Using the second definition sounds more potent right off the bat,
 but then we are faced with another issue:
How do we measure skill?

So, instead, lets use something in a similar spirit, but that we can measure.

"Imposed challenge" happens to be easily quantifiable.

For example -with a bit of telemetry- we could record each attempt at every level
 and a boolean outcome.
While at it,
 we should also grab where on the level the players fail (think MarioMaker),
 and on which levels they end their sessions.
This later, I will call the retention data.

Lets replot:
    Dark Souls 3    Witcher 3
    ▲              ▲        
    │         Fail │           
    │         rate │           
    │^^^^^^^       │        
    │              │....... 
    ┼───────▶      ┼───────▶
     Bosses         Quests

I mean, I can infer intel from that, that's good.

What if I look at my own game again?
   Curve Stomper 2
    ▲               
    │ A      C         
    │.'. ....'. .      
    │   'B     'D      
    ┼────────────▶

Not very smooth, lets review it:
    A:
      The difficulty spikes.
      Apparently, my players are twice as likely to get caught here
       by the CIA than elsewhere.
      Consulting my retention data, I can tell that people are rage quitting.
      I should lower the difficulty.
    B:
      The difficulty drops.
      I think the few people that died here, did it intentionally even? Huh.
      Again, my retention data shows that people are getting bored
       and closing the game.
      I must add stimuli.
    C:
      The difficulty spikes.
      I will have to lower the difficulty, yadda, yadda.
      Wait, no, people are not quitting!
      They appear to like the challenge.
      Thinking about it, it plays well with the story pacing.
      I won't be touching this one.
    D:
      The difficulty drops.
      Again, my retention is high.
      Now, it could be that they are still high on C,
       or perhaps they really need room to breath.
      A common pacing issue is having constant high stakes that tire people out.
      I think I will experiment with inserting more, somewhat easy sections,
       to see whether that improves the over all retention.
      I will also swap D with a normal difficulty level to see how that influences stuff.

Ok, but where does all this leave us?
I am looking at flat lines, and I'm trying to make them the right amount of scribbly.
What happened to the nice cool bell-curve subsections?

I mean, if I were to implement the increasing graph I was shown at the start,
 that would mean that the start of my game is boringly easy
 and the end is buttfuckingly-hard.

Yet, some insist that even if players dont enjoy it, its a good thing:
    https://www.moddb.com/tutorials/reflections-on-difficulty-pacing-in-puzzledorf