How to manage AI prompts like production code — and the tradeoffs involved

A software engineering post on DEV Community outlines the infrastructure gap between AI demos and production-ready AI products, focusing on prompt management, usage tracking, and safety guardrails. The author proposes storing prompts in a database using a structured schema that tracks variables, language, versioning, and category, rather than scattering them across source files. A key tradeoff highlighted is that database-stored prompts lose git-style history, meaning version numbers indicate how many revisions exist but not what earlier versions contained. The system also includes per-request token usage tracking and a character-based budget estimator to prevent runaway API spending, though the latter is acknowledged as an approximation rather than a precise measure. Safety guardrails covering PII detection, toxicity scoring, and prompt-injection checks are included, but the author notes these rely on regex and keyword patterns, limiting their effectiveness to specific formats and languages.