I noticed the end '-' thing too. I did add a section to ccn's code that handles that. Its tagged w/ the "# ADDED" comment.

++ btw ccn, took me awhile to wrap my head around this one.

my $brackets = qr(\[[^\]]*\]); # text enclosed in brackets
my $char = qr([^\[\]\s]); # not spaces or brackets

s{
  (                # group to $1
    (?:
       $char          # a char 
       (?:$brackets*) # followed by any number of brackets
    )
    {3}               # 3 times
  )
  (?!(?:$brackets*)(?:\s|\Z)) # ADDED: eliminates '-' on the end of wo
+rds w/ a multiple of 3 chars
  (?=              # looking forward to ensure that
     [^\[\]]*      # we are not inside of a brackets
     (?:
       \[
       |
       \z
     )
  )
}
{$1-}gx;

print;
[download]

Thank you very much! ccn, jbware
Your solution is great.
I've made a one-liner from it and it looks like
$text=~s{((?:([^\[\]\s])(?:(\[[^\]]*\])*)){3})(?!(?:(\[[^\]]*\])*)(?:\s|\Z))(?=[^\[\]]*(?:\[|\z))}{$1-}gx;
It's so nice! Can be used in some obfuscation code... Thanks again!

Here is non-regexp solution. It validates input in addition to the main task.

sub splitlong {
    my $text = shift;
    my @result;
    my ($cnt, $inside) = (0, 0);
    for (split //, $text) {
        ($inside = 0), next if $inside and $_ eq ']';
        $inside = 1 if !$inside and $_ eq '[';
        next if $inside;
        $cnt = /\s/s ? 0 : $cnt + 1;
        ($cnt = 0), push @result, '-' if $cnt > 3;
    }
    continue {
        push @result, $_;
    }
    warn "invalid input: $text" if $inside;
    return join '', @result;
}

my $text = 'ju[color=1]s[/color]t a sam[color=2]p[/color]le';

print splitlong($text);
[download]

This code works fine with incorrect input. The regexp version should be corrected. Problems discussed here Re^2: Problem: how to split long words. Thank you.