Stockfish analysis vs suboptimal play

From what I understand, when analyzing positions Stockfish assumes both sides will play best moves and then provides the best line based on white or black to move for the given position. It will basically play itself.

I'm creating my own studies for openings I want to learn & play and I notice something when combining Stockfish against most common responses by opponent given the position:

There are a lot more people than I thought playing Stockfish-esque lines

Most responses are reasonable...until I get a few moves deep and things clearly become non-human. This is throwing off my opening preparation.

How do I get decent analysis lines when Stockfish thinks it's playing against itself?

Most engines are always going to assume best play.

So how do I figure out how to prepare the study? If I played some of these moves, I'd accuse MYSELF of cheating- LoL.

Find lines where you understand the ideas and don't have a major weakness according to the engine.

You're not if you're playing a Daily game in the line. If you're just studying openings you certainly can

This is absolutely not true. Engines don't assume anything. They analyze all possible moves (well, not exactly, but close).

I think you miss the point. The evaluation of each possible move assumes the opponent will play the best move it has found at every step. It has looked at many inferior moves, but they are omitted from the determination of the optimal line. You can see the present state of this line for the position after each candidate move when watching an engine analysing a position.
When analysing, it is looking mostly at what it initially considered inferior lines initially, exploring them a little deeper and broader. Occasionally this leads to a change of evaluation that makes a line replace the current optimal line from some point - can be for either side.

It's worth saying that a very slow version of the same process is very much the way high quality human analysis progresses. For example in a high quality vote chess game (say minor master level), a discussion tends to select a few candidate moves then construct quite deep likely lines which are as close to optimal as can be seen. Throughout the discussion revisions occur which change the lines, occasionally changing the choice of candidate. Humans can somehow get by (if not to engine level) with extremely small numbers of lines being consciously examined. This relies on very selective choices of candidate moves throughout by a level of intuition that engines lack. (That being said, AI engines like Leelachess come closer, by having very highly developed "intuition" - NN evaluation models, enabling them to be very strong with far fewer nodes being examined than a non-AI engine).

I think you miss the point. So far, nobody was talking about the evaluation score, the discussion was about the variations themselves.

Granted, I don't understand 100% what the OP wants to say and what his problem is. I think that preparing openings by looking at Stockfish lines is a bad idea.

The discussion of using an engine on opening lines is different, short of avoiding outright blunders. Yes, there is nuance about the engine horizon whether or not a given move is best at depth, but everyone top move the engine suggests, is considered the best move, by the engine, under the constraints mentioned.

Sone engines may work a little differently, going for lines where the opponent has more chances to go wrong and not the evaluated best move, but even in those, it's still assuming the other side is going to give the best reply. If engines didn't do that they wouldn't be very useful.

Thank you

The evaluation number is based on the best defense for the opponent. But that is obvious. On the other hand, your second statement is not true. Engines don't assume anything. Instead, they try to pick the best move that works no matter what the opponent plays. If the opponent plays a lesser move, that's worse for the opponent.