Should coin flips be modeled as Bernoulli or binomial draws in RJags? What is the best way to model coin flips as a hierarchical model? Do you say coin draws are a series of draws from Bernoulli trials or as one draw from a binomial distribution?
That is something like this:
model {
  p ~ dunif( 0, 1 )
  for( i in 1 : n) {
    h[i] ~ dbern( p )
  }
}

or, this:
model {
  p ~ dunif( 0, 1 )
  numberOfHead ~ dbinom (totalTrials, p)
}

 A: One draw from binomial distribution generally is enough. But it depends of the data you have. If you have the data of how many heads in the individual coin flips have been seen in total, then binomial distribution is enough, no need for detailed model with N bernoulli flips. However, if you have data on results of individual coin flips and you need to distinguish them (e.g. because of covariates for individual coin flips), you will need more detailed model with bernoulli distribution.
A: Both models will give the exact same results.  Why? The Likelihood principle.  RJags is an R package that uses the software JAGS to conduct Bayesian inference, and any fully Bayesian procedure, one where inference proceeds from the posterior distribution, will satisfy the Likelihood principle.  Essentially, the Likelihood principle states that if two likelihood functions are proportional to each other, then the same inference about the parameters should be obtained from the two likelihood functions.
In your example we are inferring the probability of a coin landing heads up, $p$,  from $n$ independent tosses, $X_1,...,X_n$, of that coin.  Prior to tossing the coin, you assume that any value of $p$ in the interval $[0,1]$ is equally likely.  Thus the prior distribution for the parameter $p$ is $\pi (p)=1$. Suppose we observe $k$ coin tosses where the coin lands heads up, where $0\leq k\leq n$. In the case of the model using the binomial distribution, the likelihood function is
$$
l(p|X_1,...,X_n)= {n \choose k}p^k (1-p)^{n-k}
$$
For the Bernoulli model, the likelihood function is
$$
l_\star (p|X_1,...,X_n)=p^k(1-p)^{n-k}
$$
We have observed the data, so both $n$ and $k$ are fixed values and therefore ${n \choose k}$ is just a constant, and $l(p|X_1,...,X_n) \propto l_\star(p|X_1,...,X_n)$, bearing in mind $l$ and $l_\star$ are functions of $p$. Once we have our samples from the posterior distribution from RJags, we will make the same conclusion, aside from any error due to having a finite sample from a Markov Chain that has hopefully converged.
Also, if you are familiar with sufficient statistics, you could note that $k=\sum_{i=1}^n{X_i}$ is a sufficient statistic for $p$ in both models (assuming $n$ fixed).  
