Prompts Aren't Real

An engineer with 25 years of experience shares insights about building AI agents for consumer use, noting that while LLMs are impressive, they fail in unpredictable ways when deployed at scale—including inconsistently following instructions, handling structured outputs, and completing reliable tasks. Despite these challenges, he expresses genuine enthusiasm for the work, describing it as the most rewarding programming he's done in his career while cautioning that the gap between successful engineering and psychological strain in this field is particularly thin.