Would also be interesting to see how well they can do with a loop of: write SVG, render SVG, feed SVG back to LLM for review, iterate. Sorta like how a human would actually compose an SVG of a pelican.