For competive server you just make all godlike and this all. Godlike is maximal level for bots. But you can lost in gameplay variety.
For example WarCow weak when on foot, but very like vehicles and kill you when inside it.
DP|Farham is not greatly skilled in terms of skill points but very noticeable among other human bots with same skill. And so on.
So bigger skill points not necessarily mean more interested gameplay.
Even with humans - if all play same in same way with same tricks it can be boring. Variety is important things.
Also it is reason why exists things like AlternatePaths and so on.
Also in code picked all bots of bigger skill and choose random one from it - exact by this reason.
Because all time play against Xan can be boring. So there be WarBoss and XanMk2 at least for some variety. At least in names.
Biggest problem of bots - they not learn anything.
This mean - if you able win - it is forever.
If you catch them on choke spot, you will able do this again and again.
Which not happen with humans.
You almost never know what humans do next time, since they able learn.
I not found good way for change it, but have some ideas like change weight of AlternatePaths depending on success or loose for bring flags.
But this require a big count number of attempts for work. And there many pitfalls about it.
I have interesting idea about bot aim. Unfortunately this can not be implemented without replace it, so not suit to this mutator.
Idea consist in add bot aim pinged. Instead of use instant aim with zero ping and artificial aimerror, use prediction.
Bot use your past position on some moment, like 20-50 ms ago (configurable), and try predict your next pos. This pos is be aim pos with lesser aimerror from now.
This lead to more human aim.
If you move predictable - bot can more easier hit you. If you move unpredictable - bot not able so easy hit you.
Currently for bot this doesn't matter. On shot moment all world stopped and bot not care what you move be before that. Just pure random of aimerror.
Also this approach possible able fix well-known problem about dependency aim from distance to target.
Humans aim bad on long distance, good on middle and very good on close, but bad on very close. So if you get very close - hard to hit you.
Bots not experience this. Dependency from aim straight linear. If you very close bot hit you in 100% times. Which not human-like behavior.
But if use prediction, bot will not able predict your movement properly on very close range, since small mistake can make big change in result pos.
Unfortunately this can not be done with stock bots.