Swing Analyzer
The Lie
Your swing, told straight. ⛳️ It reads your body andclub, finds the faults that actually repeat — then names the one thing to fix first.
How it's different
Most swing apps are slow-mo video players that throw angles on the screen and leave you to self-diagnose, one noisy swing at a time. 🎯 The Lie does the diagnosing:
- Reads your body and club— not just lines on a screen.
- Finds the faults that actually repeat across many swings, not one noisy rep.
- Hands you a ranked plan— root cause first, weighted by impact, effort, and regression risk.
- Gates each fix on the off-course mobility work, then a “lab” confirms the change actually stuck.
Product
A decision, not a dashboard
Every other tool hands you data. This one hands you a decision— “just tell me the one thing to work on that won't blow up my game.”
- Ranks fixes and sequences them root-cause first.
- Respects regression risk — no mid-season rebuilds.
- Gets sharper the more you feed it, building a model of you as a golfer.
- A validation lab closes the loop on what actually transferred.
Design
The board, read at a glance
It leans into Augusta without ripping it off — fairway green, champion gold, scorecard red — where every color earns its place.
- Gold is reserved for earned wins. 🟢 Restraint is the whole voice.
- Red/green always pair with a glyph — colorblind-safe.
- “The Board” renders a gnarly model as a Masters-style leaderboard you read at a glance.
- Every ranking opens to show why — never a black box. Installs like a native app.
Engineering
CV measures, Claude reasons
The split that keeps it honest: computer vision measures, Claude reasons — and never the other way around.
- MediaPipe + WebCodecs pull hard biomechanics (spine tilt, X-factor, kinematic sequence) in the browser.
- Claude Opus reasons over those numbers and key frames — citing metric keys only, so it can't hallucinate an angle.
- Prioritization runs over a cause→effect dependency graph.
- A Next.js PWA on Cloudflare (D1 + R2).
How I built it
In June 2026, days after Anthropic shipped “ultracode” and dynamic workflows, I built this with them. 📈 Instead of one model writing code top-to-bottom, a coordinator writes an orchestration program that fans out a fleet of specialized sub-agents:
- Agents build disjoint modules in parallel against a frozen interface contract.
- Independent agents adversarially reviewthe work — one writes it, a different one tries to refute it.
- An integration pass reconciles everything and verifies the build is green.
Deterministic parallel barriers, typed hand-offs, loops — more a CI pipeline of agents than a chatbot. Code review and QA baked into generation, not bolted on after.
By the numbers
Read your lie.
Step in and poke around — it's my personal instance, but the ideas show. Want yours read? Send me a down-the-line clip.
Step inside