Treat prompt changes like code deploys

This article argues that prompt changes in LLM-backed features should be treated with the same rigor as code deploys, using automated evaluation gates to catch regressions. It explains why prompt changes fail silently—degraded outputs return HTTP 200 with plausible text, unlike code errors that throw or fail tests. The proposed pattern includes a prompt store, evaluation runner, and promotion gate that blocks changes unless a latest eval run passes. Design choices like 'fail closed' and actionable failure reasons are emphasized. The article also places this on a maturity ladder from no process to continuous eval in production, targeting teams using platforms like Microsoft Foundry.