You’ll need to emulate enough of the machinery to generate your own installable package, as your custom packages would be clobbered by the “real” ones. See this thread for some more discussion.
Another, far lighter-weight approach, would be to fork/extend the extensions in question, and build renamed packages which replace the core ones in the dependency injection stage at startup by specifying the package.json#/jupyterlab/disabledExtensions keyword. This will allow you to install the upstream, and maintain (hopefully) a much smaller set of changes/workflows.