Prompt Engineering in Production: Versioning, Evaluation, and Injection Defense

Prompt engineering interviews in mature teams treat prompts as versioned configuration, not artisanal text. Interviewers ask how you structure system, developer, and user messages; how you enforce JSON or tool schemas; and how prompt changes flow through CI with regression evals before reaching production traffic.

Structured output reliability is a production requirement. Candidates discuss schema validation, repair loops, constrained decoding where available, and fallback behavior when parsing fails. A support bot that returns malformed JSON breaks downstream automations—your design must handle that gracefully.

Prompt injection and instruction hierarchy are security topics, not optional extras. Explain how you separate trusted instructions from untrusted retrieved content, use delimiters without false confidence, and apply output policies that resist indirect injection via documents or tool responses.

Few-shot and chain-of-thought patterns have cost and latency implications. Interviewers want trade-off analysis: when dynamic few-shot retrieval beats static examples, when CoT should be hidden from users, and when self-consistency is worth 3x token spend.

Prompt optimization without measurement is guesswork. Describe A/B frameworks, offline rubrics, and guardrails against overfitting prompts to a tiny eval set. The best answers mention collaboration between prompt authors, eval owners, and on-call engineers when a deploy causes quality cliffs.