Making your color scheme systematic is probably the best way to get a good-looking palette. It's not pure science - so dont just plug in numbers, but structured relationships between colors definitely makes for a more pleasing map. You can use any number color functions available to CartoCSS. For categorized features, I like using spin(), for getting different hues but maintaining the exact same lightness and saturation.
As for data steps there a number of statistical + geographic approaches (jenks is popular, and QGIS will calculate this for you), but I think for non-scientific maps like these, it's really just about what represents the data most effectively at a glance - which is to say trial and error.
I used a quasi-log scale for my NYC building-size map, which you can checkout here: https://github.com/mapbox/reinv