First experiment with Windsurf: Can I build and deploy I useful site in a single weekend?
2025-06-18
0. Who’s Writing This?
I’m have spent years whipping up internal tools: everything from CLI tools to full web-apps in Backbone.js, PHP, and plain‑vanilla Python. UI polish has never been my strong suit—most of my apps lived behind a VPN, and the users didn't generally have a choice about using them. Until this project, I had never touched React and had never shipped anything truly public‑facing. Aquacost.com was my experiment to see how far AI tooling—and a weekend’s worth of grit—could take someone who’s comfortable with back‑end logic but allergic to CSS quirks.
1. The Spark
Two forces collided:
- A brand‑new pool heat pump. I wanted to know—really know—how much running it would cost under different schedules and weather patterns.
- A desire to level‑up my AI‑coding chops. Windsurf’s “single‑prompt to production” pitch felt like the perfect playground, with ChatGPT o3 ready to double‑check my physics.
Result: Aquacost.com—a web app that simulates pool‑heating costs for any location—was born over one weekend.
2. The Tech Stack (and Why)
| Layer | Tech | Why I Picked It |
|---|---|---|
| Frontend | React + Vite on Cloudflare Pages + Render.com backend | free. |
| Backend | Flask (Python, no DB) | Zero friction; familiar territory. |
| Physics Sim | Custom water‑energy model, iterated with ChatGPT o3 | Fast math and instant unit tests. |
| Dev Agent | Windsurf | Promised to scaffold the whole repo. |
| Manual AI Assist | Carefully crafted o3 prompts in ChatGPT UI | For precise, context‑rich edits. |
3. What Worked
- Lightning‑fast scaffolding. Windsurf handled the blank‑page terror.
- Physics pair‑programming with o3. The LLM caught unit mistakes I would have missed.
- Edge hosting. Static front end plus a tiny API for zero cost.
4. Where It Broke
- Complexity cliff. Once the repo topped a few hundred lines, Windsurf’s diffs touched everything—failing lint, renaming constants, undoing fixes.
- Tiny UI changes produced huge diffs. A request to center a button triggered 30‑file rewrites.
- Context drift. Windsurf’s “o3 under the hood” produced worse answers than direct o3 chats with explicit context.
- Opaque reasoning. When it went sideways, there were no breadcrumbs to debug.
Takeaway: a great intern, but a lousy senior engineer. Use agents for green‑field scaffolds, switch to surgical prompts for refactors.
5. Lessons Learned
I will say my critiques and commentary are mostly applicable to Windsurf. I can't speak to how other agentic coding apps perform at this point other than to say Claude Code is f*ing amazing and blows this experience out of the water.
| # | Insight | Why It Matters |
|---|---|---|
| 1 | Curate your prompt. Direct, context‑rich chats with o3 beat any black‑box agent. (at least for Windsurf) | Deterministic and debuggable. |
| 2 | Amazing at sketches, reckless at global refactors. | Guardrails keep projects sane. |
| 3 | Architect your project up-front into modules which will fit in a reasonable context window | Keep your code tractable and decoupled. |
6. Final Thoughts
Aquacost.com proved that a back‑end‑leaning, UI‑shy coder can ship a polished public app in a weekend—with strategic help from modern AI tools. Windsurf got me airborne; ChatGPT o3 ensured the physics held up; Claude Code helped me land smoothly.
If you have been sitting on an idea, block off a weekend, grab your favorite LLM pair‑programmer, and just ship. Worst case, you learn a ton. Best case, you create something people actually use—while leveling up your own stack in the process.
Happy hacking!