What's a peak?
 

When people say they set their peaks at -12 or -20 dbu, what's considered a peak? A yell? A gun shot? Should the meter be hitting the peak level several times a sentence, or hardly ever?

It depends on what you're recording and at what stage of the production process you're at. Doing an original recording of dialogue from a person with a steady delivery, then the peaks would occur several times per sentence. If their levels are all over the place, you'll have to give them some more leeway.
With louder sounds you not only have to watch the meters but also listen for overloading the input. Some loud sounds that are very short duration may also hit digital zero without the meters showing that high, even though they are peak meters.
It takes some experience, familiarity with both your gear and the subject. That's why it's a guide and a flexible decision on your part. After the original recording is done, then you can tweak, adjust, compress, mix etc. If I'm the only person using the material, then I'll likely mix hotter. If the tape must go to someone else, especially for broadcast, then you must work more strictly during the final stages of mastering the tape.
Unless your equipment is excellent and your subject has huge volume swings, -20 is really low for an original recording and will probably not be helpful for ordinary work.
-12 is safe for moderate volume swings. If the subject is well controlled, I will often record with peaks at -6, although obviously that leaves less room for error.