**DRAFT — pending editorial expansion.** This article is a working draft published as scaffolding for the NINtec content programme. The current version covers the substantive perspective in compressed form; the published version will expand each section to the 2,000+ word depth the topic warrants. Editorial review is required before promotion.
Production prompt engineering is engineering discipline, not creative art. Prompts ship under the same discipline as production code: versioned, tested, observable, and rolled back when bad. This piece covers the patterns we apply across NINtec deployments.
Prompt registries
Every production prompt has a version, an owner, and a changelog. Git-backed registries make prompts auditable and reverible. Semantic versioning with major.minor.patch discipline prevents accidental breaking changes.
Claude-specific patterns
Claude rewards XML-structured prompts with named tags, explicit role priming, thinking blocks for reasoning-heavy tasks, structured-output schemas with concrete examples. Prompt patterns that work on GPT often need adjustment for Claude — direct translation loses 5–15% quality.
A/B routing and gradual rollout
New prompts route to a small fraction of production traffic first. Quality and cost metrics are compared against the existing prompt. Promotion requires explicit eval-bar parity. Rollback is one configuration change away.
Adversarial discipline
Prompts that handle untrusted input must be tested against prompt-injection attacks. Adversarial test suites are part of CI. New jailbreak techniques are tested against active prompts on a schedule.
Production prompt discipline scales the engineering team's ability to iterate. Without it, prompt changes are afraid; with it, prompt changes are safe.