Realizing I’ve Built An AI System

Yesterday I posted about how much I’ve been messing around with AI. It was more of a loose brain dump than a plan. I’ve just been stacking tools, prompts, and little workflows because they seemed useful in the moment.

Today I was talking to a colleague and walked them through some of what I’ve been doing. That’s when it clicked that I’ve basically built a system without calling it that. To me, it’s just “the way I’m doing things now.” To them, it was completely new.

I showed them how I break work into pieces, how I use different tools for different stages, and how I lean on AI to draft, sort, and refactor ideas. None of it is fancy. But they hadn’t considered doing things this way at all. Watching their reaction made me realize I’ve been taking for granted how far I’ve pushed my own process.

What worked:

• Thinking in terms of outcomes instead of tools: I didn’t start by asking “which AI tool is best.” I started by asking “what do I actually need help with?”—brainstorming, drafting, cleaning up, organizing, etc.
• Being specific with instructions: the more detailed I was about context, constraints, and examples, the better the outputs got. When I treated the AI like a vague search box, the results were useless.
• Chaining small steps: instead of trying to get a perfect final answer in one shot, I broke tasks into stages. Draft → refine → restructure → summarize. It made everything more reliable and easier to fix when something went sideways.

What didn’t work:

• Expecting AI to “just know”: early on, I kept assuming the tools would infer what I meant. They didn’t. I’d get generic content or weird tangents and end up rewriting everything myself.
• Copying other creators’ hype: I watched people online talk about “one prompt to do everything.” That never worked for me. Their prompts either didn’t fit my context or produced bloated, overconfident answers.
• Ignoring my own language: when I tried to sound like “AI people” instead of how I naturally think, my instructions got worse. The model followed my fake jargon and gave me fake-jargon answers.

The big realization from the conversation: the tools aren’t the magic; the way you talk to them is. There really are endless possibilities, but only if you’re clear about what you want and willing to treat it like designing a process, not pushing a magic button.

Would I do it again? Yes, but with more intention. I kind of backed into this system by trial and error. If I were starting fresh now, I’d still experiment a lot, but I’d write down the pieces sooner and admit that I’m actually building something—not just “trying AI.”