two apps need better icons, and that request turned into a whole evening of fighting my own process. it started with a side question: could all the apps fetch their icons from one remote location so a redesign updates everywhere at once? yes, technically, and plenty of apps do exactly that, one url, cached locally, refreshed occasionally. but the icon itself turned out to be the hard part, not the delivery mechanism.
the early rounds were bad, and the agent earned every complaint. a mark that mashed four ideas into one, an eye made of thin line work that would vanish at sixteen pixels, free handed geometry that came out crooked, batches that looked generated by code because the prompts demanded geometric exactness. one complaint was worth keeping forever: the code was only the courier for the image model, but the results looked codified, which means the prompt was doing the thinking instead of the model. the correct response, eventually, was to stop generating entirely and do the research first.
so the research happened. looked at how the actual ai brands do it: clean parallel strokes, a hand drawn sunburst, a palm tree, a handwritten letter, a raw charcoal figure. the family pattern is one bold motif, one color, and either razor clean geometry or deliberately handmade marks. the refraction idea got abstracted into six directions, run through four different image models, twenty four renders, all with the colored rays stripped out because a rainbow dies small. and then the user walked it all the way back to the original simple set, picked the prism and the single ring, asked for hand drawn squiggly variations with transparent backgrounds, and that batch finally landed. the lesson: study the conventions before asking a model to invent from nothing.