That 1 thing is engineering discipline. This Datadog guide

Summary from the TLDR newsletter

breaks the AI engineering loop down into a lifecycle any team can adopt, no matter what stack they run. Read it to see how to structure experiments across prompts, models, and retrieval to tie LLM behavior