How I Vibe Code: My AI-Assisted Development Workflow

2026-02-12 · Jazz Lien

I've shipped 10+ side projects in the past year. Not because I'm a 10x developer — I majored in foreign languages, not computer science. I figured out how to vibe code effectively.

Here's my actual workflow.

What is Vibe Coding?

Vibe coding is AI-assisted development where you describe what you want in natural language, and an AI helps you build it. The "vibe" part? You're having a conversation with your tools, not fighting syntax errors.

It's not "no-code." You're still writing real code. You're just not writing it alone.

My Stack

Claude Code — My primary coding partner. I describe features, it writes code, I review and iterate.
React + TypeScript — Frontend. Type safety catches bugs before users do.
FastAPI + Python — Backend when I need one. Fast to write, fast to run.
GCP / Cloudflare — Deployment. I know GCP from my day job, so I stick with what I know.

The Actual Process

1. Start with a Clear Problem

Before touching any tool, I write down:

What am I building?
Who is it for?
What's the MVP — the simplest version that's actually useful?

If I can't answer these in one sentence each, I'm not ready to build. I'm ready to think more.

2. Scaffold Fast

I describe the project structure:

"Create a React app with a landing page, about page, and contact form. Use Tailwind for styling. Dark theme. Keep it minimal."

Within minutes, I have a working skeleton I can run locally. No boilerplate copying. No hunting for starter templates.

3. Build in Conversations

Each feature is a conversation:

"Add a form that lets users submit their email. Validate the format. Show a success message. Store submissions in SQLite for now."

The AI writes the code. I review it. If something's off, I explain what's wrong in plain English. We iterate until it's right.

The key: be specific about what you want, vague about how to implement it. Let the AI figure out the how.

4. Don't Be Precious

The AI will write code differently than I would. Variable names I wouldn't pick. Patterns I'm not used to.

That's fine. If it works and it's readable, ship it. You can refactor later. You probably won't need to.

5. Test by Using

I'm not writing unit tests for every side project. I test by using the product like a user would:

Click every button
Enter weird inputs
Resize the window
Try it on mobile

Fix what breaks. Ship what doesn't.

Real Examples

拾光 Glimmer — AI memorial video platform. Built the MVP in a weekend. I described the video generation flow; Claude wrote the React components and API calls. I focused on the emotional UX — what a grieving family actually needs to see.

Jazz Gallery — Photo gallery with admin authentication. Described the auth flow in plain English, got working JWT authentication in 30 minutes. Previously, auth was the thing I'd procrastinate on for days.

This Website — The site you're reading. React frontend, FastAPI backend, markdown content. I designed the layout; AI implemented it. We went back and forth on the dark theme until it felt right.

What AI Can't Do (Yet)

Let me be honest about the limits:

Make product decisions — AI doesn't know your users. You do.
Debug weird state bugs — Sometimes you need to actually think, not just prompt.
Replace taste — AI can implement your vision. It can't have vision for you.

I'm not a passenger. I'm the driver with a very capable navigator.

Common Mistakes I See

When I mentor people in vibe coding, they often:

Prompt too vaguely — "Make it better" doesn't help. "Make the button larger and add hover state" does.
Accept code they don't understand — Always read what the AI writes. You're still responsible for it.
Fight the AI's style — If the code works, let go of your preferences.
Skip the problem definition — Building the wrong thing fast is still building the wrong thing.

Why This Matters

Vibe coding isn't about replacing developers. It's about leverage.

I can now test ideas in hours instead of weeks. That means:

More experiments (most will fail, that's fine)
Faster validation (find out if users want it before building everything)
Less sunk cost (easy to throw away a weekend project, hard to throw away a month of work)

For someone like me — a non-CS background person who learned to code in my late 20s — this is a game changer. The gap between "idea" and "working product" has never been smaller.

Want to Learn?

I mentor people through building their first AI-assisted projects. If you have an idea and want to ship it, check out my Services page.

What are you building? I'd love to hear about it.