You won't find any such rule in all liklihood. The conventional wisdom (given that statisticians always disagree on everything) is that there is no specific level at which you can automatically remove outliers. Generally, you need to try to explain why the outliers are occuring (if they are not simply a mistake in data collection). Commonly extreme outliers reflect model mispecification, multiple populations etc. Three standard deviations is a useful tool to identify extreme outliers, but you use it to explain why they occur rather than removing them.
I spent a fair time reading various material on outliers. It is common to see warnings about automatically removing them.