CCAR-F Practice Exam 2 — 20 Questions

Question 1

You are building a document-processing pipeline with the Anthropic Messages API. Claude must extract a fixed set of fields from each document and return them as JSON. In production, you observe two problems: (1) some responses include explanatory prose before the JSON, and (2) some responses omit required fields when the document is ambiguous. You need to make the pipeline more reliable. (Choose two.)

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

Accepted Answer

Prefill the assistant turn with `{` to force the response to begin as a JSON object.. The two problems are a formatting issue and a completeness issue. Prefilling the assistant turn with `{` structurally prevents prose before the JSON, and specifying required fields with a null fallback ensures keys are always present even when the document is ambiguous. Raising max_tokens, lowering temperature, or adding a stop sequence do not enforce format or completeness and can even truncate valid output.

Answer

Add a stop sequence of '

' so Claude stops as soon as it finishes the JSON object.

Answer

Set `temperature` to 0 to make the output deterministic and therefore always complete.

Answer

Increase `max_tokens` to the model's maximum so the response has room for all fields.

Question 2

A developer is prompting Claude to classify 5,000 customer support tickets into one of six fixed categories. The first 200 classifications look correct, but the developer notices that category labels come back with inconsistent capitalization and occasional trailing punctuation. What is the simplest change that makes the labels uniform and directly usable as database keys?

Accepted Answer

Provide the six allowed labels in a fixed, lowercase form and instruct Claude to return exactly one of those labels with no other characters.. Label uniformity is achieved by enumerating the allowed values in their canonical form and restricting the output to exactly one of them. This closes off the model's latitude to change case or append punctuation, so returned strings match database keys directly. Downstream normalization, higher temperature, and reasoning text all either mask the problem or make the output harder to consume reliably.

Answer

Add a post-processing step that lowercases and strips punctuation from each label before insertion.

Answer

Raise the temperature so Claude explores more label phrasings and picks the most natural one.

Answer

Ask Claude to explain its reasoning for each ticket and include the category somewhere in the explanation.

Question 3

An MCP server exposes a `query_orders` tool whose input schema includes a `filters` object with several optional properties. The model intermittently omits required nested fields inside `filters`, and the server returns a generic 400 error. The architect wants the model to reliably produce schema-valid arguments and recover from validation failures. Which combination of MCP features should the architect implement?

Accepted Answer

Use `inputSchema` with `required` arrays and property descriptions, and return `isError: true` with a structured message naming the missing field.. The MCP contract has two halves: the input schema that declares required nested properties, and the tool result that reports validation failures. Declaring `required` arrays with descriptive properties guides argument construction, while returning `isError: true` with a structured message that names the missing field gives the model a concrete correction target. Permissive schemas, plain text errors, and tool fragmentation each leave one half of the loop open.

Answer

Return a plain text error string from the tool handler and instruct the model in the system prompt to always include all fields.

Answer

Set `inputSchema` to a permissive object and rely on the server to coerce missing fields to defaults silently.

Answer

Split `query_orders` into one tool per filter field so each tool has a single scalar input.

Question 4

A team is exposing an internal inventory database to Claude through an MCP server. They want the model to retrieve a specific item by its SKU, and they want the tool contract to be unambiguous so the model rarely mis-calls it. Which tool definition best matches this goal?

Accepted Answer

A `query_inventory` tool whose input schema has a single required string field named `sku`.. The strongest tool contract for a keyed lookup is the one with the fewest, most specific required inputs. A single required `sku` string tells the model exactly what to provide and nothing more, which minimizes ambiguity and downstream errors. Broader or freer schemas trade precision for flexibility that this scenario does not need.

Answer

A `query_inventory` tool that takes no parameters and returns the full inventory table.

Answer

A `query_inventory` tool with optional `sku`, `name`, `category`, and `warehouse` fields, none required.

Answer

A `query_inventory` tool that accepts a free-form SQL `where` clause string.

Question 5

A platform team maintains a shared monorepo. A developer runs Claude Code and wants the tool to automatically treat every file under `/infra/terraform/` as read-only for all engineers, without relying on each person to remember a flag. Where should this restriction be declared so it applies to everyone who opens the repo?

