CCAR-P sample questions with answers

• A

  In each repository's .claude/settings.json under permissions.deny, committed to version control so every team member inherits the same deny rules on checkout.

• B

  In every developer's ~/.claude/settings.json, pushed once by IT during machine setup.

• C

  In managed settings (managed-settings.json via MDM) as permissions.deny rules for Bash(curl *), Bash(wget *) and Edit(//etc/secrets/**).

• D

  In a CLAUDE.md at the organization root instructing Claude never to use curl or wget and never to touch /etc/secrets.

• A

  A guided-setup assistant that checks the customer's integration configuration and walks them through the three recurring fixes, measured by 14-day integration completion against a control cohort.

• B

  A general product assistant with retrieval over the full documentation set and every help-centre article, available on every page of the product, measured by customer satisfaction with the assistant's answers.

• C

  An agent with credentials to each customer's environment that detects configuration problems and applies the fixes automatically without asking, measured by the number of fixes applied per week.

• D

  A six-month discovery programme interviewing churned customers to confirm the cause before committing to any build, measured by interviews completed.

• A

  The new model is hallucinating; add citation requirements to the prompt.

• B

  Cap max_tokens at roughly 150 words' worth so the model cannot exceed the limit.

• C

  The prompt relied on the old model's implicit defaults; state length and tone explicitly and re-run the eval with a length check.

• D

  Roll back to the previous model permanently, since a model that changes summary length and tone without being asked is not fit for this task.

• A

  Per-segment evaluation comparing selection and error rates across applicant groups and proxies such as name, school and address, plus qualified human review of decisions.

• B

  A larger held-out test set drawn from the same applicant pool so the 91% agreement figure carries a tighter confidence interval before launch.

• C

  Removing applicant names from the prompt before ranking, which eliminates the main channel through which demographic bias could enter the model's decision.

• D

  A system prompt instruction telling the model to be fair and unbiased in its rankings.

Choose 2.

• A

  The options considered and the trade-off each makes between accuracy, cost, latency and risk, expressed in business terms.

• B

  Only the recommended option, to keep the committee focused on approving the budget.

• C

  A statement that the system prompt prevents the model from producing incorrect product claims.

• D

  The residual risks after mitigation, such as the expected rate of descriptions needing correction, and who is accountable for them.

• E

  Demo outputs from a curated set of ten products, presented as representative of production quality.

• F

  The full system prompt, so the committee can review the wording line by line.

• A

  Drop the oldest messages once the limit is near, keeping only the most recent forty turns, since the recent tool outputs and test results are what the next edit depends on.

• B

  Start a fresh session and paste in the original task description.

• C

  Increase max_tokens so the model can summarise its own history in the reply.

• D

  Compact the conversation: summarise the history into a note that keeps decisions and open items, drop redundant tool outputs, and continue from the summary.

• A

  Reference the token as ${JIRA_TOKEN} in .mcp.json and have each developer set that variable in their own environment.

• B

  Move the token into the env block of the committed .claude/settings.json so it is separated from the server definition.

• C

  Add .mcp.json to .gitignore and ask each developer to recreate the server with claude mcp add.

• D

  Keep the token in the file but restrict the repository to internal contributors only.

• A

  Place the entire handbook in the prompt as a static prefix with prompt caching enabled, and refresh the cached prefix when the handbook changes.

• B

  Build the RAG pipeline as planned so that only the handful of relevant chunks is sent on each request, keeping per-request token usage minimal as the handbook grows.

• C

  Have Claude summarize the handbook to 5,000 tokens once and answer from the summary.

• D

  Split the handbook across three specialist agents (benefits, leave, conduct) and route questions between them.

Choose 2.

• A

  Decompose by indicator: for each of the 12 indicators, send all 2,000 records in one request and ask for that indicator's total.

• B

  Decompose by data boundary: process each patient record independently, extracting all 12 indicators for that record as structured output.

• C

  Let a single agent read records in whatever order it chooses until it is confident of the network-wide totals.

• D

  Compute the network-wide totals in code from the per-record structured outputs rather than asking the model to aggregate.

• E

  Process records in prompts of 500 records at a time so the model can spot trends across patients while it extracts.

Choose 2.

• A

  Hold out a test split that is never inspected while iterating on the prompt, and report it separately from the development split.

• B

  Version the dataset alongside the prompt so every reported score is tied to a specific dataset revision.

• C

  Retire cases the model passes consistently to keep the set small and fast.

• D

  Let the prompt engineer read every failing test-split case and edit the prompt until it passes.

• E

  Regenerate the entire dataset with the latest model before each release so it stays current.