I usually break the task down into a few pieces, but I don't estimate for these kinds of things in blocks of time smaller than a half day. As long as there are at least 5 or 6 pieces to the feature after breakdown I find that the errors balance themselves out for the most part (some tasks take under an hour, etc)
Of course, the minimum time division and number of pieces required for some level of comfort needs to vary depending on the problem domain - at least 5 or 6 half day chunks seems to be about right for the stuff I've been asked about lately, but that needs to be reviewed every few months.
When I'm asked to estimate on behalf of someone else, I resist a bit more and follow a similar practice with a generous padding system ("double and add x" as mentioned above is probably a good approximation)
