Take a bell curve with mean 50 and standard deviation 10. Mark 40 and 60, one standard deviation either side: about 68% of the data, roughly two-thirds, lies between those marks. Widen to 30 and 70 (two SDs) and you capture about 95%. Widen to 20 and 80 (three SDs) and nearly everything, 99.7%, is inside.
The tails come from subtraction and symmetry. Outside ±1 SD is 100%−68%=32%, split into 16% on each side, so about 16% of values exceed 60 and about 84% lie below it. Outside ±2 SD is 5%, or 2.5% per tail; outside ±3 SD is 0.3%, or 0.15% per tail. Halves work too: between 50 and 70 is half of 95%, about 47.5%.
The percentages belong to the normal shape, not to every data set that has a mean and an SD. Suppose waits at a clinic have mean 15 minutes and SD 10 minutes. The rule would put 2.5% of waits below −5 minutes, which is impossible, so the data must be skewed and the rule does not apply. A boundary that is not a whole number of SDs, such as 'below 63' in the first example, is also outside the rule's reach; it needs a z-score and a normal table.
Think of it like this. The rule works like a recipe that scales: double the batch and every amount doubles, but the proportions stay the same. Slide the mean or stretch the standard deviation and the bell moves or widens, yet the shares stay 68, 95 and 99.7 within 1, 2 and 3 SDs. Change the shape from bell to lopsided and it is a different recipe.
Core idea. For a roughly normal distribution, the empirical rule gives the percentage of data inside one, two, and three standard deviations of the mean: about 68%, 95%, and 99.7%.