CCAR-F question bank

A data migration workflow runs nine stages in order. Stage seven calls an external service that fails about once a week. When it does, the operator re-runs the workflow from the beginning: stages one to six take 40 minutes to do their work again, and stage three sends its notifications a second time.

What does the workflow need?

A classification prompt sorts inbound messages into eight categories, one of which is other. Messages were landing in other too often, so somebody added twelve examples of messages that had been misfiled there, each paired with the category it belonged in.

The rate of other fell by half. Messages resembling none of the twelve are still landing there.

Select the two changes that would carry further.

A coordinator dispatches one worker per refund case. A worker calls the payment tool, which occasionally takes longer than the coordinator's per-worker deadline. When the deadline passes the coordinator treats the worker as lost and dispatches a second one for the same case. Sometimes the first worker's call had already gone through, and the customer is refunded twice.

The team lengthened the deadline, and a smaller number of double refunds still happen. What actually stops them?

A project has grown five instruction files beyond its CLAUDE.md. Two carry a paths field in their frontmatter, two carry no frontmatter at all, and one is a skill.

Select the two that are in context at the start of every session, before Claude has read any source file.

Two jobs are being designed. The nightly one runs the same four steps on every filing, in the same order. The other investigates a production incident, where what to look at next depends entirely on what the last thing looked at showed. Both will run unattended, and the team has been asked to cost each of them before either is built.

Which decomposition does the second job need?

Auditing what sessions search for, a team wants every find a session runs to stop for approval first. They removed Bash(find:*) from the project's allow list.

Sessions go on running find with no prompt.

What should they do?

A security team runs an agent that triages incoming malware reports. Most runs finish normally. A few come back with content empty, stop_reason set to refusal, and a stop_details object naming the policy category that fired.

The loop's error handling was written for transport failures, so anything carrying no usable text is retried on the same model with a growing backoff, three times, before the case is logged as a model outage. Nine retries later this month, not one of those cases has produced an answer.

What should the loop do with a refused turn?

An invoice extractor returns a total and a list of line items. Every record produced last month passed schema validation, and the pipeline reports a 100% success rate.

An audit of 200 records finds that 9% carry a total that does not match the invoice. These invoices print the line items, then a subtotal, then tax and shipping, then the total, and the wrong records carry the subtotal.

What check is missing?

A research agent runs an agentic loop. Each time round, it sends the conversation, and if stop_reason comes back as tool_use it runs the requested tools and goes round again. Anything else it treats as the model having finished, so it takes the last text block as the answer and returns it.

On long filings the returned summary sometimes breaks off in the middle of a sentence, and the run is logged as a success. What should the loop do differently?

An incident-response skill has grown to 900 lines. Most of it is a table of every alert the platform emits and what it means, which matters for perhaps one incident in five.

Sessions that invoke the skill now carry the whole table.

What should the author do?