A dad built a working AI skating coach for his daughter on the third try

Vision My goal was to give my daughter Hanna a personal AI coach in her pocket that complements her figure skating training — so she can always analyze her technique, track her progress step by step, and take concrete steps toward becoming an elite skater. The product's own tagline sums up the promise: "Measured technique development, not guesswork." This is the story of how that vision became a launched app: two attempts that didn't hold up, and one insight about what actually makes the difference in vibe-coding. Achievements - Built a working AI coach persona in Claude with a structured analysis protocol for jumps, spins, and step sequences, calibrated against the IJS scoring system's GOE and PCS - Packaged the protocol as a reusable skill — tested and verified in Claude - Produced a full PRD spec following the course's product-development framework - Designed a clickable mockup with 10 screens — 7 built, 3 planned - Wrote and designed a full product landing page: problem framing, mechanism, positioning, and a waitlist signup - Launched the first part of the app as a working feature, not a prototype, deployed on Vercel: video is processed frame by frame, motion is measured via pose estimation, and AI interprets the result into a GOE estimate with strengths and a confidence score - Produced a short concept film Tools - Claude (Projects) — defining the coach persona: twenty years of experience, focused specifically on developing Hanna - figure-skating-coach skill — the reusable Claude skill packaging the analysis protocol, built and verified before any app-building attempt - Manus — first app attempt, before the course's app-building modules - Claude Code — second app attempt, after the Vibe Coding course module - PRD-spec skill — Mindvalley's framework for requirements-gathering, developed further by course leaders Vykintas and Vishen - Claude — built the mockup and the product landing page, then the third and final app attempt — the version that launched - Vercel — hosting for the deployed app - Gemini — generated the concept film - Computer vision / pose estimation — measures rotation, flight time, and landing angle from video Methodology I didn't start with code. I started by defining who Claude would be: a coach with twenty years of experience, with a single mandate — developing Hanna. From there we iterated step by step until the focus landed on video analysis. I packaged that into a skill. I tested it in Claude, and it worked immediately. The next step was turning the skill into an app. The first attempt, in Manus, came before I'd taken the course's app-building modules — and it showed. The second attempt, in Claude Code after the Vibe Coding course, was better but still not good enough. It's easy to get started with vibe-coding. It's harder to make the result hold up. The difference came when I got access to the PRD-spec skill in the course's final modules — a framework for structuring requirements before you start building. I followed the framework, built the mockup, and wrote the landing page to pressure-test the positioning. Only then did I write the app a third time, in Claude, and deployed it on Vercel. That's the version that launched. It wasn't the tool that made the difference between the three attempts. It was the process. Learnings - Vibe-coding is accessible — but accessible isn't the same as easy. Two attempts without a structured spec produced two insufficient results, despite the same tools and the same goal. - Structured requirements-gathering (PRD spec) before code was what made the difference — not the switch from Manus to Claude Code. - The motivation was never the technology itself. It was giving Hanna something concrete: a tool she can actually use at every training session, built by her dad, for a sport she's serious about. That's what kept me going through two discarded attempts before the third one worked. - Writing the landing page forced a sharper answer to "why does this need to exist": most skaters get minutes, not hours, of personal coach feedback per week, and generic AI tools can't reliably count rotations or measure landing angles. Thanks Thank you to Mindvalley — and especially to Vishen and Vykintas — for building the AI Mastery course that made this possible. The PRD-spec framework from the course's final modules was the single biggest unlock in this entire project.