I'm grad student about to finish their PhD, so you may find my perspective valuable.
For reference, my undergrad degree was in physics and math, where I did big data analysis on high energy data gathered at CERN. I had some much earlier experience working in a condensed matter lab, yet gained very little (to no) chemistry experience. To date, I have never actually taken a condensed matter course -- and don't intend to. I likewise don't have any specialized training in chemistry, and never took a chemistry class.
Despite this, I feel that I've developed a strong enough understanding of the field to curate my own interest and pursue my own problems. This is in part due to a research mindset that I inherited from my supervisor, and in part due to my own personal preferences toward learning.
The mindset is as follows: Find some problem that interests you enough that you'll throw everything you got at it. Maybe you know nothing about the field -- no general knowledge, no specialized techniques, absolutely nothing. It's like staring out at an ocean -- and condensed matter is a vast ocean -- and not knowing how to pass the beach. Don't fret. Pursue your problem with as much zeal as your interest can muster. Learning everything related to this problem; and with time, effort, and some help, you'll find before you a tiny island of knowledge, deep in the scientific ocean.
Some people might say (and you might feel) that it is necessary to take a class about X, Y, and Z in order to even start doing proper research. But homework -- simple problems that have a clear answer in a book -- will only have you playing in the shallows. Yes, you should calculate and do problems, perhaps even from a book (or paper) if it helps, but always in pursuit of your problem.
These smaller problems orbiting your goal will form the beaches of this tiny island of knowledge. And it is that island which will define your unique perspective as a researcher, and allow you to see farther in the scientific ocean than you could if you stayed in the shallows. And with time, and more island forming research in the deep, you'll have an archipelago to call your own.
Essentially, what I mean by this bit of philosophical mumbo jumbo is: dive deep into an interesting problem. Your own efforts in solving the problem will provide you the ground-level knowledge needed to do other things in your field. And perhaps most importantly, it is the only way to actually learn how to research -- doing research.
This is definitely just my perspective, one which is informed by my experiences. No doubt I am on the extreme end of being class anti-oriented.
As a personal example, I started my theory career studying the Fe-based superconductors (pnictides and chalcogenides). In particular, I chose one material in that broader family -- FeSe -- and studied the hell out of it. Not only are many of the Fe-based superconductors electronically similar, in part owing to the similarities in their symmetries, but I find that the tools I learned translate to other unrelated materials as well.
In real materials, there are also many different types of measurement techniques and experiments: conductance, photoemission, scanning tunneling microscopy, to name a few. I find it can be overwhelming at times, especially when some critical new research in the field depends on a set of techniques which are entirely new to me. But it's in these novel moments that I find the chance to develop my best physics intuition. And a lot of the time, these many different measurements can be related to a few key concepts -- e.g density of states (DOS) shows up everywhere.
Lastly, as a new theory researcher you may consider exposing yourself to cutting-edge numerical methods, such as the Density Matrix Renormalization Group (DMRG), machine learning (ML), or Density Functional Theory (DFT). It's a bit of a myth that you need to invest your entire graduate career in order to learn these things, or that you won't have the opportunity to learn them if your supervisor isn't an expert. There are actually a number of readily accessible tools and codebases from which one can start: QuantumEspresso=DFT, ITensor=DMRG, and Keras/TensorFlow=ML are just a few that I've used.
These toolsets aren't just useful blackboxes either, but a fully functioning code that has the potential to teach you the method. It can be challenging to build up a codebase from scratch, as there a lot of new things to learn, and all the pieces have to fit just right in order for it to properly compute. But a fully functioning codebase can have pieces rewritten piece by piece, with the understanding that if it doesn't compute, then it's your piece that's bad. There are also tons of related workshops where you can learn from the pros.
This top-down approach to novel numerical methods is how industry teaches machine learning, since it's pedagogically faster, and allows one to get results while they learn. If you're supervisor talks about one of these techniques, but they themselves don't have experience coding them, then they'll likely be super impressed if you did. You might suddenly find yourself becoming extremely valuable to your research group.