What a spreadsheet taught me about building with AI

If you're a leader pushing your team to use AI coding tools, go build a spreadsheet with Claude first.

I've spent thousands of hours building financial models, and I've been using Claude in Excel heavily over the past few months. What I learned surprised me and changed how I think about AI-assisted building entirely.

Claude was genuinely remarkable at getting started. It helped me think through model architecture, generated complex formulas fast, and got the infrastructure of a multi-sheet financial model up and running in a fraction of the time it would have taken me. It felt like a superpower. 🦸

Then the brittleness started showing up.

Hardcoded cell ranges instead of named ranges. A change in one place quietly breaking something three tabs away. It worked. But it was fragile, and as requirements changed, I was carrying real technical debt in a spreadsheet.

That's when I understood something. Without explicit direction toward best practices, Claude builds something that works but may have the same problems. Guiding it in the right direction takes real effort. And knowing which direction is right requires having built these things from scratch before. It requires seeing around corners You need to already understand named ranges to tell Claude to use them. You need to already know what a scalable model looks like to ask for one.

That realization hit harder when I zoomed out.

I've probably logged 500+ hours building with Claude and Codex. Side projects, tools, things I'm genuinely excited about. Many work great! But I rarely look at the code. I don't understand it. I can't evaluate whether what's being produced is solid or brittle. Claude in Excel gave me the first real glimpse of what might be happening under the hood, because spreadsheets are a language I actually speak. If Claude cuts corners there without explicit direction, it's may be doing the same thing in code I don't understand.

For leaders pushing their teams to adopt these tools: this isn't a reason to slow down. It's a reason to build real respect for the expertise required to use them well. The gap between "it works" and "it's built to last" is where your engineers live every day. It takes someone who already knows how to build well to close it.

One note: Claude in Excel likely isn't receiving the same investment as Claude Code, so the comparison isn't perfectly apples to apples. But the principle holds.

Go build a spreadsheet with Claude. You'll feel it.