Did you read the rest of the post? He goes on to say you can do that. In fact the entire message of the post is that entering into the container and running commands and then committing at end is more efficient (In the terms of numbers of layers and disk space, at the cost of what else I don't know yet) than using a Dockerfile and 'docker build'.
The main cost is less clarity: the build steps aren't isolated anymore, so it's harder to pinpoint issues. There's also obvious risk of my script not interpreting all options correctly.
Actually, I've just disabled VOLUME statement in the script, as it seems to be a no-op in Docker. Only trace it leaves in the image is setting image's command to '/bin/sh -c "#(nop) VOLUME [\"/data\"]"'.
