CCAR-F Hard Questions — 25 Difficult Questions

Question 1

Refer to the exhibit. In a parallel tool-use scenario where Claude generates three such blocks in a single response, how should the orchestration layer handle the 'tool_result' messages to ensure Claude can correctly process the outputs?

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

Accepted Answer

Return a single 'user' message containing three 'tool_result' blocks with matching IDs. Parallel tool use requires that the orchestration layer executes all requested tools and returns a 'tool_result' for each one. Crucially, each result must include the unique 'tool_use_id' generated by the model. This allows Claude to correctly map the data from the external world back to the specific requests it made, maintaining the logical flow of the agent's reasoning.

Answer

Combine all three tool outputs into a single 'tool_result' block for efficiency

Answer

Send three separate API requests, each containing one tool result in sequence

Answer

Wait for the user to manually approve each tool result before sending them back

Question 2

An architect is designing an agentic workflow on the Anthropic Messages API where a single Claude model must first break a user goal into ordered subtasks, then execute each subtask with tools, and finally synthesize a final answer. The architect wants to minimize the number of API round trips while still giving the model a chance to observe each tool result before deciding the next action. Which orchestration approach best satisfies these requirements?

Accepted Answer

Use a single request with a stop sequence on the tool_use block, execute the requested tools, then send a follow-up user message containing the tool_result blocks and repeat until Claude returns a final text response.. The agentic tool-use loop is the correct orchestration pattern because it alternates model reasoning with real tool execution, letting Claude observe each tool_result before selecting the next action. This supports decomposition, execution, and final synthesis in one evolving conversation. Batching all calls, splitting into separate conversations, or simulating tool outputs each break the observe-then-decide contract or inflate cost and latency without satisfying the stated constraints.

Answer

Issue one request that asks Claude to output the entire plan plus every tool call in a single assistant turn, then execute all tools in parallel and return all results in one user turn.

Answer

Enable extended thinking with a high budget_tokens value and instruct Claude to simulate each tool's output internally, returning only the synthesized final answer without invoking any tools.

Answer

Create a separate Claude conversation for each subtask, passing the full prior transcript as a system prompt so each conversation can plan independently before returning its result to an orchestrator.

Question 3

An engineering team is building a mission-critical automated medical triage system powered by Claude. To maximize factual reliability and prevent dangerous omissions in symptom analysis, which TWO architectural strategies should be implemented? (Choose TWO)

Accepted Answer

Implement structured output validation combined with a secondary independent verification prompt to audit initial triage conclusions.. Enforcing structured output schemas guarantees programmatic validation against expected data types before downstream processing, while chaining specialized verification prompts introduces a validation gatekeeper. Together, these techniques establish robust defense-in-depth mechanisms essential for maintaining clinical safety standards and mitigating catastrophic hallucination risks in high-stakes domains.

Answer

Rely solely on open-ended conversational prompts without constraints to allow maximum clinical intuition and empathy.

Answer

Maximize the context window by injecting the entire medical history of unrelated patients to provide broader background trends.

Answer

Set temperature to 0.0 to eliminate creativity and force repetitive, deterministic generation patterns across all queries.

Question 4

An orchestration layer runs a research agent that can call a web_search tool and a summarize_document tool. The architect wants the agent to keep iterating until it has gathered enough sources, but must cap total model calls and total tool executions to control cost and prevent runaway loops. Which approach best enforces those caps while preserving the agent's ability to decide when it is done?

Accepted Answer

Wrap the loop in application code that counts iterations and tool calls, stops when either budget is exhausted, and lets Claude signal completion by returning a final answer with no tool_use block.. Hard cost and safety limits belong in the orchestrator, which can count model calls and tool executions and stop deterministically at a budget boundary. Allowing Claude to end the loop by returning a response with no tool_use block keeps the agent in control of when its task is complete, so the two mechanisms together give bounded but adaptive behavior.

