unknown

A developer shares firsthand experience vibe coding a small web project using GitHub Copilot with VS Code Agent Mode and GPT-5 mini, on a stack of Go, SQLite, React, Redux, and Sass. The positives include fast boilerplate setup, dependency installation, and incidental learning of CLI tricks like lsof and curl. The negatives are significant: the AI struggled with more than a handful of moving parts, hallucinated fixes (nearly wiping a database), installed outdated dependencies, and produced code that didn't match the versions it installed. The author concludes that vibe coding is useful for boilerplate and learning but unreliable for anything complex, and recommends treating AI output like code from a distracted junior developer with access to last year's internet.

What I learned from vibe coding a small web project with GitHub Copilot and GPT-5 mini

Keith Solomon

Maybe he should take a look here
<a href="https://github.com/GoogleChrome/modern-web-guidance/tree/main" target="_blank" rel="noopener nofollow">https://github.com/GoogleChrome/modern-web-guidance/tree/main</a>

It reads more like a benchmark of GPT-5 mini or the harness. I’m not even working with bleeding edge, at the moment I mostly use Claude Code with Opus and Sonnet 4.6 for planning and executing, and my experience is nothing like that. With good planning and guardrails they are capable of very complex tasks. I cannot remember the last time they hallucinated either.

Not sure why would you try to use the hard way rather than just plug in Claude as I use it daily in my work.
Doesn’t hallucinate at all in my and our company use case as long as you know what to request, etc.