Agents Sucked At UI (Until Today)
By turning UI design into a "hill climb" problem—having multiple AI models repeatedly score and critique designs until all gave 9.5+/10—the author completely redesigned T3 Code's interface for $400, proving agents can handle subjective creative work when given the right feedback loops.
Read Original Summary used for search
TLDR
• Created a "council" of AI models (Claude Opus, GPT-4, Muse) that score UI designs 1-10, then iterate until all agree it's 9.5+/10
• Fed competitor screenshots (Claude Code, Codex) to models for comparative scoring—scores for all apps rose together as T3 Code improved, showing relative judgment works
• Spent $400 across 17.6 hours of parallel agent runtime to go from 6/10 to 10/10 in every design category
• Same hill-climbing technique reduced MCP context usage by 52-78% while maintaining perfect eval scores—turn any problem into a measurable score and agents will optimize it
• Key insight: don't prompt for perfection, set up feedback loops where agents measure and improve their own work iteratively
In Detail
The author tackled T3 Code's poor UI by treating design as a hill-climbing problem rather than a one-shot prompting challenge. Instead of asking Claude to "make good UI," he created a "council" system where multiple AI models (Claude Opus, GPT-4 Codex, Muse, and Gemini as a "jester") independently score designs 1-10 across categories like typography, hierarchy, and color. The agent iterates on the design until all models agree it's 9.5+/10. Crucially, he fed the council screenshots of competitor products (Claude Code and Codex) so models scored T3 Code relative to real alternatives, not in isolation.
The process ran for 17.6 hours across multiple threads, costing roughly $400. Initial scores were 6-7/10. After the first round of changes, scores jumped to 8-9/10. He then raised the bar to 9.5/10, which took 10 additional iterations. Finally, he pushed for 10/10 in every category, which the agents achieved by restoring color (earlier iterations had become too grayscale) and improving contrast. Interestingly, as T3 Code's scores improved, the models also scored Claude Code and Codex higher—suggesting the comparative context was calibrating their judgment.
The same technique works for objective problems. He used hill climbing to reduce T3 Code's MCP context usage by 52-78% while maintaining perfect scores on an eval of 85+ failure cases. The key insight: turn any problem into a measurable score (whether programmatic or vibes-based), and agents will optimize it. Instead of crafting perfect prompts, set up feedback loops where agents measure and improve their own work. This makes previously impractical tasks—like manually reviewing thousands of UI iterations—trivial.