Photo by Joel Timothy on Unsplash
Today is nearly the one-year anniversary of the term “vibe coding,” a kind of light-hearted, kind of silly term that has spawned companies, bootcamps, and tech influencer videos.
At least according to the Wikipedia page (and the aforementioned X post that spawned it into being), vibe coding implies that the user does not know what he/she is doing. There is no reviewing or editing, the user just hopes things work out and more or less abstracts away the code.
I can’t help but think…what would it take to just make up a new term? Let’s try it. Let’s coin a new phrase called “vibrato coding,” and just see if anything happens.
Vibrato Coding
It kind of goes without saying that “vibe coding” is a popular term because of the person who said it. If I try to invent my own term/method, everyone will question who in the hell I am to suggest anything. So, how about this?
Vibrato Coding: A software development practice in which the user generates code using LLMs, then takes the time to understand it
That’s it. But while “vibe coding” invokes the image of an inexperienced person casually conjuring functional code (analogous to this clip of Harper Sterling vibe-trading with no stop loss during her first week on the job as an intern), “vibrato coding” invokes the image of a concert violinist skillfully closing out season one of Arcane.
A better definition is here:
While coding my AI that makes software development faster, I’ve developed a comprehensive approach that balances AI assistance with engineering excellence:
- When using AI tools, start with a clear architectural vision. Before generating any code, document your requirements, constraints, and expected behavior. This becomes your validation framework for any AI-generated code.
- Instead of accepting entire functions or components from AI, break them down into smaller, understandable chunks. I’ve developed a three-step process:
- Generate small, focused pieces of functionality
- Review and understand each piece thoroughly
- Integrate only after validation and testing
The italicized method above was not created by me, but outlined in this blog post by Namanyay Goel. I just wish it had a name. It doesn’t even need to be that specific – it simply means still using AI, but taking the time to read the code copied. Years before generative AI hit the mainstream, someone named Aphinya made a similar argument about the dangers of blindly copy-and-pasting from StackOverflow.
You can generate code with AI, and you can test it, but without reviewing each line you risk the dangers of hallucinations. AI tends to still introduce unexpected problems, which can be mitigated with code review.