Vibe Coding for Fun and (No) Profit
I vibe coded a Spanish driving course in four hours. It earns nothing, and it is production quality. The checks are why.
By Geordie Everitt
The term "Vibe Coding" had a short shelf life between the time that Andrej Karpathy coined it and the time it became a term of derision in the same vein as "script kiddie" did back in the olden days.
I recently came to the stark realization that I need to get my Spanish driver's license. Fast. Oops.
The last time I took a driver's test was in my native language in 1981 in a Mercury Capri and automatic transmission.
I failed the written test the first time because of hubris of a bright teenager who had watched his dad drive for 15 years. How hard could it be?
You may be shocked to learn that it is a bureaucratic hassle to get a driver's license in Spain. Shocked, I tell you.
So, I did the sensible thing and went down to the local driving school and plunked down €200 for a book and the prospect of getting lessons when it is convenient for them.
After looking at how terrible the book is, I did the next sensible thing and vibe coded an app to help me study.
What Claude Code came up with in 4 hours of prompting was amazing. It is production quality.
It is fully vetted, and up-to-date with recent changes to Spanish law.
It is scalable to the world and supports thousands of users. It is better than every app on the app store that claims to do the same thing.
I'm probably the only person who will ever use it.
It's called Glorieta: Guiri al Volante. Roundabout: Foreigner at the Wheel. Where an American would put a stoplight, Spain usually puts a roundabout, and I find that civilized. A guiri is what Spaniards call a foreigner, usually a Brit. Americans count.
The app has 14 modules, 65 lessons, 43 diagrams, a 30-question mock exam and a gallery of 128 traffic signs. Every string flips between Spanish and English, and Spanish is the default. Its sister app, LinguaMama, teaches the Spanish itself.
No profit, worth building anyway
The AI Triage sorts work on two things: how ready it is to hand to AI, and what it is worth. The work here was drafting a course from a regulation, and it is very ready. The material is text, the steps are written down, and a mistake is cheap to undo. It is worth little. It brings in no money, and a wrong lesson costs a learner an exam fee.
The Triage also sorts work by kind. Glorieta is tactical work: it answers an outside event, my need for a Spanish license, and not a process that runs every week. Its rule for a ready project with a small prize is to do it only if it costs almost nothing. It cost one morning.
I ask a second question, borrowed from Adam Carolla: does it make me money, and does it make me happy? Glorieta makes no money. It makes me happy, because learning new things and solving problems is enough for me. That is not the best use of my resources, and it is worth it.
No AI inside
AI is in how Glorieta was made, not in what it does. The app makes no model calls: no prompts, no completions, nothing generated while you study. A model drafted the lessons and questions, they were checked, and they shipped as fixed content. Using AI to make a thing is a different decision from putting AI in the thing. Here a mistake is caught before anyone sees it, and nothing can go wrong at runtime.
The lessons are written from the law itself. Spain publishes its consolidated legislation as open data in the BOE, the Boletín Oficial del Estado. I had Claude write a small script that prints only the wording in force today and flags repealed text loudly. Every legal rule cites its article.
Who checks the result
The first commit's own message says the question bank was "pending an independent fact-check." Seven minutes later, the fact-check landed. Two explanations were wrong, one about the color of road signs and one about yellow lines. A rule that had come into force on October 1, two weeks earlier, was missing: a 20 km/h slowdown when overtaking a cyclist outside town. The sign set included two speed limits, 130 and 140 km/h, that could not be confirmed as real signs. They looked plausible, which is how a model gets things past you. They came out.
Then a second pass read every module, by reviewers who had not seen the first author's reasoning. They found errors in every one: rules stated without their conditions, claims lifted from repealed articles, first-aid details without a source, and a sign rule that changed in the 2025 sign catalog.
The repealed articles are the dangerous category. In June 2025, Spain repealed the old articles on traffic signs and markings. The article numbers are still there, and some of them now hold different law. A citation to one of those numbers can look right and point at something else. So the check got a rule. Before every deploy, a script goes through every lesson and fails on any citation of a repealed article.
The facts page labels each fact: official source, secondary source, or verify. A new video question that the DGT, Spain's traffic authority, announced in October 2025 is marked Comprobar, verify, because its exam page doesn't mention it.
Same workflow, real stakes
No publisher was going to build this for me. The market is English speakers studying the Spanish driving test, which is me and a few people I'll never meet. Software for an audience that small used to cost more than it was worth. It costs a Saturday morning now, provided someone checks it.
A rulebook your people train on is the same workflow with real stakes, just as ready and worth far more. Make the AI fetch the current text and cite the clause. Have a second reader, one who hasn't seen the first one's reasoning, check the result. Then put a check in the path that fails when a clause has been repealed. The first gets you a draft, the second finds what's wrong with it, and the third keeps it from going stale.
The AI Triage scores your own processes on the same two questions, ready and worth. It is free.