Skip to content
Ryan James
Menu

← Newsletter3 min read

I Asked Claude to Solve a Cycling Problem. Here’s What I Actually Learned.

I Asked Claude to Solve a Cycling Problem. Here’s What I Actually Learned.
I don’t have a power meter on my bike. If you’re into cycling, you know that’s a gap — wattage is one of the best indicators of effort and fitness, and the hardware to measure it directly costs hundreds of dollars.

So I did what any curious CS teacher would do: I decided to build my own solution using AI. Specifically, I gave Claude a real engineering challenge and let it write the code while I played architect. What followed was one of the more instructive AI projects I’ve done — not because everything worked, but because of how it worked and where it broke.

The Idea

The goal was a web app where I could drop a GPS file or link a Strava activity, plug in my weight and bike specs, and get back:
• Average and max power output (watts)
• Total elevation gain and distance
• A dynamic power-over-time graph to visualize effort across a ride

Several websites already do this kind of calculation manually. I wanted something automated — something that would just run the moment I finished a ride.

The Stack

I used Claude Code to build out the architecture:
• Frontend hosted on GitHub Pages
• Backend running on Vercel serverless functions
• Database using a free Supabase SQL table for storing ride data
• Automation via a Strava Webhook that fires the instant a ride syncs from my GPS watch

The end goal was zero-touch: finish a ride, upload it, and have the app calculate my power metrics and write them automatically into the Strava activity description. No manual entry. No copy-paste.

The Part Nobody Talks About: Biprompting

Here’s a term worth adding to your vocabulary: biprompting — the practice of using iterative, context-rich prompts to guide an AI through a complex, multi-step problem rather than expecting a single prompt to produce a finished result.

It’s the difference between saying “build me a cycling analytics app” and actually sitting with the AI through each layer of the build — feeding it error logs, asking it to explain what broke, and refining the approach one system at a time.

The frontend came together quickly. The backend was a different story. At various points I was debugging three separate services simultaneously — Supabase, Vercel, and the Strava API — and no amount of clever prompting would substitute for understanding which one was failing and why. I had to read status codes. I had to pull logs. I had to know what a 200 versus a 400 response actually meant before I could give the AI useful context to work with.

What This Means for the CS Classroom

This is the part I keep thinking about in terms of what we teach students.
The temptation with AI coding tools is to treat them like magic — type a wish, receive a program. And for small, isolated tasks, that impression holds up. But the moment you’re building something with moving parts, you hit the wall fast. The AI can write excellent code. It cannot debug a system it can’t observe. That’s still your job.
What I needed to successfully build this wasn’t advanced programming skill — it was engineering literacy: understanding how APIs communicate, how to read a server log, how to isolate a failing component. The AI amplified what I already knew. It didn’t replace the need to know it.

That’s a conversation worth having with students before they assume AI has made learning fundamentals optional.

What’s Next

The cycling app works. You can see the front-end at https://ryanjames1729.github.io/bike-wattage-calculator/. And now I’m taking the same approach into a bigger project: building a live data dashboard for my department using Google’s Gemini. The goal is a system that can log into password-protected internal tools, scrape operational data, and surface insights that would take hours to compile manually.
Same philosophy: I’m not asking AI to be the engineer. I’m using it to make the engineering faster.

What are you building with AI lately? I’m especially curious how other educators are thinking about this — the wins, the walls, and what it’s changing about how we teach problem-solving. Drop it in the comments.

Keep coding!
Ryan

This post first went out on Substack. Join the conversation there or subscribe to get the next one by email.