Today I’m working on an AI try-on system for the Duke clothing store. This is one of my first real “vibe coding” experiences, and progress has been fast. With clear direction, I can guide the system smoothly and avoid most bugs.
A key lesson: output quality is highly sensitive to input specificity. When prompts are vague, results degrade fast, with mismatched clothing or missing elements. Understanding model behavior and constraints (especially with Claude) makes a big difference in reliability.
On the generation side, even strong models like Nano Banana Pro are still inconsistent. I’m currently running prompt experiments to improve stability and visual accuracy.
Longer term, I see a shift toward personalized avatar models capturing attributes like skin tone, height, and body shape. That could reduce generation latency, lower costs, and create a defensible layer for integration with retailers. For now, starting with user-uploaded images is the most practical path.
Meeting with the Duke Store team today to evaluate next steps.
Prompt testing:
