I've spent some 40 years learning electronics. Started with the hydraulic analogy, but didn't really have any deep understanding. In college they made us do a physics lab with an ocope and RC circuits, and I was miserably. Who cares if you can use discrete components to make a curve when you have a 486?
Years later I landed on a team of makers and started to work on ambitious projects- various high power LEDs, motion control systems, etc. This is where my gap in knowledge- especially wrt high power electronics, diodes, and any chip-based component- became a real problem. So I built ambitious stuff and when it didn't work, or was flaky, I'd show it to somebody who knew electronics deeply and they'd explain whatever the next thing on my list to learn was- pull-up resistors, constant current supplies, MOSFETs to control high power devices, connecting up an SPI bus between a microcontroller and sensor, schmitt triggers, bias, etc.
I got really good at making small repros of larger projects, handing them to somebody, having them solve the basic problem, then taking the learnign from the repo and putting it into the real project.
Eventually, after doing that a lot I was able to to read Art of Electronics and got a lot more out of it. most recently I was building a custom circuit to drive a vacuum tube and had some problems, and somebody mentioned Spice. So i got LTSpice and put my circuit in, and learned just enough to have it spit out what I was seeing in real-life on my oscope. OH MY GOD, it was a revelation. The simulation produced exactly what I saw and I quickly debugged the problems. This has always been true- if I have a simulation, I can learn to intuit how things work faster than if I have to assemble them manually.
My mental model now is all about modularity- building the individual bits of a larger circuit, debugging those, then integrating them together. There are so many details in analog that you have to be aware of; if you're missing a pull-up resistor or a schmitt trigger or have too much EMI, you might get it to work 30% of the time or have to do bounce elimination in software.
Another thing to be aware of is in the past 20-30 years, a lot of discrete components gained alternatives that were chip-based and move a lot of the smarts into the chip. Sure, you can build an H-bridge from components to make a bidirectional motor driver. But that motor driver from Pololu has decades of intelligence about driving motors- and reverse polarity protection (I plug things in backwards all the time) and self-limiting (if you push too much power, it shuts down, instead of frying). I've had problems with components that could have been solved by an EE.
I don't know if I could actually design any non-trivial circuit, but then, what exactly do you need to design today? Most of the work is in identifying what your problem is, then finding the existing solutions.