The distance between having an idea and being able to show it, built with Kpler's actual components, was too long. Figma gets you a picture of the idea. It doesn't get you the idea running in the real environment, with real data shapes and real constraints, which is usually where an idea turns out to be wrong.
I built Chef, an internal AI agent that lets designers and PMs prototype directly against Atlas, our real design system, instead of approximating it. I used it on my own work first. It took about three weeks from first build to getting visibility up to CTO level.
Once it held up, I didn't just hand it off — I ran training sessions and sat with people while they used it. People built real pieces of product with it, including a color switcher for our graphs: test colors, build sets, apply them. I also walked people through DialKit, a tweaker tool built by Josh Puckett, when that workflow needed to click. Some of those prototypes grew into their own tools, including a more ambitious dataviz app that went well beyond what Chef started.
It ended up integrated into the web app repo, and the pattern I'd built it with gave engineering a blueprint for other agents afterward — including how to handle accessibility and PR checks. It got presented at Chapter Front, our recurring meeting for everyone who touches the front end: design, PM, front-end, back-end, full-stack. It was also proposed for the Athens offsite, which ended up canceled — not because of Chef, but because of the situation at the Strait of Hormuz. Kpler has an office in Dubai, close enough to that conflict that the offsite got pulled. Around the same time, a lot of people came to Kpler specifically for information on what that situation meant for world trade, which was a strange, oversized moment to have happening in the background of a design tool rollout.
It moved from a personal tool to something with real organizational visibility, and it gave other people a working starting point instead of a blank file.
Two things didn't work, and both were more honest than any of the wins. A handful of users were on Windows. The technical ones figured out workarounds; the rest found it enough of a headache that they just dropped it. And some people stopped using it after the initial excitement, and when I looked into why, it wasn't the tool — it was that I hadn't taught file management well enough. People weren't sure where their work lived or how to start a new idea from scratch, and that friction was enough to quietly kill adoption for them.
Building the tool was the easy part. Getting people to change a daily habit, on a normal day, under normal deadline pressure, is a slower and much less satisfying problem — and it's mostly a teaching problem, not an engineering one.