Gave Gemini 3 a vibe-coded login app and asked how I could improve it.
The response was thorough: "This is a solid, functional foundation. It has all the necessary elements and clear labeling. However, it currently feels a bit dated and utilitarian. To elevate it to a modern, professional, and trustworthy standard, we need to work on visual hierarchy, use of white space, and modernizing the UI components."
Then it broke down specific improvements, ranging from quick fixes to larger design shifts. It provided two alternative layouts showing different approaches to spacing and hierarchy.
The interesting part: I used a tool to screencast my iPhone to my laptop, showed the mockups visually to Claude Code, and told it to update the app to match the new designs. Within minutes, I had a modern version that Gemini had conceptually guided me toward.
What struck me was how the model personalities differ. Claude is thorough and careful. Gemini is creative and sometimes surprising. GPT is confident — sometimes right, sometimes wrong. Learning to match models to tasks, like choosing the right tool from a toolbox, is becoming its own skill. And the toolbox keeps getting upgraded every few weeks.
The whole cycle — concept to design critique to implementation to polished UI — took a fraction of the time it would have taken me to reason through all those design decisions alone.
Part of the #100DaysToOffload documenting agentic development in 2026