How I screwed over my client last week

Last week I picked up a small gig and wanted a simple order page. I had AI build it in three hours, sent it to the client — looked pretty decent. Next day, after the client paid, orders were a mess — same order billed three times, two just disappeared.

I got stuck at this exact step: I figured "if AI can write it, why bother learning?" But clients don't care whether you used AI. They'll just say "your thing is broken."

The real trap with AI coding: not the tool, but "I don't get it"

This play is called "Vibe Coding" in Silicon Valley — you tell AI in plain words what you want, it builds the page or mini-app for you. Indie dev Pieter Levels shipped several profitable micro-products solo using this approach. A friend of mine, Xiaoli (a side-hustler), used it to spin up a client intake page.

But an article on ssp.sh punctured one truth: AI's code itself isn't broken. The problem is the user doesn't understand "system architecture" and "intent" — simply put, how data flows from the client's hands to your account, and what potholes sit in between.

Analogy: you ask AI to set up a small shop, it builds you a checkout counter. But you don't know how money gets logged, how inventory gets deducted, where data lives. One power outage later — money and goods gone.

Cost to replicate this "AI-built system" today

Money: $0 to start. Cursor, Claude, Windsurf all have free tiers; ~$20/month subscriptions are enough.
Time: three hours gets you a decent prototype, but actually understanding "how it runs" takes at least another week.
Technical barrier: you just need to know how to chat. But you have to be willing to learn basics like "where does data live, how does the flow work."
First step: open Claude.ai and ask it: "I want to build an order page. What's step one for designing how data flows from order to payment?"

Playbooks by stage

Just starting out: Don't touch AI writing full systems yet. Let it write a landing page or copy for you — saves you the anxiety of learning basics the hard way.

Got 1-2 clients: AI-written tools can save time here, but after every step AI writes, push back: "If the user pays, where does this money go? Where does the data live?" If you can't explain it, don't ship it.

Scaling up: You can seriously use pro tools like Cursor, but find a tech-savvy friend for a one-hour weekly chat to review your architecture. Don't skip that spend.

This tool isn't for everyone. If your business runs fine on WeChat groups and manual bookkeeping, AI-built systems are a luxury, not a necessity. Skipping it for now is fine.