Don't try to write it yourself. Keeping up with the spammers is a full-time job.

The hot technology right now seems to be Bayesian filtering, which looks at the types of words that appear in an individual's legitimate mail versus the spam he receives to determine the likelihood something is spam. I personally use SpamAssassin, which uses Bayesian filtering in addition to a variety of heuristics. SA is available as part of the hosting capabilities provided by pair Networks -- which, incidentally, also hosts PerlMonks. Once I trained it, SA seemed to capture more than 90 percent of spam, with zero false positives. The results could probably be better, but I would rather wade through a few unmarked spam than risk false positives.

I've also heard some folks rave about SpamBayes, which is available as a server-side filter or Outlook (if you must ;-) plug-in. I have no personal experience with it.