Comment by theptip

Did you eval using screenshots or some sort of rendered visualization instead of the CLI? I wonder if Claude has better visual intelligence when viewing images (lots of these in its training set) rather than ascii schematics (probably very few of these in the corpus).