Somewhere between watching my first college gymnastics meet and my fifteenth, I started actually caring about the scoring. Not in a frustrated fan way, more like a genuine curiosity about how you even judge something like that. What makes one score different from another? How does anyone agree?
The more I watched, the more I wanted to understand it from the inside rather than just observe it from the couch. That ended up being the seed of a project — an app that lets you score each gymnast in real time and compare your calls to the official judges at the end of the meet. I'd also been wanting to find something real to test how far I could push AI as a collaborator, and this felt like the right excuse. I worked with Claude throughout, and it turned into one of the more interesting things I've worked on.
The first useful conversation wasn't about features. It was about what to leave out. Every time I tried to add complexity early, it got in the way. The version that actually got built was the one where we agreed to keep things simple and focus on the core experience first. That constraint turned out to be a design decision as much as anything else.
From there I learned pretty quickly that jumping straight into building was the wrong instinct. For anything with real design complexity, describing what I wanted and expecting a solution backfired every time. What actually worked was treating bigger features as conversation problems before they became building problems. I'd ask for reactions before asking for solutions, give specific notes, look at things a few different ways. Only after a few rounds of that did anything get made, and it went faster and cleaner because of it. Every time I skipped that step I ended up in the same conversation anyway, just with more to undo.
A lot of the work itself isn't exciting to describe. Bugs that came down to one wrong assumption. Errors that took several rounds to untangle. Moments where something looked right but felt off and it took a while to figure out why. But the loop of fix, check, fix again built something I didn't expect: a better instinct for where things go wrong. I started forming hypotheses before describing the problem out loud. That felt like progress separate from the app itself.
The question I started with, how far can I push this, gave way to a different one pretty quickly. What does this actually require from me? The calibration took time. Too specific and you get exactly what you said but not what you meant. Too open and you get something plausible that misses the point. Finding that middle is genuinely a skill. I also had to learn to say things I'd never say to a person. "Don't change this." "Assume I haven't considered that." With a human collaborator a lot goes unsaid. With Claude you have to surface assumptions that would normally just be understood. It felt blunt at first. Skipping it cost me time.
The other thing worth mentioning is knowing when to reset. The longer a project runs, the more old decisions accumulate and create drag. There were moments where starting fresh moved faster than trying to patch what was already there. Recognizing that point turned out to be its own kind of skill.
What surprised me most was realizing Claude isn't a replacement for what I do. It fits into how I work. And it rewards a specific kind of engagement. Show up without a point of view and you get output that's convincing but slightly off in ways that are hard to name. Show up with taste, with opinions, with the willingness to push back, and it's genuinely fast and genuinely good. It didn't change what good work requires. It just made the gap more obvious when I wasn't bringing it.
Thirty days in, I'm proud of what came together. My best guess is this would have taken months without AI in the mix — being able to quickly brainstorm, build, and test an idea in real time changed the math entirely. If you're into gymnastics or just curious, check out the app at the link below.
