20 free CCAR-F sample questions

Q1

The agent resolves refunds and shipping questions correctly, but in 40% of billing disputes it invents account credit balances that do not exist instead of reading them. What is the best way to fix this?

• AAdd a fetch_account tool call so the agent reads the customer's real credit balance before responding to any billing dispute.correct
• BLower the model temperature toward zero so its decoding becomes near-deterministic and it stops sampling fabricated numeric balance values during billing replies.
• CAppend a strongly worded system-prompt line that explicitly forbids inventing balances and instructs the agent to never state a credit figure it is unsure about.
• DRoute every billing dispute through hand_off_to_agent so a human pulls the real balance and the agent never answers credit questions on its own.

Q2

Each sub-agent currently returns a free-form paragraph, and the coordinator's merge step fails to align findings in roughly 1 in 3 runs because it cannot tell which claim came from which sub-agent. What is the most effective approach?

• AHave sub-agents return a structured summary with explicit source, claim, and confidence fields.correct
• BGive the coordinator a larger context window so it can hold every sub-agent's full paragraph in memory at once.
• CInstruct each sub-agent to write much shorter paragraphs so the coordinator can read and merge them faster.
• DRun the merge with a second model pass that votes across the paragraphs and keeps claims both passes agree on.

Q3

You are building the agentic loop for a support assistant that answers order questions with find_order and check_shipping. The current loop ends whenever the assistant's reply text contains the phrase "all set", and tool outputs are pasted back into the conversation as plain narrative text. Transcripts show premature endings and misread tool data. Which two changes should you make? Select TWO.

• AContinue the loop when the API returns a stop_reason of tool_use and end it only on end_turn.correct
• BCap the loop at a fixed ten iterations and treat hitting that ceiling as the normal, expected completion signal.
• CAppend each tool's output to the conversation as a tool_result block paired with the call that produced it.correct
• DTighten the phrase-based completion check so "all set" only terminates the loop when it appears in the final sentence.
• EInject each tool's output back into the conversation as a JSON-formatted user message so the model reads structured data.

Q4

An internal chore assistant calls list_stale_branches once, receives the result, and then the run simply ends with the branches never deleted. Reviewing the harness, you see it makes a single API call, executes whatever tool the model requested, and exits. Which two changes turn this into a working agentic loop? Select TWO.

• ARaise max_tokens so the model can finish the whole cleanup in its first response.
• BAfter executing the requested tool, append the result to the conversation and call the model again.correct
• CAsk the model to list every tool call it will need up front so the harness can run them in one pass.
• DKeep iterating until the API returns a stop_reason of end_turn.correct

Q1

A developer keeps re-typing the same instructions about the project's import ordering, preferred test runner, and the fact that the API client lives in src/lib/api/. These facts apply to every session in the repo. What's the most effective way to make Claude Code apply them automatically?

• ARecord the conventions in the repo's CLAUDE.md so they load into context for every session.correct
• BCreate a path-scoped rule under .claude/rules/ targeting src/lib/api/ so the notes load when relevant.
• CEncode each convention as a separate slash command under .claude/commands/ and invoke them in sequence before coding.
• DAdd the conventions to ~/.claude/CLAUDE.md so they persist automatically without touching the repository.

Q2

A developer is about to ask Claude Code to migrate authentication across roughly 18 files and wants to review the full sequence of intended edits before any file is written. What is the best way to do this?

• AUse plan mode so Claude proposes the steps first and waits for approval before editing.correct
• BHave Claude work on a fresh branch so the edits can be reviewed as a diff before merging.
• CSet permissions to deny every write tool so Claude can only describe the intended changes.
• DSplit the work into 18 separate single-file requests so each edit can be reviewed individually.

Q1

A pipeline turns insurance claim PDFs into records, and roughly 9% of outputs come back with trailing prose like "Here is the JSON you requested" wrapped around the object, breaking the downstream parser. What is the best way to force clean, parseable output?

• AConstrain the response with a JSON schema so the model emits only a valid object with no surrounding text.correct
• BAdd a regex post-processor that strips any text appearing before the first { and after the last } character on each response.
• CAdd a few-shot example of a bare, correctly formatted JSON response so the model imitates that output shape.
• DFine-tune a dedicated extraction model on several thousand labeled claim PDFs so it learns to omit the surrounding prose.

