Binder needs only one thing for images to work:

• to be able to launch jupyter notebook with jupyterlab as a specified user (passed via docker build args as NB_UID/NB_USER)

What this means in practice is that the notebook and jupyterlab packages must be installed and on PATH:

RUN pip install --no-cache notebook jupyterlab

Note that if you are not installing jupyterlab in addition to the classic notebook, you'll need to change your mybinder.org URLs from /lab to /tree as described here. Otherwise, you might get a "404: Not Found" error when launching your project on binder. Now that's almost everything.

The remaining piece is that the specified user must be able to start the notebook, which requires certain permissions like being able to write to the home directory.

The absolute bare minimum for this is to set HOME to /tmp so that it's writable, which would make the shortest, smallest Dockerfile that works on Binder:

FROM python:3.9-slim
RUN pip install --no-cache notebook jupyterlab
ENV HOME=/tmp

which you can try out:

However, it would be better to consume the NB_UID/NB_USER arguments and create a real user:

# create user with a home directory
ARG NB_USER
ARG NB_UID
ENV USER ${NB_USER}
ENV HOME /home/${NB_USER}

RUN adduser --disabled-password \
    --gecos "Default user" \
    --uid ${NB_UID} \
    ${NB_USER}
WORKDIR ${HOME}

From this point, you can start adding files, installing packages, etc.