You use the available functions to build up a generator for your data type.

First, let's give the instanc itself. For this I'm just going to use the frequency function to use "featureGenNormal" five times more often than "return FeatureError". This means that will get a FeatureError every now and then, but mostly you'll get featureGenNormal (see below). You can change these frequences, of course.

instance Arbitrary Feature where
    arbitrary = do
        frequency [ (5, featureGenNormal),  (1, return FeatureError) ]

In order to write featureGenNormal, we need to be able to generate random values of each of the parts of the data type. Often these types will already have Arbitrary instances, so generating an isntance for your type is quite often just a matter of calling "arbitrary" for each component, and then returning a datatype. However, there is no Arbitrary instance for String, which is a bit annoying, so let's write our own generator for strings.

First a generator for a single letter:

letterGen = oneof $ map return $ ['a'..'z'] ++ ['A'..'Z']

Then a combinator for generating a list of values given a generator for a single value:

listGen :: Gen a -> Gen [a]
listGen g = do
    x <- g
    xs <- frequency [ (1, return []), (10, listGen g) ]
    return (x:xs)

And then we use this to build our "stringGen" generator.

stringGen :: Gen String
stringGen = listGen letterGen

Now, we have all we need to write the featureGenNormal generator:

featureGenNormal = do
    id <- stringGen
    name <- stringGen
    featuretype <- arbitrary
    grouptype <- arbitrary
    children <- arbitrary
    properties <- listGen stringGen
    return (Feature id name featuretype grouptype children properties)