Q2

The support agent resolves billing questions but occasionally invents a refund policy detail that does not exist in the account data returned by fetch_account. What's the most effective approach to reduce these fabricated policy claims?

• ALower the temperature on every model call to zero so that answers become deterministic and the model stops inventing policy details.
• BInstruct the agent to answer only from fetch_account data and say it must escalate when the data is missing.correct
• CAdd a second model that fact-checks the first model's policy statements after each reply and rewrites any unsupported claims it finds.
• DAdd few-shot examples of correctly answered refund questions so the model internalizes the real policy wording.

Q1

A frontend repo's checked-in CLAUDE.md mixes team-wide conventions (component naming, the test runner to use) with one engineer's personal preferences ("respond tersely", "show diffs vim-style"). Teammates complain that their sessions now pick up the personal preferences. Which two changes should you make? Select TWO.

• AMove the entire file into that engineer's ~/.claude/CLAUDE.md so the repository stays clean.
• BRelocate the personal preferences into the engineer's own user-level ~/.claude/CLAUDE.md.correct
• CConvert the personal preferences into a project slash command that teammates simply avoid running.
• DKeep the team-wide conventions in the project CLAUDE.md so every teammate's sessions load them.correct
• EDuplicate the team conventions into a CLAUDE.md in each subdirectory so they take precedence.

Q2

In a monorepo, guidance about the build tool and the PR checklist applies to work anywhere in the tree, while packages/api/ has handler conventions that only matter when editing that package. Which two placements scope this guidance correctly? Select TWO.

• APut the build-tool and PR-checklist guidance in the repository root CLAUDE.md.correct
• BPut the API handler conventions in the root CLAUDE.md so they can never be missed.
• CPut the API handler conventions in a CLAUDE.md inside packages/api/.correct
• DPut the repo-wide guidance in each engineer's ~/.claude/CLAUDE.md for reliability.

Q1

Your CI review bot posts findings under the instruction "be conservative and only report issues that really matter." Developers now dismiss nearly everything it posts, largely because its dependency-vulnerability category is wrong about 60% of the time. Which two changes should you make? Select TWO.

• AAdd a sentence to the instruction telling the bot to report only the findings it is at least 90% confident about internally.
• BReplace the vague instruction with explicit criteria defining which specific conditions to report and which to skip.correct
• CLower the sampling temperature on the review requests so the bot's flagging decisions become more repeatable across runs.
• DHave the bot state each finding twice in different words and post only findings that survive both phrasings.
• ETemporarily disable the dependency-vulnerability category while its criteria are rewritten.correct

Q2

A QA prompt reviews support transcripts and flags "unprofessional tone." Team leads overturn nearly half the flags: the model marks casual-but-friendly greetings, while leads only care about dismissive or hostile replies. Which two prompt changes should you make? Select TWO.

• ADefine the violation as specific behaviors, such as dismissing a stated problem or blaming the customer, with one concrete transcript line per behavior.correct
• BTell the model to flag a transcript only in cases where it is very confident, based on its own judgment of the exchange, that the tone is genuinely unprofessional.
• CList acceptable informal patterns, such as casual greetings and light humor, that must never be flagged.correct
• DHave the model rate each transcript's overall tone on a numeric scale from 1 to 10 and automatically flag every transcript whose score comes out above 7.

Q1

The MCP server exposes a tool described only as "issue_refund — refunds a customer." In 31% of refund attempts the agent omits the order_id because the description never states which inputs are required or what they mean. What is the best way to improve reliability?

• AExpand the tool description to name each parameter, mark order_id as required, and state when refunds apply.correct
• BAdd a system-prompt note telling the agent to always pass a valid order ID whenever it calls the refund tool.
• CAdd server-side validation that rejects calls lacking order_id so the agent can retry after each failure.
• DSplit issue_refund into start_refund and confirm_refund so the agent is forced to make a second confirming call.