Answer

Set a very high max_tokens on every request so Claude naturally stops once it has written enough, avoiding the need for an explicit iteration limit.

Answer

Instruct Claude in the system prompt to stop after at most ten tool calls and trust the model to count its own invocations accurately.

Answer

Configure the tools with a strict input_schema so malformed arguments cause the API to reject extra calls once a threshold is reached.

Question 5

You are building an application that uses the Anthropic API to classify customer support tickets into exactly one of five categories: Billing, Technical, Account, Feature Request, or Other. You require the output to be a JSON object with a single field "category" and a value from that list. During testing, you notice that for ambiguous tickets the model sometimes returns a different key name, such as "ticket_category", or wraps the JSON in markdown code fences, causing parsing failures. You need to enforce the schema reliably while keeping latency and cost low. Which approach is most effective?

Accepted Answer

Use the Anthropic API's tool use feature by defining a tool with an input schema that specifies the required "category" field and its allowed enum values, then force the model to call that tool.. Forcing the model to use a tool with a strict input schema ensures the output conforms exactly to the required JSON structure, including the key name and allowed enum values. Unlike prompt instructions or temperature adjustments, tool use provides a hard constraint enforced by the API, making it the most reliable and efficient method for schema adherence in production.

Answer

Add a system prompt instruction that explicitly states the exact JSON schema and warns the model not to include markdown formatting.

Answer

Set the temperature parameter to 0 and include a detailed example of the desired JSON output in the prompt.

Answer

Post-process the model's raw text response with a regular expression that extracts the first JSON object and remaps any key to "category".

Question 6

A financial services firm is using Claude 3.5 Sonnet to analyze quarterly earnings reports that exceed 150,000 tokens. Which TWO strategies are most effective for maintaining high reliability when the model must extract specific data points from the middle of this large context?

Accepted Answer

Encapsulating documents in XML tags like and . Long-context models like Claude 3.5 Sonnet are highly capable, but performance can still vary based on data placement. Using XML tags helps the model distinguish between different sections of the input, while placing the specific question or 'call to action' at the very end of the prompt ensures the model focuses on the task after processing all context.

Answer

Using markdown headers exclusively for sectioning

Answer

Placing the user query at the beginning of the prompt

Answer

Disabling the system prompt to save token space

Question 7

A developer is building a structured output pipeline where Claude must return a list of support tickets, each with 'id', 'priority', and 'category'. The pipeline must be robust against malformed responses. Which TWO practices most directly improve the reliability of this structured output? (Choose two.)

Accepted Answer

Validate the parsed output against the schema in application code and retry with a corrective prompt on failure.. Reliability comes from structural enforcement plus defensive validation. Defining a tool 'input_schema' with required ticket fields and forcing tool use guarantees a structured tool call, while validating the parsed result in application code and retrying with corrective feedback handles residual edge cases. Together they make the pipeline robust against malformed or invalid responses.

Answer

Set 'temperature' to 1.0 to encourage diverse ticket categorizations.

Answer

Ask Claude to 'be careful' and 'double-check' the output before responding.

Answer

Append a stop sequence of '}' to end the response as soon as the JSON closes.

Question 8

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.)

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 9

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 10

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 11

A security-conscious team is configuring Claude Code for a repository that contains both application source and generated artifacts. They want committed project configuration to keep the assistant focused on source files and to prevent accidental modification of generated output. Which two configuration actions should they take? (Choose two.)

Accepted Answer

Add deny rules for the generated paths in the committed .claude/settings.json permissions section. Two complementary controls are needed: an ignore mechanism that removes generated paths from the assistant's view, and permission deny rules that deterministically block tool access to those paths. Together they keep Claude Code oriented toward source code and prevent edits to build output for every developer using the committed project configuration, without disabling capabilities the team still relies on.

