Drop thousands of balls through a Galton board, a board of pegs where each ball bounces left or right at random at every row. Most balls get a mix of lefts and rights and land near the middle; a ball that bounces the same way every time is rare and lands at an edge. The pile that forms is a bell. Heights behave similarly: a height is the sum of many small, separate influences, and sums of many small random pieces tend to be normal.
The bell's shape is fixed; only its center and width change. Change the mean and the bell slides along the axis; increase the SD and it spreads wider and flatter, because the total area under it is always 1. That fixed shape is why the 68-95-99.7 rule fits every normal distribution. If heights in a group have mean 162 cm and SD 7 cm, about 95% lie between 162−14=148 cm and 162+14=176 cm.
Not every single-peaked pile is normal. Household incomes peak at the low end and trail off in a long right tail, because a few very large incomes stretch it. For skewed data the mean sits away from the peak and 'about 95% within 2 SDs' can fail badly. Also, probability for a continuous variable is area: the height of the curve at 170 cm is not the chance of being exactly 170 cm, which is effectively 0.
Think of it like this. Picture where a steady archer's arrows land, left or right of the aim point. Most land close, misses to either side are equally common, and wild misses are rare: a normal curve. A shakier archer makes the same shape, only wider, like a bigger SD. Gusts that only ever blow from one side would push some arrows far that way and skew the pattern.
Core idea. The normal distribution is the bell-shaped curve where values cluster at the mean and thin out evenly on both sides.