The biggest change for me has been separating planning from execution.
Previously, I would describe a task and immediately ask the model to generate code. Now I first ask it to inspect the requirements and relevant project context, then create an ADR or implementation plan. A developer reviews that plan before the model writes any code.
After implementation, tests, linters and static analysis provide another feedback loop before the change reaches code review.
This makes the process feel less magical, but much more reliable. AI has not removed the need for developers. It has pushed us toward clearer tasks, smaller changes and faster validation.