tune Claude on the entire policy corpus to embed the knowledge directly into the model weights.

Answer

Cache all previous answers and serve them for repeated questions to reduce API calls.

Answer

Set a low temperature to make answers more deterministic and consistent.

Question 13

An architect is defining the input schema for an MCP tool that schedules a meeting. The schema currently accepts a free-form string for the start time, and the model frequently sends ambiguous values such as "next Tuesday afternoon." Which change best improves reliability?

Accepted Answer

Change the start time field to a JSON Schema string with format "date-time" and add a separate timezone field with an enum of IANA time zone names.. JSON Schema formats and enums turn an ambiguous free-form field into a precise contract the model can satisfy and the server can validate. A date-time format standardizes the timestamp, while an IANA time zone enum removes offset and daylight saving ambiguity. This reduces invalid calls and eliminates the need for fragile natural-language parsing on the server.

Answer

Add a natural-language description telling the model to always use ISO 8601 format.

Answer

Accept the free-form string and have the server parse it with a natural-language date library before scheduling.

Answer

Replace the start time with separate integer fields for hour and minute and let the client supply the date.

Question 14

An architect is onboarding a new developer to a repository that uses Claude Code. The developer wants to understand how the tool determines which project-specific instructions to load automatically at the start of a session. The repository already contains a CLAUDE.md file at the root. What is the correct behavior of Claude Code regarding this file?

Accepted Answer

Claude Code automatically reads the CLAUDE.md file from the project root and includes its contents as context for the session.. The CLAUDE.md file at the project root is automatically ingested as context when Claude Code starts in that directory, which is how teams distribute conventions and architectural notes. No manual import or settings.json entry is required, and the file is not merely a write-only scratchpad. This automatic loading is what makes project memory reliable across sessions.

Answer

Claude Code only reads CLAUDE.md if it is referenced inside a .claude/settings.json configuration file.

Answer

Claude Code ignores CLAUDE.md unless the developer explicitly runs a command to import it at the beginning of each session.

Answer

Claude Code treats CLAUDE.md as a scratchpad that is only written to and never read as context during a session.

Question 15

An architect is designing an MCP server that exposes several tools for a document management system. The tools include 'create_document', 'update_document', 'delete_document', and 'search_documents'. The architect wants to minimize the risk of the model accidentally invoking destructive tools during exploratory or read-only tasks. Which two design choices should the architect implement? (Choose two.)

Accepted Answer

Implement a human-in-the-loop approval step for destructive tool invocations, where the MCP server pauses and requests confirmation from the user before executing.. Separating read-only and write tools into different servers, and requiring human approval for destructive actions, are both preventive controls. They ensure that during exploratory tasks, destructive tools are either not exposed or cannot execute without explicit human consent. These measures directly reduce the likelihood of accidental invocation.

Answer

Add a 'confirm' boolean parameter to each destructive tool that the model must set to true, and document in the tool description that it should only be set after user approval.

Answer

Log all tool invocations and review the logs periodically to detect accidental destructive actions after they occur.

Answer

Rename destructive tools with a prefix like 'admin_' to signal their risk, and rely on the model to avoid them unless instructed.