Accepted Answer

In the project's `.claude/settings.json` using a `permissions.deny` rule targeting the Terraform path.. Shared enforcement belongs in the committed project settings file, because it is version-controlled and applied to every session opened in that repository. A deny rule scoped to the Terraform directory stops edits deterministically, independent of individual developer setup. Per-user settings, Git ignore files, and memory documents all fail to give the team a uniform, repository-level guarantee that survives onboarding and local configuration drift.

Answer

In the repo's `.gitignore` by adding the `/infra/terraform/` directory.

Answer

In the `CLAUDE.md` memory file with a sentence asking Claude not to modify Terraform files.

Answer

In each engineer's `~/.claude/settings.json` with a matching deny rule.

Question 6

A team runs a Claude agent that orchestrates a multi-step data pipeline. The agent uses the Messages API and passes the full conversation history on every request. After the history grows beyond roughly 180,000 tokens, the agent begins to fail intermittently with context length errors even though the configured model supports 200,000 tokens. Logs show the agent appends each tool result verbatim, including large JSON payloads from a database tool. Which change best resolves the failures while preserving the agent's ability to reason over prior steps?

Accepted Answer

Summarize or truncate older tool_result blocks before appending new turns, keeping only the fields the agent needs from large database payloads.. Context length failures stem from unbounded accumulation of verbatim tool results, not from the model's advertised maximum. Compressing or summarizing older tool_result blocks, and trimming large JSON payloads to only relevant fields, lowers input tokens per request while preserving the reasoning trail. Caching and max_tokens tuning affect cost or output reservation, not the input footprint, so they cannot resolve the overage.

Answer

Reduce max_tokens on each request so the combined prompt plus completion fits under the model limit.

Answer

Enable prompt caching on the system prompt and tool definitions so repeated prefixes are billed and processed more efficiently.

Answer

Move the entire conversation history into a single user message and ask Claude to ignore irrelevant portions.

Question 7

A developer wants a single Claude Code slash command available only to their own user account on a shared Linux workstation, not to other engineers who log into the same machine. Where should the command's Markdown file be stored?

Accepted Answer

In ~/.claude/commands in the developer's home directory. Slash commands are Markdown files discovered from command directories, and scope follows location: files under the user's home .claude/commands are personal, while files under the project's .claude/commands are shared through the repository. To keep a command visible only to one developer on a multi-user workstation, it must be stored in that developer's home directory command folder rather than in any project-level location.

Answer

In the project's .claude/commands directory at the repository root

Answer

In a CLAUDE.md file committed at the repository root

Answer

In the settings.json file under the project's .claude directory

Question 8

A production agent uses the Claude Messages API with tool use to manage cloud resources. During a run, Claude returns a response whose stop_reason is "tool_use" containing three tool_use content blocks that are fully independent of one another. The orchestrator loops and processes them one at a time, executing each tool and sending a separate user message containing a single tool_result block before each next call. Runs that used to complete in seconds now take minutes and occasionally exceed the request timeout. What is the most likely architectural cause of the slowdown?

Accepted Answer

Each tool_result block was sent in its own user message, so Claude had to re-ingest the full conversation and re-plan after every individual tool instead of receiving all three results together in one user message.. When a single assistant turn requests several independent tools, the client should execute them and return every matching tool_result block in one user message keyed by tool_use_id. Returning results piecemeal forces additional model calls that each reprocess the conversation, multiplying latency and token cost. Batching the results restores one planning step per assistant turn.

Answer

The orchestrator should have set the max_tokens parameter higher before executing the tools, because low token limits cause the API to throttle subsequent tool_result submissions.

Answer

The stop_reason value "tool_use" indicates the model is waiting for the client to switch to the streaming endpoint before any tool can be executed.

Answer

Independent tool_use blocks must be executed in the exact order they appear in the assistant message, and any out-of-order execution invalidates the context window.

Question 9

A team wants Claude to call an internal `create_invoice` tool only after the user has explicitly approved the line items. Which mechanism should the architect use to enforce this gate?

Accepted Answer

Withhold the `create_invoice` tool definition from the request until the orchestrator records explicit user approval, then include it in a subsequent turn.. Enforce business-critical gates in the orchestrator, not in prompts. If the create_invoice tool is absent from the request, the model cannot emit a tool_use for it. Once the host records explicit approval, the tool definition is added and the conversation continues, giving the architect deterministic control over a sensitive side effect.

Answer

Use a system prompt instructing Claude to always ask for confirmation before calling `create_invoice`.

