I have a side project. It's called Project Sub Three, and it does one thing: help runners break a three-hour marathon.
It's a weekend project. It's also decades in the making.
Twelve attempts
My first marathon started at a blazing sub-three pace. Until mile 18. The wall. I finished in four hours, the hardest way possible.
I tried eleven more times. All failures.
Here's the strange part: during those years I was at Adidas building miCoach, working alongside some of the top running coaches in the world. Later Nike, later Bowflex. I was literally building the coaching tools while failing to coach myself.
On attempt #12, after a year of focused, honest training, I broke through. 2:58.
The wall every training tool hits
Every training-plan product I worked on hit the same wall, and it wasn't the science. The science is solid — Jack Daniels' VDOT system has been getting runners to the line for decades.
The wall was the inputs.
Heuristic algorithms need clean data. Runners don't have clean data. They have skipped days, self-reported paces, a lingering calf, a marathon PR from ten years ago with no context, and a long tennis match the night before a long-run day.
A good coach handles all of that with one question: "How are you feeling this week?" An algorithm can't ask it — and can't use the answer.
Why now
Modern language models change the math. They accept messy, partial, natural-language input and produce useful output. They can hold context. They can follow real training philosophy instead of flattening it into rules.
A prototype that would have taken a team of five and a research budget a decade ago is now something one product person can build in evenings.
So I built it. Not a chatbot — a companion. It knows your training history, adapts when life happens, and takes action: log a run, rebuild a taper week, recalculate your probability of breaking three. The plan bends. The goal doesn't.
How it works
The architecture came out of getting it wrong first. Version one asked the AI to do everything — conversation, coaching, and generating a full 56-week plan in one streaming response. It fell apart at scale.
The fix: AI decides, tools execute. Claude handles the conversation and the coaching decisions. A deterministic plan engine — plain TypeScript — builds and rebuilds the entire plan in milliseconds via tool use. Change one thing, and the ripple effect shows up everywhere: probability, injury risk, phase transitions.
Underneath it all sits a single runner context — a structured-but-forgiving record of who you are as a runner right now, read and written by every conversation.
The stack is a real product, not a demo: Next.js on Vercel, Claude doing the reasoning and tool orchestration, Supabase for accounts and persistence. That's the other half of why this project exists — testing how far I can take a full AI-native build solo: auth, backend, APIs, context, tools, skills.
The full breakdowns live on the site: the method and the tech.
Early days
It's early. It's rough in places. That's the point — it's a living prototype, and every conversation with a real runner teaches it (and me) something.
If it helps even one runner get to 2:59:59 in fewer than twelve attempts, it's worth every hour spent building it.