It works best on a long-running project. When you plan a new feature, you can gauge really accurately at what speed your current team developed a feature of comparable complexity, so as you go forward you get better and better estimates.
Joel Spolsky explained it there: http://www.joelonsoftware.com/items/2007/10/26.html
I know this because every line of code I've written (using that company's process) for years has been measured and timed. So now the estimation process becomes a matter of figuring out how many LOC (as a proxy for effort) a given feature/project will be and then dividing that number by 30.
It is scarily accurate, but like I said, it's been aggregated over thousands of lines of code for years.