Vibe coding takes effort to produce production code, despite opinions to the contrary.
I’ve been playing a lot with code generation via LLM recently. As such, I’ve had a few conversations of late that have gone something like this:
Me: I vibe coded this cool thing
Them: Whoa? Really? How long did that take you?
Me: Welp, around 18 hours!
Them: 18 hours?! Why so much time?! I heard that ChatGPT produces code in an hour!
Let’s set aside the fact that a few hours to develop a full production version of a system is nothing short of a modern miracle. A machine that outputs a complete application is pretty neat. However, it still needs guidance … and therefore, (some) time.
Even with the most advanced models, I still get my best results by systematically walking the LLM through the thought process I myself would use to implement/evolve a system.
It turns out that practicing evolutionary design, while adding a few extra coding hours, leads to higher quality code and more consistent output from a coding agent.
Giving discrete, succinct, and precise directions leads to better code every single time.
Solid, production-grade code has always taken longer than spaghetti “demo” code … even with an LLM.
Be wary of anyone claiming a production system emerges from a simple prompt, unattended and uninformed.
Be wary also of anyone claiming an LLM can’t produce a system capable of serving real-world customers … these systems are capable, they just need a little more guidance.
Originally shared on LinkedIn, May 26, 2025.