Vibe coding: Observations of a practitioner
As someone who love building few things, and breaking few others I went into the deep-end the vibe coding. After spending long time experimenting and longer time debugging, I realized that Vibe-coding is a double edged sword. And there are quite a few caveats in using these cutting-edge tools.
Right balance of Context and instructions: With the right amount of context and prompt the latest models will be able to generate some impressive pieces of code. But the pitfall is provide the right amount amounts.
LLMs are giant probability machines: LLMs models predict the next word, sentences by calculating and refining probabilities. And based on model evolution the probabilities might change, that means same input can give different outputs.
Efficiency of workflows: These models sure can generate 1000's of lines of code, and 100's of files. That is great if you are building from scratch. If you are working on a codebase with preset coding patterns, design choices and organizational constraints, will the model able to discern these and help you follow those patterns.
Co-dependency and Trust: In the early days of ChatGPT craze, I once saw a dev generate an SQL script for a report. As time progressed he considering ChatGPT as a tool, then a co-worker, then as always subject matter expert. The more I use these tools, the less docs I read, less experiments I conducted and less I learned or levelled up.
Further Readings:
- Other great article I liked is https://cendyne.dev/posts/2025-03-19-vibe-coding-vs-reality.html?ref=dailydev
