Having looked at the RecDescent man page just now, I'd say that the use of the "or" conjunction causes the first element (before "|") to be tried first, and the latter to be tried only if the first does not match. In your case, the LITERAL condition matches not only the content between tags, but also everything except the "<" of the tags themselves. When you test for LITERAL first, TAG would never have an oppurtunity to match.

Being less familiar with RecDescent, I thought it might be easy to make a suitable regex for use with split() to get the same result as what you want, but it seems that this alone would not be enough -- you'd have munge the data a little, before or after the split, to get your array. For example, assuming the whole string with tags and content is now in $_ (and hoping there are no "spurious" angle brackets), here's a method that inserts split-able strings around tags to make split produce intended array:

s/(.)</$1=!=</gs;  # use some distinct pattern to mark open brackets
s/>(?!=!=)/>=!=/g; # and close brackets that aren't adjacent to "<"
@tokens = split( /=!=/ );
[download]

If the string begins with initial "<", the first substitution won't match that, so we won't get an empty (fictitious) first element in @tokens. Likewise, the second substitution makes sure that no empty tokens are generated where the input had "...><..." (nothing between adjacent tags) or a string-final ">".