A plain summary, so you can get the gist here without leaving.
In 2025, researchers Advait Sarkar and Ian Drosos published the first empirical study of vibe coding, the new habit of building software mostly by talking to an AI rather than writing code by hand. Instead of guessing how people do it, they watched them do it and listened to them think out loud.
What the paper studied and how
Vibe coding is the practice of steering a code-generating AI through conversation, prompting it, glancing at what it returns, and prompting again, often without reading every line closely. Because the practice was so new, almost everything written about it was opinion or anecdote. This 2025 study set out to gather real evidence.
The method was careful and human. The researchers gathered curated recordings of people in extended vibe coding sessions who narrated their thinking as they worked, a technique long used in studying how people program. They then analyzed these think-aloud sessions systematically, looking for recurring goals, workflows, prompting habits, debugging moves, and the points where things went wrong.
The core finding: a loop, not a straight line
What they saw was not a tidy march from idea to finished program. It was a repeating cycle. A person sets a small goal, prompts the AI, then quickly judges the result by scanning the code or simply running the application to see if it behaves. Based on what they observe, they prompt again, edit by hand, or back up and try a different angle.
Each turn of the loop nudges the work a little closer to what the person wanted. The researchers also recorded the friction in this cycle, the moments where the AI misunderstood, where bugs hid behind plausible-looking code, or where the developer struggled to express exactly what they meant. Naming these pain points is part of what makes the study useful.
Why this matters for building with AI
This paper matters because it turns a buzzword into something you can examine and teach. When you can describe the actual loop people follow, you can spot where the loop breaks and design better tools, prompts, and habits around it. That is far more useful than slogans about the future of coding.
For a community learning to build with AI, the gentle lesson is reassuring. Vibe coding is not a single magic trick; it is an iterative conversation full of small checks and corrections. Getting good at it means getting good at setting clear goals and verifying results, the same disciplined habits that show up everywhere else in working with these tools.
- Published in 2025 by Advait Sarkar and Ian Drosos, this is the first empirical study of vibe coding.
- Evidence came from watching curated think-aloud sessions of people coding by conversation, not from opinion or theory.
- The central finding is an iterative loop: set a goal, prompt the AI, evaluate by scanning or running it, then prompt or edit again.
- The study names common friction points, including misunderstandings and bugs hidden inside plausible-looking code.
- Seeing the real loop makes vibe coding teachable and points toward better tools, prompts, and the habit of verifying results.
arXiv
New to this? Come build with us.
Reading is good. Building with people is better. Our drop-ins are free and open to total beginners.