Do TAS runs exist that only use tricks a human could do?

In general, as serious TAS goes, they will disregard human limitations, using tricks that often go beyond frame-perfect and pixel-perfect (things theoretically achievable by humans but too improbable) and go into inputs that would require hardware modifications - for example simultaneously inputting left and right, or providing analog stick output corresponding to angle that would require removing the case from the controller, because the stick just doesn't normally bend that far. They are a speedrunning category of their own.

Of course there are games simple enough, that a TAS won't have any opportunities to do things a human can't - say, the Atari 2600 racing game "Dragster", where the number of inputs is so limited a human can achieve the perfect set of inputs and there simply isn't anything left a TAS could do better.

But TAS isn't only used for competitive TAS speedrunning. It's also a tool that is used in assisting development of regular speedruns - one could say, as a more graceful alternative to video editing. A recent example of this use is LOTAD (Low-Optimization Tool Assisted Demo) of Ocarina of Time by ZFG, which is a demonstration of new routing, a "blueprint" of upcoming speedruns. It's a hybrid of a segmented speedrun and TAS; ZFG performs a segment multiple times until he achieves desired result (not necessarily in optimal time but without major errors), and records the inputs in the process. The sequence of inputs from a successful attempt goes into the TAS script, then he moves on to the next segment. Obviously the final result is neither a legit speedrun submittable to the leaderboard, nor a competitive TAS, being less optimal than dedicated ones. But it's a valuable guide, a resource for speedrunners to use as a guide in developing and training the route, a complete low-level documentation of all inputs required to finish the game in record time. And indeed, using this route, ZFG has already managed to snag a legit 3:23:55 WR in the 100% category.

No, Tool Assisted Speedruns rely on the same inputs a player can perform, the difference is that the TAS can perform frame-perfectly every time. It is assumed that TAS can perform tricks that players cannot but with enough practice, skill and luck all inputs that a TAS makes should be repeatable by a human player too.

For an excellent example of this, I'd like to showcase the ability for Mario to jump through walls in Super Mario Bros (1985). This game is perhaps the most popular game for speedrunning and tool assisted speedruns had made use of tricks that players were assumed incapable of repeating. One trick that the TAS used was to clip through walls by timing a frame-perfect, pixel-perfect jump, and immediate crouch which pushes Mario through a wall. It was assumed for a while that players could not repeat this because the timing was so extreme. It has since been repeated by players and has helped shaved about a second off of one of the most optimized speedruns in existence.

For a quick demonstration you can skip to 1:59 in this video:

The point being, it was assumed that this was impossible for humans to replicate, but since TAS runs have the same tools as humans then humans are subsequently able to replicate the TAS moveset.

yes, there is one type, and it's not really considered tas, but well, it's still tool-assisted.

it's called a LOTAD. a Low Optimization Tool-Assisted Demonstration.

essentially, it's showing that something is possible to do by humans, but it's not going frame by frame. usually, it's done by slowing the game down and using save states. sometimes it uses on-screen speed displays or location displays, but it shows that something is possible by humans with enough practice.

there are many videos on the art of tasing, and most of them are really informable.