A fixed, curated set of inputs (and expected outputs or rubrics) used to regression-test prompts, models, or tools on every change.
Unlike ad-hoc vibe checks, a golden set makes quality comparable across time and versions. It pairs with LLM benchmarking habits and reduces verification debt by catching regressions in CI.
Design tips
Represent real customer phrasing, edge cases, and failure modes; version the dataset alongside prompts and schemas.