The best characterization I can currently work out is: Fix a proper (not necessarily nontrival) coset $H$. Form a disjoint union of sets of the following form: Choose a subgroup $H'$. (not necessarily proper) that strictly contains $H$. Choose a coset of $H'$, and for each coset of $H$ within that coset choose a representative. Then this disjoint union is not good.

Proof: Let $k$ be the smallest element of $H$, then the elementary cyclotomic polynomial of $k$ divides the polynomial of each such set of coset representatives, so it divides the polynomial of a disjoint union.

I do not know if this characterization is complete, but it seemed too long to fit in the comments.

I don't think there is any nice characterization for this problem.

$1+x+x^{7}+x^{13}+x^{19}+x^{20}$ isn't relatively prime to $x^{30}-1$ despite not being characterized by any sort of reasonable construction.

It comes from adding two cosets and subtracting one, like so: $(1+x^{10}+x^{20})+(x+x^7+x^{13}+x^{19}+x^{25})-(x^{10}+x^{25})$. You can't get it by just combining cosets. Choosing elements from cosets doesn't help because that would be the wrong elementary cyclotomic polynomial.

This is equivalent to the problem: When does a (multi)set of roots of unity sum to $0$? You shouldn't expect a particularly good answer. Compare it to the problem "when does a set of complex numbers of norm $1$ sum to $0$", which just asks you to find polygons of side length $1$. There are quite a lot of these.

This problem is more tractable when you can mandate that the size of the set is small. Then you can give an explicit construction for all bad sets. But as far as several graduate students who worked on the problem in the second form can tell, there is not any construction that works in general that is better than "take multiples of an elementary cyclotomic polynomial that have only $0$ and $1$ for coefficients."