Top Down. To some it probably feels gamy that a wall should become harder to climb as you level up, but I create obstacles where they are meaningful. And if the DC of a climb becomes meaningless, the obstacle may as well just be part of the narrative.
The DC's are for actual challenges (whether it be an obstacle in the middle of combat or a skill challenge). I decide, this wall is going to be hard difficulty to climb, PC's are level 14, DC should be 23. At this point, I decide what the DC 23 wall should look like. It's a stone wall that's been weathered down with wind and sand, so the stones are smooth, what may once have been seams in the wall are filled with sand and calcified, hand holds are difficult to find.
Then I begin to cheat, and look at the Athletics skills for the PC's, three out of five have +18 or higher, and the lowest score is a +12. So this is not going to be very challenging for them. To add a little more excitement, I decide it's going to start raining just as they get to the wall, making the climb harder, increasing the DC by 2 for the first half the encounter, and by 5 for the second half.
If this was say a task requiring Thievery skill, and I realized the highest bonus in the party was a +9, I might be inclined to change something in the opposite direction, perhaps part of a mechanism is rusted or cracked, making it easy to smash and take out of the equation, and I'd lower the DC by 2 or 5.
So I basically start with the numbers, double check what the PC's are capable of, and then build in the picture based on the setting the challenge will be in.
I have the same thoughts as you do on this subject.
Regarding shifting how difficult something is according to how good the party is at the skills - well, I don't do it too much. I like it when a player goes: "I can do it, I am great at this stuff!". I might just shift it +/- 2.
If I want something to be more difficult I am more inclined to add other problems with different skills into the mix.
My players sometimes want to use different skills, for instance using bluff instead of diplomacy. I usually allow it, but up the target DC by 5 and adding some amusing point if it fails. It can't really fail doing it like that. Either the players fail in some spectacular fashion, or they succed in the same way. Both me and my players usually get some giggles out of it.
I remember a gully dwarf in a dragonlance campaign that bluffed some guards into taking care of the problem upstairs, while he would take care of the alert-gong. The DM was really good at explaining why the bluff worked and it was such a memorable moment.
