But to take something I have recent experience with: to get a 5-month-old baby to learn to balance his torso using his core muscles, spend time holding him by the thighs and let him flop around (it doesn’t help to support the baby at the top, since that is a passively stable configuration). After a couple weeks he’ll be somewhat able to stabilize himself and compensate for small perturbations, so start tilting him in various directions so that he’ll have to learn to compensate for being more aggressively off-balance.

After a month of occasional practice (a few minutes at a time every once in a while) the now-6-month-old baby will be stable enough to sit up on Dad’s shoulders without falling off, if held by the legs. Then just walking around with the baby on your shoulders will be great balance practice, and requires no special effort at all (but occasionally rocking a bit will help him to keep improving his strength and balance). By 8 months old the baby will be very good at balancing sitting on shoulders.

[I mostly care about this because carrying a baby on my shoulders is much less tiring than carrying the baby in my arms; whatever benefit the baby gets out of it is somewhat incidental.]

This kind of process works pretty well for learning to cook a recipe, solder electronics, play an instrument, touch type, or the like.

If you want to learn to solve math problems or write poetry (i.e. something that is largely non-physical) then the context, actions, and feedback are rather different. But still, you want to get your brain as much data coming in as possible, strengthening connections between ideas and concepts, experimenting and doing more of what works, ...

One of the big problems with formal schooling for many students is that feedback comes much too slow to be usefully integrated. The expert advice from teachers (results of graded papers) is not interactive, but is low-bandwidth and very delayed. Regularly sitting down 1:1 with an expert is a much faster way to improve.