TL;DR: I'm an email developer who had an idea for 2 years, couldn't afford engineers, and used Claude Code to build a live product with AI vision, payments, and real users in less than a week.
The Idea That Wouldn't Go Away
For 3+ years, my entire world was HTML tables, inline CSS, and making buttons render in Outlook. I've built over 3,000 email templates. I know what mso-line-height-rule: exactly does. That's the kind of developer I am + a passionate food blogger.
But I had this idea that kept nagging me.
Every week, the same thing happened in my kitchen: I'd buy fresh herbs, vegetables, ingredients for recipes I was definitely going to make. By Friday, half of it was wilting in the back of my fridge. The parsley died. Again.
I wanted an app where I could just photograph my fridge and get instant recipe suggestions — prioritized by what's about to expire. No barcode scanning. No typing expiration dates. One photo.
For two years, this idea lived in my head. I looked into hiring developers. The quotes made me close my laptop and walk away. As a solo creator, it felt impossible.
The "Why Not" Moment
Then I tried Claude Code with a Max subscription.
What happened next is honestly still hard to believe. In under a week, I went from idea to a live, deployed, payment-processing product.
Here's what I built:
✋Don't Touch It! — Snap your fridge, get recipes sorted by what expires first.
Live at: https://dont-touch-it-app.vercel.app/
The Tech Stack
| Layer | Tool | Why |
|---|---|---|
| Framework | Next.js 14 (App Router) | SSR, API routes, everything in one repo |
| Database | Supabase (PostgreSQL) | Free tier, auth built-in, real-time |
| AI | Claude Vision API | Best image understanding I tested |
| Payments | Stripe | One-time credit packs, webhooks |
| Hosting | Vercel | Auto-deploy from GitHub, PWA support |
| Styling | Tailwind CSS | Fast iteration, mobile-first |
Total monthly cost to run: ~$120 (API budget + subscription). Everything else is on free tiers.
How the AI Magic Works
The core feature is deceptively simple:
User uploads a fridge photo
Claude Vision API identifies every visible ingredient
AI estimates perishability for each item (days until it should be used)
Ingredients get color-coded: 🔴 use today, 🟡 use this week, 🟢 no rush
Recipes are generated prioritizing the most perishable items
The tricky part was prompt engineering. My first attempts identified maybe 50% of visible items. After iteration, I got it to scan every shelf, drawer, door compartment, and even partially hidden items. A well-stocked fridge now returns 15-25 identified items.
The perishability estimation was another rabbit hole. Getting AI to accurately know that fresh basil lasts 2-3 days while eggs last 3 weeks — and everything in between — required very specific prompting.
The Hardest Parts (That Nobody Warns You About)
- Webhooks in Production Testing Stripe webhooks locally with the CLI? Easy. Getting them to work reliably in production? I lost an entire afternoon. The signing secret alone had me questioning my career choices.
- Image Handling on Mobile Phone cameras produce massive images. Uploading a 4MB photo over a cellular connection is painful. Compression, resizing, format handling — this was way more work than I expected.
- Environment Variables I had 15+ env vars by the end. Keeping local, test, and production variables in sync — and remembering which Stripe keys are test vs live — nearly broke me.
- The Hat-Switching Tax The hardest part wasn't any single skill. It was switching between them 15 times a day:
- 9 AM: debugging a Stripe webhook
- 10 AM: picking a button color
- 11 AM: writing Instagram copy
- 12 PM: restructuring a database table
- 1 PM: filming my fridge with the cat in the background
Each context switch costs brain power. By day 4, I was running on coffee and stubbornness.
Monetization: Keeping It Simple
I didn't want subscriptions for v1. Too complex, too much commitment for users.
Instead: credit packs.
| Pack | Price | Per Scan |
|---|---|---|
| 3 scans | $0.99 | $0.33 |
| 10 scans | $2.49 | $0.25 |
| 15 scans | $3.99 | $0.27 |
Everyone gets 3 free scans per day. Power users can buy more.
What I'd Do Differently
- Start with the webhook setup — Don't save payments for last. Set up Stripe, webhooks, and credit logic early.
- Mobile-first from day one — 90% of users will be on their phone. I caught layout issues late that could've been avoided.
- Prompt engineering deserves its own sprint — The AI prompt is the product. Treating it as an afterthought was a mistake.
- Ship the ugly version faster — I spent too long on button colors when I should've been validating the concept.
The Numbers (Week 1)
- Development time: ~6 days
- Total cost: $120 (API credits + subscription) + $0 (all free tiers)
- Lines of code: I honestly don't know, Claude Code wrote most of it
- My contribution: product vision, design decisions, prompt engineering, and a lot of debugging, marketing, social media engagement.
Try It
🍳 https://dont-touch-it-app.vercel.app/
3 free scans every day. Works on any phone or computer.
If you're an email developer, a designer, a marketer, or anyone who thinks "I'm not a real developer" — you can build products now. The barrier has never been lower.
And if you've ever thrown away forgotten parsley — this app is for you.
Top comments (0)