Answer

Disable all Bash tool usage in the project settings to protect the artifacts

Answer

Create a CLAUDE.md note asking the model to avoid touching generated files

Answer

Store the generated artifacts inside the user-level ~/.claude directory so they are hidden

Question 12

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 13

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 14

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 15

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.

Question 16

A large-scale agentic system uses Claude to manage a supply chain. The agent must interact with a legacy SQL database, a modern REST API, and a real-time streaming service. What is the most robust way to handle tool definitions in this heterogeneous environment?

Accepted Answer

Define separate, strongly-typed tools with clear descriptions and parameter schemas. In complex environments, tool definitions act as the interface between the LLM and various technical stacks. To ensure robustness, the architect should define clear, narrow tool schemas that abstract the underlying complexity, providing Claude with a consistent JSON interface while the orchestration layer handles the specific protocols and authentication for each backend system.

Answer

Provide the model with the raw SQL schema and API documentation in the system prompt

Answer

Create a unified 'execute_command' tool that accepts raw strings for any service

Answer

Use a single 'agent_proxy' tool that asks a human to execute the tasks manually

Question 17

You are building an agent that needs to summarize 50 different documents. Which orchestration approach provides the best balance between speed and the model's ability to cross-reference information between all documents?

Accepted Answer

Parallel processing of summaries followed by a synthesis step. For a task involving many documents, a hybrid approach is best. Parallel processing allows for fast individual summarization, but a final synthesis step (Orchestrator) is required to cross-reference and unify the information. This prevents the model from being overwhelmed by too much text in a single context window while ensuring global coherence.

Answer

Feeding all 50 documents into a single prompt for one summary

Answer

A sequential chain where each document is added one by one

Answer

An autonomous agent that decides which document to read next

Question 18

Refer to the exhibit. The model consistently provides the ID as a string instead of a number. How should the architect fix this?

Accepted Answer

Update the JSON Schema to explicitly define the 'id' field as 'number' and add a prompt example.. The most robust solution is to modify the tool's input schema to include a 'type' constraint and potentially add a 'pattern' or 'coerce' logic if necessary. By explicitly defining the type as 'number' and ensuring the model is prompted with examples, the architect fixes the root cause at the contract level. This provides the model with unambiguous instructions on the expected input format for the 'process_data' tool.

Answer

Modify the server code to cast the string to a number inside the function.

Answer

Change the tool definition to accept both 'string' and 'number' types in the schema.

Answer

Restart the MCP server to clear the previous tool registration cache.

Question 19

An architect is implementing a complex multi-turn conversation where Claude uses multiple tools. The model returns a response containing both a `text` block and a `tool_use` block. What is the correct way to handle this response in the application logic?

Accepted Answer

Display the text to the user immediately, then execute the tool and append the result to the conversation.. Handling mixed-content responses is a common requirement in tool-augmented applications. Claude often provides a rationale or update to the user in a text block while simultaneously initiating a tool call. Properly managing these blocks ensures a seamless user experience and maintains the integrity of the conversation history for subsequent turns.

Answer

Discard the text block and only process the tool_use block to minimize latency and focus on task execution.

Answer

Wait for the tool to finish executing and then combine its result with the model's text into a single user message.

Answer

Error out and ask the model to re-generate the response with only one block type to avoid ambiguity.

Question 20

A Lead Developer is troubleshooting why Claude Code is not picking up changes made to a local file. The file is not in .gitignore or .claudeignore. What is the most likely cause related to the tool's workflow?

Accepted Answer

The model's internal context hasn't been refreshed to reflect the new disk state.. Claude Code maintains a local index and context of the project to optimize performance. Understanding how the tool refreshes its view of the filesystem is key to ensuring that the model always works with the most up-to-date code, preventing frustrating errors where the model suggests fixes for already-resolved issues.

Answer

The file size exceeds the maximum individual file limit for the CLI.

Answer

The developer has not yet run the 'claude sync' command.

Answer

The file is being blocked by the system's antivirus software.