I used Nano Banana 2 to make perfect sketchnotes: 5 lessons learned (and how I broke Gemini)

Editor's take

A writer found that by carefully prompting and iterating, they could achieve impressive results with the Nano Banana 2 model for creating sketchnotes, even to the point of causing unexpected behavior in Google's Gemini.

This demonstrates the ongoing challenge of fine-tuning LLM outputs for specific creative tasks. While Nano Banana 2 showed promise, its susceptibility to certain prompt structures highlights the critical role of user skill in eliciting desired outcomes from even advanced models, and the potential for unexpected interactions between different AI systems.

Future developments will likely focus on improving model robustness against adversarial prompting and developing more intuitive interfaces that abstract away some of the complexities of prompt engineering. It will be telling if future iterations of Gemini or other models offer more resistance to such prompt-induced failures.