Answer

Set `tool_choice` to force a specific tool so Claude cannot pick anything else.

Answer

Set `tool_choice` to `auto` so Claude decides when to call `create_invoice`.

Question 10

A support agent has two tools: `search_knowledge_base` and `escalate_ticket`. For simple questions it should answer directly; for unresolved issues it should escalate. Which `tool_choice` setting best supports this behavior?

Accepted Answer

`{"type": "auto"}`. The scenario needs the model to choose between answering directly and invoking one of two tools. The `auto` setting lets the model decide per turn, calling a tool only when the question requires it. Forcing any specific tool, or disabling tools, would remove one of the required behaviors.

Answer

`{"type": "none"}`

Answer

`{"type": "any"}`

Answer

`{"type": "tool", "name": "escalate_ticket"}`

Question 11

A document-processing agent reads contracts and must extract a list of payment milestones, each with an amount, currency, and due date. Downstream billing systems cannot tolerate hallucinated fields, and the extraction must be validated programmatically before anything is written to the database. The architect wants Claude to emit data that can be parsed and checked with certainty rather than relying on post-hoc regex over prose. Which approach best satisfies this requirement?

Accepted Answer

Define a tool whose input_schema describes the milestone array with required amount, currency, and due date fields, and force Claude to call it using tool_choice.. Tool definitions carry a JSON Schema in input_schema, and tool_choice can force the model to emit arguments conforming to that schema rather than free text. The client then validates the structured fields before persisting them, giving a deterministic contract that prompt wording alone cannot provide. This is the standard pattern for extraction feeding strict downstream systems.

Answer

Instruct Claude in the system prompt to reply only with JSON and add a reminder to double-check every field before answering.

Answer

Raise the temperature to encourage variety and then filter the results with a second Claude call that rewrites non-conforming records.

Answer

Use a very large max_tokens value and ask Claude to include the milestones in a markdown table so a parser can split on pipe characters.

Question 12

A research agent must answer questions that require looking up several unrelated facts, such as the population of three different cities. The architect observes that Claude frequently emits one search tool call, waits for the result, then emits the next, repeating until all facts are gathered. This serial pattern roughly triples latency versus what the workload should allow. Which change most directly enables the model to request all three independent lookups in a single assistant turn?

Accepted Answer

Add a system instruction that encourages Claude to batch independent lookups by requesting multiple tool calls in the same turn when the queries do not depend on each other.. Parallel tool use is a capability the model exercises when the calls are independent, and explicit guidance in the system prompt reliably encourages it. The client executes the returned tool_use blocks concurrently and submits all tool_result blocks together, cutting three sequential round trips down to one. No API parameter forces this; it is a prompting and orchestration practice.

Answer

Switch the search tool from a single-query parameter to an array parameter so one call can carry all three queries.

Answer

Set tool_choice to force a specific search tool on every turn so the model cannot choose to wait between calls.

Answer

Increase max_tokens so the model has room to write more tool_use blocks before the response is truncated.

Question 13

A developer wants to use Claude Code to perform a large refactor across multiple files. What is the most efficient workflow to ensure the model has the necessary context without manual file-by-file reading?

Accepted Answer

Provide a high-level description of the refactor and let the model use its search and read tools.. Efficiency in large-scale refactoring depends on the model's ability to discover and navigate the codebase. Rather than the user manually feeding files to the model, the best practice is to describe the scope of the change and allow the model to use its internal tools to explore the relevant dependencies.

Answer

Manually run 'cat' on every file in the project and paste the output into the prompt.

Answer

Use the /read-all command to force Claude to load every file into memory at once.

Answer

Copy the entire project into a single .txt file and upload it using the /upload command.

Question 14

An engineering team wants to customize Claude Code behavior across their enterprise repository by enforcing specific project-level guidelines and system prompts. Where should the lead architect place these configuration directives so that Claude Code automatically detects and applies them upon initialization in the workspace root?

Accepted Answer

Inside a CLAUDE.md file located directly in the repository root directory. Placing configuration guidelines in the CLAUDE.md file within the workspace root allows Claude Code to automatically load custom project instructions, codebase conventions, and behavioral constraints during startup. This mechanism ensures consistent AI assistance tailored precisely to the specific engineering practices and architectural patterns utilized across the development team's codebase.

