Engines are weak at openings, and they are weak at evaluating endgames. For example, it will say that Bishop of wrong color and Rook pawn versus lone king is something like +4. Those that do know that it's a draw have tablebases built into them and it's not the bot figuring it out. Same thing with openings, they have an opening encyclopedia programmed in.
Engines value material too much over positional advantage, and they lack positional understanding.
Engines are best at calculating concrete combinations, possibly even 30 moves deep if it's all for the most part forced.
You are best off analyzing your own games at a board, away from the artificial intelligence, making notes, and only then turn on the machine, and enter each move slowly to see whether the evaluation of the position changes drastically. For example, if you enter a move for White, and it goes from +1 to -6, White almost surely blundered, and you can use the bot to find the combination for Black that wins. However, if it changes from say, +0.24, and after a move by Black, it went to +0.36, there likely is nothing wrong with Black's move, and maybe the bot realized something that it couldn't see without the knight moved. Again, it lacks strength in positional understanding unless you program additional information programmed into it, similar to opening books and tablebases (for endgames). Engines alone are horrible at such evaluations.
While analysing games , it has been observed that not always does computer point out the best moves and also misses some of the creative moves? In such a case should an amateur only focus on blunders pointed out by the player or also look at the inaccuries and mistakes (As sometimes it seems that computer does not look into the game plan or sacrifices in a holistic manner or as a part of the whole ). What is the best way in such a case ?