Dan Shiebler, Abnormal Security’s head of machine learning, thinks of the approach to coding as “based on the inputs and outputs, rather than how things will work.”
Janet Worthington, senior analyst for security and risk at Forrester, sees it as a way to iteratively prompt a GenAI editor “so you don’t have to have any understanding of what the underlying code is.”
For Andrej Karpathy, former director of AI at Tesla, it’s…
…vibes, man.
“There’s a new kind of coding I call ‘vibe coding,’ where you fully give in to the vibes…and forget that the code even exists,” he wrote in a Feb. 2 post on X.
Shiebler and Worthington shared with IT Brew whether their development work lately has taken on the style of “vibe coding.” Both said the prompt-heavy style is helpful for small, low-stakes projects where a wrong iteration won’t have catastrophic consequences.
“It just takes things that otherwise would have been too small and too annoying, and reduces them to the point where now you can do that,” Shiebler said.
Vibe check #1. Shiebler recently wanted to write a script that took specific spreadsheet columns, sent them to OpenAI’s embedding system to be arranged in training “clusters,” and rewrote the cluster assignments in a CSV format—a coding task where he used text prompts rather than traditional programming.
He employed “AI code editor” Cursor, which turns natural language inputs into code.
With Cursor, Shiebler iterated on the description of the task until the results looked right.
He had to modify an “insufficiently specific” prompt, he said, like when he wanted the results to return in their own folder versus a new directory altogether.
The script was not tremendously complex, Shiebler said, making it well-suited for a text-to-coding approach.
“If the script might run only once, then I’m probably not going to want to spend an hour-and-a-half writing it. But if I can prompt…it into existence in five minutes, I might write it otherwise.”
Programmers have increasingly accepted a level of AI into their workflow. A GitHub survey of 2,000 developer team members, released in July 2024, found more than 97% had used generative AI and LLMs to provide software-development assistance.
Top insights for IT pros
From cybersecurity and big data to cloud computing, IT Brew covers the latest trends shaping business tech in our 4x weekly newsletter, virtual events with industry experts, and digital guides.
Vibe check #2. Worthington recently built a website displaying weather based on zip code. Though Worthington declined to name the AI-integrated development environment (IDE) platform used for the job, she said the tool answered her prompts with recommendations: JavaScript, for example, when Worthington rejected the platform’s initial idea of using React, as well as weather-site API keys and ways to test functionality.
“You’re creating something without having to get into the minutiae of settings, without having to know all the technical pieces of setting things up. And I didn’t even look at any of the code,” Worthington told IT Brew.
Like Shiebler, Worthington sees the tool as valuable for prototyping scenarios—mockups of a front-end design and user interface, for example—but insufficient for jobs that handle customer data, where security mechanisms like encryption must be implemented with insight from human developers.
IT Brew recently wrote about low-quality software reports created by large language models. One Cursor output, in March 2025, reportedly refused to assist a vibe coder.
“If I put my money in a bank and I was using that mobile app, and they told me that somebody vibe-coded it, would I use it? No,” Worthington said. “But if they told me that they vibe-coded it to figure out what the front-end is, and then they actually did software development, maybe where they were using generative AI as part of the process, sure.”
Along with Cursor, Shiebler has used Lovable, which offers chat-based software engineering and integrates directly into infrastructure, he said. With those kinds of tools, “you can get pretty far without thinking about the code,” he told us.
“Those tools are not quite at the point yet where someone who is totally non-technical, without any coding experience, is going to be able to build things that are more than toys…but the future is very clear.”