Basically you feed it a massive volume of application code. It turns out there is a lot of commonality and latent repetition that can be teased out by LLMs, so you can get quite far with that, though it will fall down when you get into more novel terrain.