We found having about 50% free disk space is a good rule of thumb. Cassandra will typically use less than that when running compactions, however it is good to have free space available just in case it compacts some of the larger SSTables in the keyspace. More information can be found on the Datastax website 
If you have a situation where only one node in the cluster is running low on disk space and all other nodes are fine for disk space, there are two things you can do.
1) Run a 'nodetool repair -pr' on each node to ensure that the token ranges for each node are balanced (this should be run periodically anyway).
2) Run targeted compactions on the problem node using 'nodetool compact [keyspace] [table]', where [table] is the list of the SSTables tables on the node that need to be reduced in size.
Note that having a single node that uses all its disk space while the other nodes are fine implies that there could be underlying issues with the node.