I was once charged with maintaining a piece of Perl code that was responsible for parsing two CSV files, extracting and combining interesting information, and writing out a new "result" CSV.

The original author had made some assumptions about the CSV format that were simplistic (and he hadn't used Text::CSV_XS or it's compatriots). So, when the application that generated the two source CSV files found commas in the data for the first time, it did the right thing and quoted the whole value; the code died.

I set out to fix this bug by replacing the homegrown CSV parser with Text::CSV_XS. In the process, however, I discovered a bug in the result-set generator that had been skipping important information for several years -- I'd never have noticed if I hadn't constructed test data with which to test my changes. The "2-hour bugfix" I'd pitched to management turned into a two-week re-write.

It was that experience which started me down the road to test-driven development...