Hi Ahmet:

1) Remember that the assumption of normality that most of the parametric tests require, is on the residuals of the model, not on the data itself. For example, you perform an analysis of variance to compare the means of 3 groups, then you take the residuals of the anova and perform a shapiro-wilks (or anly other normality test) to test the null hyphothesis that those residuals come from a normal distribution. If they are normally distributed you stick to the anova, if not, you might think to perform a Kruskall-Wallis test.

If you directly choose the non-parametric you need to explain why. Sometimes normality is not the only reason, it might also be non-homogeneous variances, low sample sizes, or that the purpose of the test is not making predictions, but only detecting a difference. Keep clear in mind what is the purpose of your test.

2) It also depends on the purpose. You can make univariate normality tests (only one item) or multivariate normality tests (more than one item). If you will conduct univariate tests like Kruskall-Wallis, then stick to the the first, but if you are willing to conduct multivariate tests like manova, permanova or so, choose the second.

It is not the parametric or non paramemetric, it is the fitness of method your problem.Regards