Making magic, AI tools that amaze!

I always start my articles by thinking "What do I know that most don't but would be useful and interesting to them?" This story I think qualifies.

I was helping a friend recently - a top-class performance marketing specialist - use AI to run automated platform data analysis. He'd gotten a long way on his own. Had some successes. But also a lot of frustration with things he just didn't know were limitations.

He didn't know that a custom GPT will work well initially but will eventually run out of context. That custom instructions max out at 8,000 characters - enough for a one-off tool demo, sure, but nowhere near enough for a real product.

He didn't know that if you use the API in a simple agent, there's no shared memory unless you build it yourself. So simple agents often disappoint. To get them to work well, you need to use a framework like LangGraph or build out the infrastructure yourself.

He didn't know that you can use Claude Code or similar tools to do much more powerful data analysis than you can in a custom GPT (or Gem if you use Gemini). It gives you proper use of Python scripts, can use your computer's files as memory to save intermediate states, machine learning can be invoked. Custom templates and outputs designed.

We got there though. And here's the thing - his world-class knowledge of performance marketing meant that the resulting tool is unbelievably powerful.

Now imagine trying the same trick with an AI expert who didn't know marketing. They'd solve for all the AI challenges. Build a beautiful system with proper memory management and context handling. But they'd have no clue how to actually get world-class insights out of the data. They wouldn't know which metrics matter. Which combinations reveal hidden patterns. What questions the data should answer.

The magic happens when you combine the skills and experience.

The AI literacy you need isn't "how to prompt" anymore. It's understanding the scaffolding. Knowing when you've hit a tool's limits. Recognizing that context windows matter, that memory needs to be architected, that the right tool for a quick demo isn't the right tool for production.

But all of that is just plumbing. The real value - the stuff that makes a tool go from "neat" to "I can't work without this" - that comes from deep domain knowledge. From knowing your field so intimately that you can design the right analysis, ask the right questions, and spot the insights that matter.

You don't need to become an AI engineer. But you do need to understand enough about the constraints to avoid the common pitfalls. And then you need to be world-class at your actual job.

That's where the magic is.