On one project I worked on, we divided all feature requests into small, medium, large and huge (expected to be roughly 2, 5, 10, 20 days worth of effort respect

On one project I worked on, we divided all feature requests into small, medium, large and huge (expected to be roughly 2, 5, 10, 20 days worth of effort respectively). This allowed us to come up with a quick overall costing of the project before starting - the estimates were gradually refined as we learned more about the system we were building. Looking back over the project, once most of the features were implemented, we could see that we had mostly underestimated the small features (they took maybe 3 or 4 days) and mostly overestimated the large or huge features.