Treating LLM prompts like code: a regression catalog for AI failures

A field note on turning prompt-engineering folklore into structured, regression-tested artifacts. A few months ago, an LLM call in the system I work on confidently paired a refund with a beneficiary