AI Experiment 002: I Quit Asking AI to Draw Logos
My LinkedIn banner looked fine at a glance, but every icon was janky — AI drew the logos from memory. So I quit asking AI to draw logos and rebuilt the banner with official vector logos instead.
My LinkedIn banner shows my current stack: 20 tools in a tidy grid. It looked fine at a glance, but the icons were off. LinkedIn, Buffer, Slack, and Google Analytics all looked janky. The Loom logo was too small. Nothing was quite symmetrical.
The experiment
Rebuild the banner so every icon is crisp, official, and exactly the same size.
What I tried
First pass: AI generated the icons. That is the default move now. Describe what you want, get an image back. The layout was right and the vibe was right, but every logo was an AI approximation. Close enough to recognize, wrong enough to notice once you looked.
The Claude icon was the worst offender. The AI drew an asterisk with ten blobby petals. The real Claude asterisk has twelve sharp rays. It was not even close, and it sat in the first tile of the grid, right where everyone looks first.
What happened
I quit asking AI to draw logos. Instead I pulled the official vector logo for every tool, from Simple Icons, Wikimedia Commons, and the brands' own sites and CDNs, and composited them with code. Every tile exactly 72 pixels. Every mark normalized to the same visual size. Loom blown up to fill its entire tile. The whole banner built at double resolution, then shrunk down for clean edges.
Three rounds of picky feedback later, v43 was the keeper.
Key Takeaway
AI image models are composition engines, not precision instruments. They approximate. They will get you 90 percent of the way on layout and vibe, then quietly mangle the details that matter, like a logo. When precision matters, let AI write the code and feed it the real assets.
The rule I am keeping: never ask AI to draw a logo.
