When you submit a round in GolfStack, an AI coaching story appears within seconds. Three paragraphs in the voice of Harvey Penick, specific to how you played, never the same twice.
Here is what is actually happening behind that — what we send to Anthropic Claude, how the prompt is structured, and why this approach changes behaviour more reliably than a stat dashboard.
What gets sent
After you submit a round, the GolfStack backend assembles a structured payload and sends it to the Claude API. The payload is small — a few KB — and contains:
- Your hole-by-hole scorecard (par, score, putts, fairway hit, mental responses)
- Course metadata (name, tee selection, total par, total yardage)
- Weather at tee time (temperature, wind direction, wind speed, condition)
- Your pre-round intention if you wrote one
- Your post-round reflection if you wrote one
- A short summary of your last 3 rounds for trend context
No identifying information is sent. No images. No location beyond the course name.
The Harvey Penick voice profile
The system prompt to Claude includes a voice profile we have iterated on for months. The key elements:
- Warm, grandfatherly tone. Never clinical. Never robotic.
- Pattern-focused, not stat-focused. Names what happened, not what the percentages were.
- Concrete and specific. Refers to actual holes, actual shots, actual moments — never "you should focus on putting."
- Forgiving of bad rounds. Penick was famous for finding the one good thing in any round. The AI does the same.
The result is a story, not a report. You read it once, you put your phone down, and you carry one or two things from it onto the next round.
Why a story beats a dashboard
This is the part that took the longest to figure out. Dashboards do not change behaviour. They show you a number that goes up or down, and you forget about it before the next round.
A story — even an AI-generated one — does something dashboards cannot. It names what happened in a way you can repeat to yourself. "The 5-through-9 stretch was the rhythm I want to remember." That sentence travels with you. The "consistency: 67" number does not.
This is not a new insight. Coaches have known it for a century. Penick's Little Red Book is the bestselling sports instruction book of all time precisely because it is stories, not mechanics.
The new thing is that AI models like Claude are now good enough to do the storytelling reliably across millions of rounds, in a consistent voice, for users we have never met.
Why Claude specifically
We tried multiple models. Claude Opus produces the most consistent voice and the most useful pattern observations on golf-specific data. It also has the best behaviour around uncertainty — when there is not enough data to draw a conclusion, it says so, rather than confidently making one up.
For Scorecard OCR (a different feature, also powered by AI), we use Claude Sonnet's vision-capable variant. Same model family. Different model size matched to the task.
What it is not
The AI coaching story is not a swing diagnosis. It cannot see your swing. It cannot watch you putt. It cannot fit your equipment. Anything technical — grip, posture, plane — needs a PGA professional in person.
What the AI can do, reliably, is read your round and tell you what it noticed. Patterns. Tendencies. The one good stretch worth remembering. The one decision worth questioning. That is enough for most amateurs to improve, round over round, without ever taking a lesson.
Log a round and read your first one.