1) Slow, untimed solves. This is effective regardless of the method used. Work on consistency of steps, looking ahead, not turning the whole cube very much, and turning smoothly. If you wanted to take this idea to the next step, then start solving with a metronome dictating your speed, and then gradually increasing the speed while maintaining smoothness.
2) Work on your first four edges as an automatic reaction to what you see during inspection. Lots of random scrambles done slowly may start to get you familiar with what you have to do to keep your movecount down and cube turns to a minimum. Not everybody is going to start color neutral, so maybe at first you're just going to look at your target center and the four edge pieces that go with it for a lot of scrambles and see what the easiest way there is. However, some of them are not intuitive at first, and you might gain that knowledge by watching some machine solved cases. One of the extra features of the Android app Twisty Timer is that when it generates a scramble for 3x3x3, you can press the "Hint" button and it offers optimal solves for the first four edges for any of the six colors. Once you think that you can formulate a good plan, the next thing to do is to figure out if you can execute the plan without having to look at those pieces, because you need to start looking for what's next - the first corner/edge pair. For Roux solvers, apply this to the first block, and start looking for the second block.
3) As you execute the moves for each corner/edge pair to build the F2L, you need to locate the next pair of pieces for the next corner/edge pair. If you can see farther ahead than that, that's great. If you feel you're not doing something smoothly, look up a couple of different algorithms for that case and see what feels good. While the F2L should be done intuitively much like the first four pieces, it still may be helpful to learn some algorithms for the more difficult cases. For Roux, the next step is the U layer corners, and is its own very difficult subset of algorithms, nearly as difficult as the last layer algorithms in CFOP, and probably will get accomplished using the main feature of item #4 below.
4) For the last layer, my suggestion is drill, drill, drill.
Once you learn an algorithm, go on youtube and watch a couple of different people do the moves to look at move groupings and hand positions.
While the OLL is done before the PLL, and it might make sense to learn things in order, it's more important to learn the PLL first. Early on you will only make use of only three or four of the OLL cases, having to use them in combination and even multiple times in some cases.
Once you feel like you have enough moves put together to finish the last layer, use the qqtimer subset feature to generate scrambles for the last layer.
For Roux, there is a different subset you can use that just scrambles the last six edges - it's labeled "Roux-generator
Don't forget, there's always more to practice (and not just at cubing).