For me vivid imagery works best.

I found there's too much room for error in my recollection when I use phrases, getting the precise words right etc. And I also find them harder to remember than images once they move into the last stack and the review intervals are longer. The other thing is that my ultimate goal here is to learn another language, so I want to try and eliminate english as much as possible from this process.

'Course, some kanji have such abstract collections of primitives that only a phrase will work, so I use them then, but try to keep it to a minimum.

I also find that I need really solid noun-y images to get the best results, the more concrete, the better. So with a keyword like fortune, I'll visualise a fortune cookie, rather than imagining the abstract concept of fortune.

It's interesting, because if I have a hardcore day of adding new cards, that night my dreams tend to be waaaaay more vivid than usual. I guess that's a sign that the visual/imaginary part of my brain is in overdrive mode.