The best practice would be to use a virtual environment (e.g. conda, pipenv etc.). This way you can delete the environment and create it from scratch in case of breaking modifications.