A new category of software exists now, and it had no name for most of computing history. The program written for exactly one user, by that user, who cannot program. A script that renames your photo library the way you specifically like it. A page that tracks the one habit you actually care about. A tool with a market of one, built in an afternoon, by describing it.

The workflow got its name in February 2025, when the AI researcher Andrej Karpathy described a style of building where you “fully give in to the vibes, embrace exponentials, and forget that the code even exists.” His summary of the method was cheerfully unrigorous. “I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works.” The phrase vibe coding stuck hard enough that Collins Dictionary made it a word of the year, and the practice has since matured toward something the professionals call agentic engineering. This post is about what the amateur version is actually like, because the amateur version is the revolution.

What It Feels Like

You open a chat with a capable model and describe an outcome. Not code, an outcome. I want a script that takes every image in this folder and adds my site’s title style over a blue gradient. The model writes something, you run it, and one of two things happens. It works, which is startling the first dozen times, or it fails, and you paste the error back like a patient describing a symptom, and the second version usually runs.

Steering feels less like engineering than like directing a contractor who is impossibly fast, weirdly obedient, and occasionally wrong in confident ways. Your job is taste and verification. You know what the tool should do, you recognize when it does it, and you keep asking until the gap closes. The site you are reading runs on tools built exactly this way, including the script that draws every post’s cover art and the checklist that audits every draft. Neither existed as a product anyone sold. Both were described into existence and refined by pointing at what was wrong.

What It Is Genuinely Good For

The sweet spot is software for one. Personal tools sat below the economic waterline for the entire history of the industry, since no developer could be hired for a task worth twenty minutes a week, and now the waterline is gone. Automations for your own files, calculators for your own strange spreadsheets, small sites for your own projects, scripts that glue two apps together in the one way you need. If you once learned a little Python and stalled, this is the on-ramp that was missing, because the machine handles the syntax that stalled you while you supply the intent that it lacks.

There is also a quieter benefit, and it sounds old-fashioned. Making your own tools changes your posture toward all tools. The person who has built even one small program stops experiencing software as weather and starts experiencing it as decisions, and that shift in posture survives even when the code was mostly the machine’s.

Where It Quietly Falls Apart

Honesty requires the other column, and the old stories supply the right warnings.

The first failure is instructions followed too literally. Every golem tale turns on this, and vibe-coded tools fail the same way, doing exactly what you said in the sentence you did not think hard about. The script that renames files will happily rename the wrong folder. Test on copies. Ask the model what could go wrong with its own code, a question it answers surprisingly well.

Unread contracts are the second failure. Code you cannot read is a promise you are taking on faith, which is tolerable for a photo renamer and reckless for anything touching money, passwords, other people’s data, or anything you cannot afford to lose. The amateur rule is a bright line. Vibe code the tools whose worst failure is an annoyance, and buy or borrow the ones whose worst failure is a disaster.

Failure number three is maintenance, the tax nobody mentions. A tool that worked in spring breaks in autumn when something it leaned on changes, and you will not remember how it worked, because you never knew. Keep each tool small enough to rebuild from its description, and keep the description, since in this workflow the prompt is the real source code.

Starting This Week

The method rewards starting stupidly small.

  1. Pick an irritation, and never a project. The best first build removes a splinter, the file you rename weekly, the text you copy between two apps. Projects stall. Splinters get pulled.

  2. Describe the outcome, and let the implementation be the machine’s problem. Say what goes in, what comes out, and what done looks like. When you catch yourself specifying how, stop, since the how is the part you hired out.

  3. Make it explain itself. Before running anything, ask for a plain-language walkthrough of what the code does and what could go wrong. You are checking the shape, and the habit is where your real learning accumulates.

  4. Test where failure is free. Copies of files, throwaway folders, fake data. Only promote the tool to real work after it has bored you with reliability.

  5. Keep the off switch obvious. Know how to stop, undo, and delete what you built. The golem stories made the erasure a ritual equal to the making, and they were right.

The Maker’s Posture

A person who builds tools without fully understanding them has joined an old and cautionary lineage, and what separated the craftsman from the fool in that lineage was posture rather than total knowledge, which nobody in any story possessed. The careful makers stayed curious about what they had made, tested before trusting, kept the limits in view, and never mistook a working golem for a finished understanding.

Vibe coding hands that ancient posture to anyone with an afternoon and an irritation. The tools are small, the stakes are chosen, and the education is real. Describe something into existence this week and watch your relationship with every screen in your life shift a few degrees toward authorship.


Andrej Karpathy coined “vibe coding” in February 2025, and Collins Dictionary named it a 2025 word of the year. His later reframing toward “agentic engineering” dates to talks given in 2026.