Which best captures the key idea of "Validation and retries"?

ASeparate instructions from data, and tell the model to treat data as untrusted BValidate structured output and re-ask on failure CGenerate a first draft you then shape - beat the blank page DLog requests, assess impact, decide formally — protect scope and sanity