Over the summer I spent twelve weeks building out the most incredible app for my best pals. It had all the bells and whistles I could dream up in late-night sessions in Claude and Claude Code. When it was finally ready for a beta rollout, I was so excited to show the guys.
It was one of the most complicated things I have shipped. (If you want to know about the stack, I built a site showing what’s under the hood.)
This essay is not about the perils of building, it’s about the perils of overbuilding.
When my pals saw the fancy app I slaved over for them, it became clear pretty fast they wanted… no… part… of… it.
It was my first lesson in Rule #1 in the AI Building era we are in now: Build what they need and want, not what you think they need and want. Just because you can build it, doesn’t necessarily mean you should.
So here’s the deal: Me and my three best guy friends have known each other for almost 30 years. There is barely a day that passes that we don’t message each other in a snarky group chat that has become one long running conversation. We are such good pals we can rip each other, constantly poke fun at each other, to the point that if it was anybody else you would block them. But with these guys, we all shrug our shoulders and smile, and think of a good comeback.
Ten years ago one of the guys created what is known as a “confidence pool” we play each week, for like twenty weeks, during football season. It’s a fun thing. Bragging rights only. No money changes hands.
The pool has evolved over the years, but the one guy who runs it kept score by hand and entered it into a spreadsheet on his laptop each week, then messaged us updates in a group chat throughout the week. This is how it’s been for a decade. It worked, yes.
But then AI came along.
I decided to take a decade’s worth of spreadsheets and build a private app in Claude that would keep score on its own each week, generate a snarky, AI-driven newsletter each week for the guys, automated emails that would fire off a couple times a week with scoring updates, a bespoke website and domain name… and that is just for starters.
When the guys finally got their hands on it, it was met with a thud.
All three of them are AI skeptics, some more than others. So there was natural resistance already baked in.
One pal didn’t want any of his personal info fed into AI, and I understand and have to respect that, so I had to quickly dial back a part of the app that would make it the most personal.
Another pal, I found out later, hates emails with a passion. And he is no fan of AI. So an AI-generated automatic email was like a right jab and a left uppercut every time he opened his inbox. He asked (actually demanded!) to be removed from the emails. So I had to build out a workaround to mute his emails.
Anytime there was a minor glitch during the beta rollout phase, the guys would pounce like a lineman on a fumble. They saw any glitches or bugs as evidence AI was not up to the job and would fail.
I won’t bore you with all the play-by-play, but suffice to say I had to unwind and unplug about 50% of what I built in the app just to prevent a revolt and keep my friendships.
At one point, when I started asking them what exactly they wanted, one of them said all that was needed was an app that would add up scores in a spreadsheet each week! I was thinking to myself, “If that’s all you wanted, I could have one-shotted that in Code in like an hour!”
The first couple weeks of the app beta rollout were rough. The guys didn’t particularly like it. Some of their concerns were legit, others were (in my opinion) just guys used to routine, stubborn and anti-AI. I was doing my best to A) Not take personal offense to their disinterest in my beloved app I built for them and B) Unwind everything they didn’t like.
We are now in Week 10 of the season. The app runs itself. I don’t have to touch it. It’s working. Most glitches were smoothed out weeks ago. The guys mostly are happy with the scaled-back version of the app. Maybe happy isn’t the right word. Maybe “accepting” is better.
And they don’t make fun of mistakes the app makes anymore because, well, to be honest, there aren’t many mistakes. It’s humming along. After Week 1, the guy who kept the weekly spreadsheets admitted the self-updating, AI-powered spreadsheet was just as good if not better.
Four lessons
I learned several lessons from this build. The technical kind, yes, but more importantly the human kind.
People who know and love you so much they are not afraid to tell you “this is stupid.” It’s honest, and that sort of bluntness is often needed when building something in AI. Often we get so deep in the weeds in a model and in code, we need to see the forest for the trees, as the old saying goes.
Not the nonessential frills of what you think they want.
Some people are ambivalent, and some are downright hostile to it. Know your audience, respect their views, and build with that in mind.
Find out the pain points first, before you make the first pull request in Code. I was deep into the rollout of this app before I learned the one pal hated email. Also, I built the app to be mobile-first, assuming all the guys would look at it on their phones. Wrong. Two of the guys actually prefer to look at it on desktop. Terrible assumptions. Cost me a lot of time.
I am now working on a different project, building AI tools that will help a relative with his business. The first thing I did was interview him to find out exactly how he works, what he wants, and what pain points I can help him solve. Do this on the front end, not on the back end.
This was a low-stakes project, a football pool among friends. But the lessons aren’t small. Swap the pool for a newsroom or any large organization and it plays out the same way: people wary of AI, colleagues who don’t want one more automated email, and a single glitch taken as proof the tool doesn’t work. Getting people to actually use what you build matters more than building it.