Why I did AI Prompting Wrong (And What Actually Works)

The moment I realized I’d been solving the wrong problem

Image generated by author using ChatGPT

Picture this: you’re trying to optimize the perfect sentence to get ChatGPT to write your marketing copy. You spend 45 minutes tweaking words, testing different approaches, burning through API calls. Sound familiar?

I just spent two months building an “academic” framework to solve exactly that problem. I called it CERT (Combinatorial Enhancement and Recursive Transformation) — a sophisticated system to scientifically optimize prompts for maximum performance.

The results? Complete failure.

But here’s the thing about failures — sometimes they teach you more than successes ever could.

The Academic Rabbit Hole

I started with what seemed like a solid approach. I had this intuition that prompt optimization was like exploring a systematic transformation search.

I formalized it mathematically. Defined transformation sets, created enhancement matrices, and built algorithms to detect synergistic combinations. On paper, it was elegant.

I brought geometric insights (I’m a big fan of representing complex problems with simple geometric structures), thinking about intention spaces…