Working with AI: From Vague Intuition to Real Software
How I transitioned from hesitant experimentation to building full educational tools with Google Gemini using rapid conversational iteration.
It has been quite a while since my last post, but something genuinely transformative pulled me back to the keyboard: diving deep into building software with modern AI.
Over the past few months, working alongside LLMs has become the most compelling workflow shift I have encountered in my career as both an educator and a tools builder. Even though I haven't published every small experiment online yet, I have been deeply engaged in engineering new systems behind the scenes.
Finding the Right Partner: Why Google Gemini Stands Out
I spent considerable time experimenting with different foundational models across various tasks—from mathematical reasoning to frontend scaffolding. So far, my go-to choice has consistently been Google Gemini.
Its combination of large context awareness, fast token turnaround, and nuanced understanding of multi-step instructions makes it exceptionally capable when translating conceptual physics ideas into working code.
The "Vague Prompt to Precision" Method
When I sit down to build something new, I rarely start with a 20-page technical specification. More often than not, I begin with an intuition—a mental sketch of an educational mechanic or a testing workflow.
My opening prompt is often surprisingly raw and conversational:
"Here is what I'm trying to achieve conceptually. Make something like this, keeping performance lean and zero-dependency."
From that initial prototype, the real craft begins. I treat the model like an eager junior engineer sitting beside me:
- Deconstruct the Initial Scaffold: Examine the initial layout and isolate architectural bottlenecks.
- Iterative Refinement: Sharpen the math models, refine the DOM structure, and eliminate unnecessary abstraction layers.
- Stress Testing: Push edge cases—such as responsive breakpoints or LaTeX equation rendering.
Step by step, a loose concept crystallizes into high-performance, functional code that I can deploy straight to production.
What's Next
As I continue building educational software, canvas simulations, and the testing platform at SMA Physics, I'll be sharing the complete behind-the-scenes engineering process right here.
I will be posting real-world build logs, prompts that worked (and those that failed miserably), and the actual code produced during these sessions. Stay tuned—there is a lot of exciting engineering ahead.