Depends on the task but some combination of asking it for skeleton code for new tasks and sending it my written code or error messages and asking for corrections or potential solutions.
It's very effective, if you are atleast semi-new to technology you are using it will explain and teach you things you didn't know before, and if you know the tech by heart it saves you from having to type it out.
For example, yesterday I had to make a custom container with some pretty involved dependiencies that also had to be be runnable on AWS Lambda (which I haven't used much before), me and gpt4 went back and forth with Dockerfile code and error messages for a few hours and then it ran like charm. Would probably have taken me 1-2 days of regular coding and googling otherwise.
It also can't actually start anything on its own. That is, it has no motivation, like an actual AGI would. Don't pollute that term also. An Artificial General Intelligence is something like Asimov's robots or R2D2, not search-on-steroids.
