From what I've understood it is just like creating sprites for any other game type, although the perspective is more difficult and you would need more animation frames than a simple side view game. But the basics I'm understanding are similar assuming you understand art, perspective, etc...which I personally don't really....

Me....I can't sprite in any perspective very well yet, but I can knock out acceptable 3d models. Since screen resolutions are pretty big these days, if you wanted to do a low-rez pixel art style you would probably need to scale up. But there are enough pixels for other styles. I make 3d models and render them out, all with equal lighting and similar shaders. Blender's materials can give me about anything I want, including the cel-shaded look, nice dithered lighting, outlines if wanted.... The modelling is much easier for this purpose than for full-on 3d rendered games because you can skip details that wouldn't be seen at lower resolutions anyways, and you can do tricks with procedural textures and modifiers that you wouldn't be able to do if you were wanting a model for the final result. If it is easier to create the model, you can easily make a bunch of polys that you don't need to optimize out later since you are just rendering them for sprites.

There are plenty of other advantages to this method besides what I mention above, though the biggest is probably the ability to easily make changes and re-render. If you have a bunch of sprite animation frames and need to change something, say make the character different, change color of clothes, change lighting direction or intensity, whatever..you have to add/redraw all those animation frames. With 3d, you can just make the change once and then re-render. So though the time to make the 3d model can be longer than the time to simply sprite out a single animation frame, the benefits far outweigh that loss later as it is much easier to animate and make changes to as needed.

I am currently making a simple asteroids clone using Gamemaker: Studio. The point of it isn't the game so much, rather proving how the art pipeline can give really nice results. I'm using 3d models for sprites, and I'm rendering normal maps for them as well. The game at run-time will have lights in 3d space which using normal-mapped shaders on my sprites will give a nice 3d look. Of course, I'm using pretty high resolutions for sprites(256x256 for some of them), and I'm using lots of animation frames too in order to keep a really nice smooth look. One point to this project is to test how well the engine can handle the large sizes and animations of the sprites. I want to see if this art pipeline can work well with good performance on "casual gamer" kind of harder, like laptops with integrated graphics chips. If it does, than I'll know it is worth doing it this way. If it doesn't, than I'll know I might as well just make the game itself in 3d with Unity because I'm not getting any "performance" advantage by faking the 3d in 2d with normal maps. I'm pretty sure I'm right though, and if so, it will open the door for future projects and the way I can make them look while still having "2d platformer" type of performance. This is all brought possible by making sprites using 3d models.