LLMs reward expertise

Where the conversation got more useful was in drawing the boundary. For frontier tasks, legacy systems, and domain-heavy work, people were blunt that shallow prompting falls apart fast. You can get a model to produce code, but not necessarily code that fits the codebase, respects the tradeoffs, or survives contact with reality. That is why experienced users described the workflow less as “prompt engineering” and more as supervising a fast junior. They iterate, test, reject bad ideas, and keep the model pinned to evidence. At the same time, a lot of people pushed back on the stronger version of the thesis. For routine work like CRUD apps, dashboards, HTML and CSS, and one-off internal tools, LLMs already flatten the field. They let people skip learning parts of the stack they do not care about and still get useful output. Several commenters argued this is not a bug but the whole point. The practical distinction that emerged is not “experts win, novices lose.” It is that LLMs compress simple, well-trodden tasks while making judgment more valuable on anything messy, novel, or expensive to get wrong. The anxiety running underneath all of this was about skill formation. If AI handles the frustrating middle, how do juniors build the depth that seniors now use to steer the tools? Multiple people worried that teams may produce more tickets and more slop at the same time. The short-term productivity is real. So is the risk of a generation that can ship with AI but cannot recover when the model misses, hallucinates, or quietly bakes in bad architecture.