Bayesian networks (BNs) are compact representations of probability distributions that allow for supporting reasoning and decision making under uncertainty. Their interpretable structure and probability parameters allow for integrating human knowledge in their construction and explanation. For BN construction, reusable building blocks, or idioms, exist that describe the dependencies and reasoning patterns among small sets of variables. In this paper we formalise the concept of an idiom, explicitly including qualitative constraints that capture the reasoning patterns among variables as stated in the informal descriptions that accompany the idioms in literature. Our proposed formalisation ensures that idioms can be applied more consistently and reliably, improving the BN’s accountability.