Free CCAR-F practice test: 30 questions with answers

A stop_reason of 'tool_use' means the model is waiting for a tool result, so the application runs the tool and returns a tool_result tied to the tool_use id, then calls the model again. Closing the session ends the loop too early, and restarting with tool output in the system field or changing temperature does not complete the exchange correctly.

Concise structured summaries keep the hub's context small while preserving findings and sources for synthesis. Full transcripts and raw pages flood the coordinator's context, and copying the whole history into each subagent wastes context on detail that most subagents do not need.

Validated structured returns and bounded tasks with explicit inputs keep the coordinator in control of each step. Letting subagents talk to users or to each other directly, or flooding every prompt with all other outputs, removes the coordinator's oversight and bloats context.

A check in application code cannot be talked around, so refunds stay blocked until the verification tool succeeds. Prompt sentences and self-reported reasoning are advisory and can be skipped, and temperature does not enforce any business rule.

Forking creates a separate continuation from the shared point, so the two explorations do not contaminate each other while keeping the earlier context. Asking Claude to disregard earlier work is unreliable, while pasting transcripts into CLAUDE.md or starting fresh sessions either bloats instructions or throws away the shared context.

Independent subtasks on a checklist, each verified before the next begins, keep errors contained and make progress visible. Sending everything at once, pausing for every rename, or running an unreviewed bulk replace each fail in a way that makes errors hard to find or the work too slow.

Path-scoped rule files load conventions only when Claude works in the matching area, which keeps context focused. A single large CLAUDE.md loads everything every time, slash commands depend on people remembering to run them, and shell variables on a developer's machine are not read as convention rules.

Plan mode separates investigation and proposal from editing, so the developer can approve the approach before anything changes. Auto-accepting edits and reviewing afterward, or running a temporary-branch command, allows modifications before the approach has been reviewed.

Project slash commands are markdown files committed to the repository, so every teammate can invoke them by name. A CLAUDE.md instruction is passive guidance rather than a named command, and hooks or personal aliases do not give the whole team a shared, invocable workflow.

Print mode runs non-interactively, so the permissions the job needs must be granted ahead of time rather than approved at a prompt. Pauses, interactive sessions, or deleting CLAUDE.md do not address the missing approval mechanism in an unattended pipeline.

Committed project files share instructions across the team, and nested CLAUDE.md files add guidance for their own directory. CLAUDE.md is context that guides behaviour rather than a guaranteed enforcement mechanism, custom commands and skills live in their own folders, and the file is plain Markdown rather than JSON.

Concrete failing output and a specific expected result give Claude a verifiable target to iterate against. Repeating a vague request, rewriting everything without tests, or adding adjectives does not supply the feedback needed to converge.

Explicit, testable criteria with severity definitions and named non-issues tell the model what to exclude, which cuts false positives. Asking for thoroughness, longer output, or confidence scores that are still all posted does not narrow what counts as a bug.

Forcing a tool call whose input schema defines the fields makes the output follow the structure the pipeline expects. Requests in prose, regex extraction, or code-fence stripping leave the structure unenforced, and a persona instruction alone does not guarantee schema conformance.

Feeding the specific validation error back with a bounded number of retries gives the model a targeted chance to fix the output without looping forever. Storing invalid records, discarding the batch, or halting for manual work at every failure ignores the feedback loop the pipeline can use.

The Message Batches API suits large asynchronous workloads where latency per item does not matter and cost does, and custom IDs let each result be matched back to its ticket. One-at-a-time calls and per-ticket streams are slow, and one oversized prompt is unreliable and cannot return clean per-ticket results.

An explicit schema, worked examples, and a closed label set all narrow the output space so results are consistent and parseable. Higher temperature increases variation, and a free-form paragraph followed by a second conversion step adds an unstructured stage that is hard to make reliable.

A worked example that shows the missing-field case teaches the model the exact output for absence, which the existing examples did not cover. A lower temperature, a bare prohibition, or removing the examples does not show the desired behaviour for missing data.

Clear descriptions that state what each tool returns, the input it expects, and when to avoid it give the model the signals it needs to route correctly. Emphasis words and a higher token limit do not resolve the ambiguity, and removing a tool that some workflows still need breaks legitimate use.

Setting isError to true with an informative message lets the agent recognise the failure and decide whether to retry or report it. Reporting success with an empty result, crashing the server, or hiding an error word inside normal content all keep the failure away from the agent's error handling.

A project-scoped .mcp.json committed to the repository shares one server definition with everyone who clones it. Personal settings and local scope keep the configuration private, and a URL in CLAUDE.md is not a registered MCP server, so the tools would not be available to the agent.

Glob matches file paths by pattern, which is exactly what locating test_*.py files requires. Read opens a file that is already known, Grep searches file contents rather than names, and Edit modifies files, so none of those fit the listing step.

Precise parameter descriptions and actionable error results help the agent call the tool correctly and recover from problems. Unlabelled raw output, free-form multi-purpose tools, and vague names all make correct selection and use harder.

The any setting forces a tool call while letting the model pick the matching extraction tool for each document. Auto allows a plain text reply, none forbids tool use, and naming one tool would force the wrong extractor for other document types.

Pinning critical facts in a structured block that is re-supplied each turn keeps them exact instead of relying on paraphrase or recall across a long conversation. Asking for restated figures, trimming the system prompt, or removing the lookup tool does not preserve the facts reliably.

An explicit request for a person, or a clear gap in documented policy, are observable conditions that justify a handoff. Sentiment scores are unreliable proxies, a token count has nothing to do with the customer's need, and escalating every refund mention ignores whether the case is already covered.

Structured error context with the failure type, the failing step, partial results, and retry guidance lets the coordinator decide whether to retry, reroute, or report partial coverage. A bare failure word hides what was done, a fake success corrupts results, and a raw log dump buries the information the coordinator needs.

Targeted search plus a notes file with file references keeps context small while letting later sessions pick up where earlier ones stopped. Loading everything, returning all file text, or relying on one sentence either overflows the context or leaves out the detail needed to answer accurately.

Keeping each claim attached to its source and date, and flagging the disagreement, preserves provenance and shows the uncertainty. Averaging, silently keeping only the newest figure, or choosing by plausibility hides the conflict and presents guesses as fact.

When the target is ambiguous and the action changes data, the agent should confirm the identity before acting rather than guess. Picking a match, updating both records, or relying on the requester to undo changes risks modifying the wrong record.