Answer

Inside a .claude/config.json file in the user home directory

Answer

Within the package.json file under a dedicated claudeConfig key

Answer

As environment variables defined inside a local .env file

Question 15

Refer to the exhibit. The model fails to return accurate data when the user asks for metrics over a 'five minute' interval. What is the most likely cause?

Accepted Answer

The inputSchema lacks explicit constraints or a more descriptive instruction for the argument.. The tool definition requires a string input but does not specify the expected unit or format (e.g., ISO8601 or seconds as an integer). Because the model cannot infer that 'five minutes' must be converted to '300' seconds, it passes the raw input, causing the backend to fail. Improving the schema with a description specifying unit requirements is essential for robust tool interaction.

Answer

The tool name 'get_system_metrics' is not unique enough to be invoked correctly.

Answer

The MCP server is missing a required 'outputSchema' field in the definition.

Answer

The model is unable to handle more than one argument for an MCP tool.

Question 16

Which of the following describes the purpose of the 'resources' feature in MCP?

Accepted Answer

To provide the model with read-only access to external data entities.. MCP resources are designed for the model to read and utilize data from external sources, such as configuration files, logs, or documentation. Unlike tools, which are meant for taking actions, resources represent data objects. They allow the model to ingest context from the environment efficiently, which is a fundamental requirement for building agents that need to be aware of their surroundings to perform accurate task execution.

Answer

To allow the model to execute long-running background tasks on the server.

Answer

To allow the model to modify persistent settings in the server's configuration.

Answer

To maintain state between different MCP client sessions across the network.

Question 17

Refer to the exhibit. What is the correct next step for the application developer to ensure the agent continues its task?

Accepted Answer

Execute the function and return a tool_result block with the tool_use_id. The exhibit shows a standard tool_use response from the Messages API. To proceed, the developer must execute the 'get_stock_quote' function locally using the provided input, and then send a new request to Claude containing a message with the 'tool_result' block that references the unique ID 'toolu_01A2B3C4'.

Answer

Ask the user to manually verify the ticker symbol before proceeding

Answer

Send a new user message containing only the raw numeric stock price

Answer

Re-submit the original prompt with a higher temperature to get the price

Question 18

An agent tasked with data analysis repeatedly fails because the 'tool_result' it receives is too large for the context window, causing subsequent calls to truncate. What is the most effective architectural adjustment to resolve this?

Accepted Answer

Use the 'tool_result' to return a summary or a file path instead of the full raw data.. Large tool outputs can quickly consume the context window, leaving no room for the model's reasoning or future tool calls. Architects should implement a 'Summarize-or-Reference' pattern where the full data is stored externally and only a concise summary or a pointer is returned to the agent within the message history.

Answer

Switch to a model with a smaller context window to force more efficient coding.

Answer

Format the tool output as a series of multiple user messages to spread the load.

Answer

Request that Claude only uses the last 100 lines of the tool output in its reasoning.

Question 19

Which THREE strategies are effective for reducing hallucinations in RAG (Retrieval-Augmented Generation) systems?

Accepted Answer

Instruct the model to say 'I don't know' if the context does not contain the answer.. To reduce hallucinations in RAG, you must constrain the model to the provided context, encourage it to admit ignorance if the answer is missing, and verify its reasoning. These techniques transform the model from a creative writer into a grounded processor of provided facts. Architects must build these guardrails into the prompt to ensure the output is factually tethered to the retrieved source material, which is critical for trustworthy systems.

Answer

Increase the temperature to 1.5 to allow for more creative exploration.

Answer

Use a complex system prompt that describes the model as an 'Omniscient Expert'.

Question 20

When designing tools for a highly regulated environment, what is the most important architectural goal?

Accepted Answer

Providing complete observability and strict access controls for every tool call.. In regulated environments, auditability and deterministic control are the primary architectural drivers. Every tool invocation must be logged with its full context, including the user, input, and response. This provides a complete audit trail that is necessary for compliance. Coupled with least-privilege design, this ensures that every model interaction is traceable and restricted, meeting the strict security standards required by enterprise compliance frameworks like CCAR-F.

Answer

Maximizing the model's autonomy to perform complex, multi-step tasks.

Answer

Ensuring the model can access as much data as possible to be effective.

Answer

Optimizing for the lowest possible latency in every tool interaction.