Lessons from Using AI Tools in Actual Engineering Work

I spent six months integrating AI into my daily engineering workflow. Not as experiments or side projects—as the primary way I shipped production code, debugged systems, and made architectural decisions. This wasn't about maximizing AI use or proving it could replace developers. It was about finding where AI actually made me faster versus where it created new problems I didn't have before. The results were uncomfortable. AI transformed some parts of my work and made other parts demonstrably worse. The difference had nothing to do with prompting skill or model choice. It had everything to do with understanding which engineering tasks are actually about pattern matching and which require something AI fundamentally cannot provide. The First Uncomfortable Truth AI is exceptional at tasks I already know how to do. It's nearly useless for tasks I don't understand yet. When I asked Claude Opus 4.6 to write a data validation function for a REST API, it generated clean, working code in seconds.