Every golfer knows WITB. What's In The Bag. It's the little ritual we do with each other and with the pros: the 14 clubs, the shaft flex, the grind on the wedges, the ball. We geek out over the setup because the setup is where the round is won or lost before you ever tee it up.
So let me do the golf-nerd thing, but for the part of HoleRecord you never see. Not what's in the bag — what's in the backend. The stuff that turns a spoken sentence — "pushed my drive right, punched out with an 8-iron, hit the green with a 7, two-putted for bogey" — into a filled-in scorecard. And, more honestly, the stuff that broke when we finally stress-tested it, and what fixing it taught me.
We built a driving range for the AI
You don't groove a swing by hitting three balls and calling it good. You go to the range, you hit a bucket, you watch the ball flight, and you adjust.
So we built a driving range for the AI. Five rounds. Ninety holes. And here's the important part: every one of those holes came with a known-correct scorecard — a real answer key. The exact score, the putts, the fairway, the green, the penalties, the clubs. Then we fed the app the spoken version of each hole and graded what came back against the truth.
That's the whole discipline, and I want to be plain about why it matters: without an answer key, "the voice thing feels pretty good" is just a vibe. With one, every miss has a name. You can see exactly which hole broke, exactly which stat it got wrong, and exactly whether your latest fix helped or quietly made something else worse. It's the same reason we built a golf-specific eval harness for the transcription layer a while back — you can't improve what you refuse to measure, and you can't trust a measurement you haven't checked.
The biggest surprise: it wasn't the microphone
Going in, I assumed the failures would be about hearing. Golf is a linguistic minefield — "putt" turns into "pot," "OB" turns into "obi," "birdie" comes out "birdy." I figured the course, with its wind and its background chatter, would be where it all fell apart.
It wasn't. Even on audio recorded outside, near noise, the speech-to-text held up. The transcript was fine. The words made it through.
The misses were about understanding golf, not hearing it. The model wasn't listening wrong — it was thinking wrong. It needed to think more like a golfer, not listen harder. That reframed the entire project for me. This wasn't an audio problem. It was a golf-knowledge problem. And golf, it turns out, is full of little rules that are obvious to us and completely non-obvious to a machine.
Here's what the range session actually surfaced.
Up-and-down isn't just "you got it on and putted"
The model was handing out up-and-downs like a generous starter handing out mulligans. It counted a two-putt after a chip as an up-and-down.
Any golfer knows that's wrong. An up-and-down means you missed the green and still made your par — you got "up" onto the green and "down" in the hole in one putt. Chip on, two-putt? That's not a save. That's a bogey with extra steps.
So we taught it the actual rule. You have to get down in a single putt from off the green for it to count. Small thing. But if you're a golfer trying to figure out whether your short game is actually saving you strokes, an inflated up-and-down number is worse than no number — it's a lie that feels like progress.
"Two-putt" vs. "to putt" — a homophone that cost people strokes
This one's my favorite, because it's so golf.
The model kept hearing "two putt" as "to putt" — and logging a single putt. Say "I had to putt from off the fringe" and "I two-putted from the fringe" out loud. Your ear knows the difference from context. A transcription model, mid-sentence, doesn't always. And because putts feed into the score, a couple of these didn't just miss a stat — they threw off the number on the card.
We taught it to disambiguate from the surrounding words. And we learned something we now quietly pass along: saying "two-putted" — past tense, as a verb — is clearer than "two putt." The machine hears it better, and honestly, so would your buddy.
Off the tee and the approach — capture both
Your scorecard has room for two clubs on a lot of holes: the club off the tee, and the club into the green. When you say "driver off the tee, then a 7-iron in," you're handing over two useful facts.
The model was only writing down one of them. Now it captures both — the driver and the 7-iron — so when you look back and wonder why you keep missing greens from 150, the data's actually there to tell you.
Grading against your goal, not against par
The last fix isn't about extraction — it's about respect.
The app grades your round. But golf grading against par is brutal and, for most of us, meaningless. If your goal is to break 100, shooting 94 at a hard track is a genuinely great day — and the app now knows that. It grades you against your goal, not against a scratch-golfer standard you never signed up for. Break 100 and shoot 94, and that's an A. It'll still point out, in the coaching notes, that your three three-putts are the first thing to sharpen. Encouraging and honest aren't opposites. The best playing partners are both.
About that "essentially perfect" number
We shipped a real win earlier — upgrading the transcription engine from Whisper's original model to gpt-4o-transcribe. On a clean-audio baseline, that took golf-term accuracy to essentially perfect: every golf word survived the trip from mouth to transcript.
I want to be careful with that word, perfect. Those were clean conditions — good mic, no wind, clear diction. Deliberately so; we wanted to measure vocabulary handling on its own. Real rounds are messier: wind on the mic, a foursome laughing, a cart rattling by, a dozen accents, whatever phone's in your pocket. Messy on-course audio is the next frontier, and I'll say the quiet part out loud — I distrust our own measuring stick about as much as I distrust the model. A great score on the wrong test is just a comfortable lie. We'd rather find the ugly numbers ourselves, on the range, than have you find them on 14.
The point of all of it
That's the WITB. Not carbon shafts and tour-grind wedges — an answer key, a hard rule about up-and-downs, a homophone we finally pinned down, a second club on the card, and a grade that meets you where your game actually is.
None of it shows up when you tap the gold button and say how the hole went. That's the whole idea. The tech should disappear so the round can stay. Because HoleRecord was never about capturing a number — it's about keeping the whole record: the story, hole by hole, told the way you'd tell a friend in the parking lot. Our job is just to make sure that when you tell it, we get it right.