Foundation models have become ubiquitous in computational biology research. These models take large-scale biological data such as DNA sequences or quantified single-cell RNA molecule counts as input, and learn high-dimensional embeddings that capture patterns in the data through self-supervised training, which are then used for a variety of downstream tasks. Despite their great promise, we found single-cell foundation models were difficult to install and run, and lacked adequate documentation, which we outline in several observed drawbacks. To alleviate these problems, in this work we propose six recommendations for better development and deployment of these foundation models to enable ease of use by both computational and non-computational expert end-users, as well as a framework for improving the state of open-source software maintenance in computational biology as a whole.