Q2

Invoices are extracted into records, but downstream systems reject 18% of them because total_amount arrives as strings like "$1,240.00" or sometimes null. The extraction tool's schema currently declares total_amount as {"type": "string"}. What is the most effective way to block malformed output?

• AAdd a post-processing script that parses the strings and strips currency symbols after extraction.
• BDefine total_amount as {"type": "number", "minimum": 0} and make it a required field.correct
• CInstruct the model in the prompt to always return numeric totals without symbols.
• DAdd a regex pattern to the schema so total_amount strings must match a currency format.

Q3

A support agent's issue_refund MCP tool is described in one line: "Handles refunds." Transcripts show the agent omitting the required order_id in a third of calls, and also invoking the tool when customers ask for exchanges, which it does not support. Which two changes to the tool definition should you make? Select TWO.

• AAdd a system-prompt instruction reminding the agent to include an order ID with every refund call.
• BDocument each parameter in the description and mark order_id as required.correct
• CHave the server default a missing order_id to the customer's most recent order so incomplete calls succeed.
• DSplit the tool into separate request_refund and approve_refund tools so every refund takes two calls.
• EState in the description when the tool applies — refunds only — and that exchanges are out of scope.correct

Q4

An internal toolkit exposes lookup_docs and search_docs, both described as "Finds relevant documentation." The assistant's system prompt also says "when in doubt, use search_docs." Engineers report documentation requests landing on the wrong tool about half the time, including cases where lookup_docs was clearly the fit. Which two changes should you make? Select TWO.

• ALower the sampling temperature on requests so the model's tool selection becomes more deterministic.
• BRewrite each tool's description to state the distinct corpus it covers and when to choose it.correct
• CRemove the system-prompt sentence that steers the model toward search_docs regardless of the request.correct
• DRe-register the two tools in the opposite order so the preferred one is encountered earlier in the tool list.

Q1

The coordinator dispatches 6 parallel sub-agents, each researching one supplier. Each sub-agent returns its full browsing transcript of roughly 18,000 tokens, and the coordinator's window overflows before it can merge them. What is the best way to keep the merge within budget?

• AHave each sub-agent return a structured summary of its findings instead of its full transcript.correct
• BIncrease the coordinator's context window and feed all 6 raw transcripts unchanged.
• CHave each sub-agent write its full transcript to a file and hand the coordinator just the file path.
• DRun the 6 sub-agents sequentially instead of in parallel so their transcripts arrive one at a time.

Q2

A developer keeps pasting the same 400-token explanation of the project's folder layout and naming rules into every Claude Code session before asking for changes. What's the most maintainable way to make this guidance persistent?

• AMove the layout and naming rules into the project's CLAUDE.md file.correct
• BAdd the rules to ~/.claude/CLAUDE.md, which loads automatically at session start.
• CCreate a custom slash command that inserts the 400-token explanation on demand.
• DHave Claude infer the layout from the directory tree at the start of each session.

Q3

A support agent compresses its conversation history into a rolling summary every 10 turns. An audit of escalated billing cases shows handoff notes saying things like "customer disputes a recent charge of around $180" when the customer actually stated $183.42 and was promised a callback by March 12. Which two changes prevent these details from being lost? Select TWO.

• AAdd an instruction to the summarization prompt telling the model to be extra careful when carrying over numbers and dates.
• BExtract exact amounts, dates, and identifiers into a structured case-facts block maintained outside the summarized history.correct
• CDisable the rolling summarization entirely and retain the full transcript in context so that nothing is ever paraphrased away.
• DBuild escalation handoffs from the structured case record rather than from the rolling prose summary.correct

Q4

A research coordinator concatenates eight sub-agent reports into a single unbroken block before synthesis, and reviewers keep finding that findings from the third through sixth reports never appear in the final brief. Which two changes address this? Select TWO.

• AOpen the aggregated input with a key-findings digest that surfaces each report's main conclusions up front.correct
• BAppend a brief recap at the end of the input and rely on it alone, leaving the reports unlabeled in arrival order.
• CInsert explicit section headers marking where each report begins and which sub-agent produced it.correct
• DAdd an instruction telling the synthesis model to pay equal attention to every report regardless of its position in the input.
• ERandomize report order on each run so no report is permanently stuck in the middle.