🤖

My ultimate Vibe Code setup (conversational programming/vibe wrangling) (30 july 2025)

Date: 30 July 2025
Here's a glimpse into my current coding setup for samvannoord.nl, embracing maximum automation, minimum effort, and just enough AI hype to make me feel cutting-edge. Welcome to my sarcastically efficient "vibe coding" workflow.

How Does It Work?

Let me break down my groundbreaking workflow step by step:
  1. Craft a Perfect Prompt for Codex
    1. Because writing actual code is so last decade, I carefully feed my wishes into Codex. It’s like outsourcing your brain—but cooler, obviously.
  1. Automatic Pull Requests
    1. Codex spits out some code suggestions. If these don’t immediately explode on sight, I bravely command it to create a Pull Request (PR) on GitHub—because who manually opens PRs anymore?
  1. Vercel Previews Everything
    1. Vercel obediently watches my GitHub repo and builds a preview environment for every PR—saving me from tedious manual deployments.
  1. A Quick Sanity Check
    1. Within minutes, I inspect the changes in the Vercel preview. Just a quick visual check, because obviously, rigorous testing went extinct with the dinosaurs.
  1. One-Click Merge & Live Deployment
    1. If everything looks shiny enough, I click once or twice in GitHub, and Vercel magically deploys my changes to the live site. No stress, no fuss—just vibes.

Tech Stack for Maximum Vibes

Efficiency: A Vibe Check

Here’s the harsh reality of how long this workflow actually takes:
  • One carefully crafted prompt to keep Codex happy.
  • 3-minute coffee break while Codex generates something probably usable.
  • Brief glance at the diffs to pretend I understand what happened.
  • 1 effortless click to make a PR on GitHub.
  • Another 1-minute break while Vercel builds the preview.
  • A lazy visual inspection because "it works on my machine" is now officially replaced by "it looks good enough on Vercel."
  • 2 casual clicks to merge the PR and get back to more important things.
  • Automatic deployment, because manual deployments belong in a museum.

Multitasking Like a Pro

Thanks to this almost-too-easy setup, I can parallelize multiple tasks. While one change is busy building itself, I effortlessly move on to another prompt. Multitasking has never felt this effortless—welcome to the "productivity" revolution.

Screenshots or It Didn't Happen

How the tasks look in Codex
How the tasks look in Codex
Vercel automatically sees a new branche and builds it, as you can see usually <1 minute. A preview is ready for me to review, and test is the website performs as it should.
Vercel automatically sees a new branche and builds it, as you can see usually <1 minute. A preview is ready for me to review, and test is the website performs as it should.
And maybe the worst part, you can multitask in codex, for extra fast slop creation.
And maybe the worst part, you can multitask in codex, for extra fast slop creation.
This is a finished task in Codex, here you can review the generated code, see the changes and publish (and view) a pull request in the top right.
This is a finished task in Codex, here you can review the generated code, see the changes and publish (and view) a pull request in the top right.
2 clicks, merge & commit the pull request, and It’s done. It’s now in the main, and Vercel will build & deploy this live.
2 clicks, merge & commit the pull request, and It’s done. It’s now in the main, and Vercel will build & deploy this live.
 

P.s.

If it wasn’t apparent to you yet, I had chatGPT 4.5 generate this entire page. I put in some whatsapp messages I sent to my friends, talking about this pipeline, instructed it to leave some room for screenshots and be a bit sarcastic about it all, and ta-da.
 
This experiment, as I would like to call it, came to be by accident. I was working with Next.js anyways. I had prepros set up so that I didn’t have to manually upload using FileZilla all the time. At some point I got bored and wanted to see what Vercel could do, and I found out it could watch GitHub repos.
Some more time later, I was playing around with ChatGPT, as I do more often, and a popup stating Codex was released appeared. Of course I wanted to play around with this, as it was included in my Plus subscription. And wow, it was able to connect with my GitHub and access the entire repo. It actually understood the repo, what was in it and the logic behind it, without me instructing it at all. I was impressed.
So some more fiddling around with it, and I had accidently created this pipeline.
 
Being honest, it’s fun for me to experiment with ai and setting up this pipeline and seeing it’s functionalities. I hate and love it at the same time. For me this is a more fun way of coding, because a lot of the simple work is being done. No more remembering syntaxis (except for some basics of course), but looking for design patterns and training myself in describing what I want. It feels more like a puzzle now then it has ever felt before.
It sometimes impresses me how easily and perfectly it can create something from extremely simple prompts, but other times it messes up the simplest changes. The capabilities it has change a lot, and with the past year the improvements are very notable. Also, the better I get at prompting (which reminds me of getting the hang of some unknown video game with really strange controls and even more strange outputs, it’s never as expected), the better the results actually are.
 
To get it working properly, I definitely need to edit some code it generates and nudge it in the right directions a lot of the times, but for me it still takes time and complexity off the task of programming as a whole. (web dev = programming? that’s another discussion)
 
There are definitely downsides to this, and I definitely use it carefully. Out of 5 quick use cases I tested it with, 1 was definitely not doing what I wanted. Getting the navbar to stay “sticky” on top was a challenge, it did it 90% right the first time, but kept editing just a bit too much. I could’ve done this faster myself.
Security is also an extra thing to be careful with in my opinion, I’m currently only hosting a front-end and use my notion pages as my only back-end data source. This I carefully put together, storing my API secrets appropriately and of course setting the API up in such a way that it is read only for the files I chose, and nothing more.
 
Knowledge of (basic) security principles is a must before going into this. Having to use less brainpower to create applications is a bliss and a burden at the same time.

P.p.s

I was debugging something, this is in CLI with GPT-5. He is just so excited to get into it all. Funny thing.
I was debugging something, this is in CLI with GPT-5. He is just so excited to get into it all. Funny thing.
 
My ultimate Vibe Code setup (conversational programming/vibe wrangling) (30 july 2025) - Sam van Noord | Sam van Noord