Docker is a tool for lightweight virtualization through images and containers. Docker images are created by performing a build, controlled by a source-level artifact called a Dockerfile. We studied Dockerfiles on GitHub, and—to our great surprise—found that over a quarter of the examined Dockerfiles failed to build (and thus to produce images). To address this situation, with the goal of reducing the number of broken Dockerfiles on GitHub, we created Shipwright, a human-in-the-loop system for finding repairs and suggestions: after a specific build fails, one runs Shipwright to obtain either repairs or suggestions for possible fixes. We used a modified version of the BERT language model to embed build logs and to cluster broken Dockerfiles. Using these clusters and a search-based procedure, we were able to design 63 rules for making repairs and suggestions. With the aid of Shipwright, we submitted 45 pull requests (with a 42.22% acceptance rate). Furthermore, in a ``time-travel'' analysis, we found that Shipwright proposed repairs that match ground-truth patches in 23 out of 102 cases. Finally, we compared our work with recent, state-of-the-art, static Dockerfile analyses, and found that static tools detected possible build-failure-inducing issues in 20.6–33.3% of the cases we examined, whereas Shipwright was able to provide repairs or suggestions in 73% of the cases.