Current image models are terrible at this. When DALL-E 3 comes out, people will have 1 week to find a prompt that works. (The timer starts when I get access, or when someone else gets access who's willing to test out prompts from anyone in the market.)

The prompt should be some string of text such that the name of a polygon can be inserted, and at least 50% of the generated images are the correct shape.

It must succeed on all polygons from the triangle to the octagon. It's ok if it fails on more esoteric ones.