I agree. I have provided Dockerfiles in work projects, and it was successful. However, I found it difficult to run tasks using just the Dockerfile. The containers became too large. And I had to write too long arguments for volume mounts and environment variables. In the end, I had to include the steps for starting and running the containers in the README.

(I am deeply concerned about this issue. So I implemented a little tool called Lask, and it tried to solve this kind problems. https://github.com/lask-task-runner/lask )

I forgot to mention it but a docker compose is even better to mitigate incantations. It seems every project I use that has a docker file also has some custom script around it that boils down to mounts and variables

I agree. I was writing a task runner, and when I used generative AI to write Haskell code, it tended to produce higher-quality code compared to other languages. In particular, I’ve rarely encountered runtime errors when testing the code. On the other hand, when it came to design, there were errors where my intentions weren’t conveyed properly.