It's based on an improv sketch format I saw back at university: a scene, usually based on audience input, would be acted out in two minutes. Then, the actors would repeat the scene in one minute, then thirty seconds, fifteen, eight, four, two and one. The one-second version would inevitably consist of each actor just screaming a single word and making a single gesture that had come to sum up their performance.
Let's say you have eight hours available for the jam. In the first four hours, you make a game (perhaps subject to further restrictions or inspiration by theme or mechanic). Then, in the next two hours, you re-make the game. You are not allowed to reuse any of the code of the previous version, and should try to distil down the core of the game. Then you redo it in an hour, in half an hour, and finally in a frantic fifteen minutes.
Restarting game development from scratch is sometimes tempting but rarely useful. In this sketch format, it's mandatory. In each iteration, you can leave out more things that turn out not to be part of the core of the game. And of course, you get faster at implementing the game itself. The final fifteen minutes are a kind of "code kata" where you distil down your knowledge of what the game is and how to write it into a single clean experience.
Maybe, at the end, you are left with nothing. Or maybe you have managed to refine the core of your game idea into something that's worth pursuing. Or the pressure to be efficient has made you try out some new way of writing games you hadn't considered before.
As an extension, you could have a final round which is as long again as the first one. This time, you can reuse code if you want to, and your goal is to produce a more full and polished version of your concept.