Most people meet Andrew Ng through the Coursera Deep Learning Specialization, which is gentle, well produced, and deliberately holds your hand. CS230 is what happens when the same person teaches the same material to Stanford students who are expected to keep up, and the difference is instructive. The course runs on a flipped model, so the heavy lecturing on mechanics happens in the Coursera videos you watch beforehand, and the in-class time goes to the things that are hard to get from a textbook, how to choose a project, how to read a learning curve and decide what to do next, how to set up a train and dev split that will not lie to you. That emphasis on strategy and judgement is the real value here, because the mechanics of building a network are now well covered everywhere, and the part that separates people who get models working from people who do not is exactly the decision making that this course foregrounds.
The lectures move through convolutional networks, sequence models, and the practical engineering of training, and Ng's gift for finding the one explanation that makes a concept click is fully on display. The catch is structural rather than intellectual. CS230 is not a single tidy package, it is a set of lectures, slides, and projects that assumes you are also doing the Coursera assignments for the coding, and if you are auditing from outside Stanford you have to stitch that together yourself and you will miss the graded projects and the teaching assistants who make the experience cohere. It also assumes a real starting point, comfortable Python and the basics of machine learning, so it is not a first course.
My honest recommendation is to treat the Coursera specialization as the on-ramp and CS230 as the thing you graduate to, because once you have the mechanics down, watching Ng talk about how to actually run a deep learning project is some of the most useful free education available anywhere.