I tried vibe coding with AI after 10 years away from IDE. It took me 2.5 months to ship one small tool.
I didn’t plan to write this, but after 2.5 months, I feel I should.
I used to be a developer. I loved coding — but I also hated it.
In my younger days, coding meant overnight marathons, pressure, coffee, burnout. I eventually escaped into management: meetings, emails, endless PowerPoint decks. Busy, but far from building.
That’s why I didn’t really code for more than 10 years.
A few months ago, I started hearing about “vibe coding”.
People said AI could help you build again, without destroying yourself.
I wanted to test that.
I also love images, design, and colors. I’m an amateur photographer (once won a National Geographic photo competition), so I picked a visual problem as my first experiment.
It’s called HexPickr, but the output is mostly OKLCH and Tailwind CSS. HEX is where people start, but it’s no longer where modern CSS should end.
I thought AI agents would make this smooth.
That was naive.
I started with Codex v5.1. It wrote code fast, but had serious memory issues. A real example: my Prisma field was clearly updatedAt, but Codex kept hallucinating updated_at, even after I pasted the schema and said “DO NOT rename fields”.
Then came i18n. One wrong key = white screen. Codex basically refused to help.
This is where ChatGPT (used as PM) helped me slow down and plan changes safely.
Later I used Codex and Antigravity in parallel. Antigravity reasoned better, but trial tokens ran out at the worst moments. Codex v5.2 improved things, but every prompt became a legal contract.
Anyway, HexPickr is live: https://hexpickr.com/?utm_source=indiehackers&utm_medium=community&utm_campaign=launch_story
I’m not selling. I’m curious:
Did AI help you enjoy coding again — or just change the pain?
This resonates hard. I also stepped away from coding for years (management track), and "vibe coding" got me back into hacking. Your experience with Codex hallucinating field names is so real. I've found the same. AI writes fast but doesn't understand context. The "updatedAt vs updated_at" thing happens constantly, especially with Prisma/ORMs where naming conventions matter.
The ChatGPT as PM insight is good. I do the same... use it to think through architecture before coding, not just to generate code.
Did you find yourself enjoying the process again, or was it more "finally, I can build without the old burnout"?
The renamed fields thing drives me mad. I've had the same experience where you explicitly paste the schema and say "this is the source of truth" and it still invents its own version. The prompting style that's helped me is treating it like a very literal junior dev - I'll write out exactly what I don't want it to touch, almost like a contract.
The i18n breakages are real too. One missing key and suddenly you're staring at a blank page with zero useful error message.
Interesting that you used ChatGPT as a PM layer. I've started doing something similar - using one model to plan the approach and a different one to execute. Keeps the execution context smaller and the reasoning cleaner.
Congrats on shipping though. 2.5 months for a solo project while relearning the whole ecosystem is actually pretty solid.
Thanks man, I really appreciate that! 2.5 months felt like an eternity while I was in the thick of it, but looking back, it was definitely a crash course in the 'new way' of shipping.
The 'Literal Junior Dev' analogy is spot on. I eventually started using a similar 'Contract' approach—basically telling the agent: 'You are allowed to change the logic in this specific function, but if you rename a single variable in the export object, the whole system breaks. Do not touch them.' Even then, you have to watch them like a hawk.
That i18n strike was the funniest and most frustrating moment. I think the model just hit a 'context wall' and decided it was too bored to map 50 keys, so it just gave up. You’re right about the 'blank page'—debugging a missing translation key in a silent AI-generated build is a special kind of hell.
I’m definitely sticking with the 'PM Model vs. Execution Model' split for my next project. I will further add the role of image designer and a design director, in charge of the overall site design. It’s the only way to keep the execution agent from getting 'distracted' by the big picture and hallucinating the small details.
Are you using Claude Code or Cursor for your execution layer? I found Antigravity was the only thing that could actually save me on the complex math bits.
I've been in a similar boat recently with returning to code and wrestling with AI agents. What worked for us was using specialized .cursorrules and keeping the context window very small for the agents to prevent those 2.5-month-long hallucinations. Also, for the environment issues, Dockerizing the dev-stack helped us avoid those Node version nightmares! Good luck!
Thanks for the tips. I didn't know about cursorrules yet, let me check how to stop these long hallucinations
Do you find that keeping the context window small makes the agent forget the bigger picture of the app? I am still learning how to talk to these agents better
Docker is also a good idea, I had many Node version headaches, thanks
BTW, good luck with your projects too!
This comment was deleted 2 months ago.