The Three Little Pigs always cracks me up. People who know nothing about something keep telling the wrong story about it. I've built my own straw-bale house, and it's more comfortable and more pleasant than the cinder-block houses I've lived in.
Same thing with vibe coding.
TLDR: Everyone is burying vibe coding in April 2026. Karpathy rebranded it as "agentic engineering." 70% of builds stall at the demo, a number now repeated everywhere without anyone bothering to source it. The going diagnosis: the method is bad. I had to lay 1,880 bales of straw on my own house to figure out that everyone got the diagnosis wrong.
When I told my neighbors I was building my house in straw, I got the exact same monologue I get today about vibe coding. It burns. It won't hold. You're naive. Five years later the house is standing, and the same voices have moved one decade over to AI-generated code. Same sentences, different material. And the same reasoning error behind them.
What People Said When They Saw the Bales Going Up
A neighbor stopped his truck on the road, looked at the bales stacked under the tarp, and asked me with no warning if I knew that mice eat straw and that fire eats straw faster. He wasn't malicious. He was certain. Same energy as every commenter on this site who's certain that AI cannot ship.
It burns. It won't hold. You're naive.
I gave the same answer I keep giving today on Medium comments. This is not a faith argument, it's physics. Compressed straw is dense enough that fire cannot find oxygen inside the bale. The lime-and-clay plaster wrapping the wall seals what little air remains. The wood frame carries the load. A bale laid wrong burns. A bale laid right outlives you.
If you want a date: the first European straw-bale house, the Maison Feuillette, was built in 1920 in Montargis, France. Still standing. Still inhabited. Older than my grandfather and in better shape than his last apartment.
The house is fine. The method is fine. What's missing in the conversation is the person who has actually built one.
Five years later I hear the same script about vibe coding.
What People Say When I Tell Them I Vibe Code in 2026
It's dead. It doesn't ship. The serious people moved on to "agentic engineering" (same thing, longer name). 70% of vibe-coded apps stall at the demo, an industry stat now appearing in every think-piece without ever pointing back to the survey it came from. There's a viral Medium article currently telling readers it's all over. Bloomberg ran a piece blaming AI coding tools for a productivity panic, diagnosing the wrong disease entirely.
It burns. It won't hold. You're naive.
Same script, ten years later, applied to JSX instead of straw. And the answer is the same. This isn't a faith argument either. It's a method-and-reps argument.
Vibe coding done badly snaps in half. That part is true. The 70% number isn't pulled out of thin air. People do hit the wall. Plenty of dead repos on GitHub prove it.
Done badly means done by someone who has typed three prompts in their life and expected a finished SaaS at the end. Someone who has never specified a feature in writing. Someone who has never seen what an unbroken loop of generate-test-fix-test looks like across twelve iterations on the same project.
The method matters. The method without reps is a piece of paper.
That part is settled.
Three Years Reading. Two Years Laying 1,880 Bales. The Wall Still Stands.
I read on straw-bale construction since three years before touching a bale. I'm not proud of that. I would have learned faster by laying ten of them badly. But I had a kid, a job, a budget that wasn't ready, and I needed the theory to settle before I could justify buying the land. Three years of books, weekend workshops, two months at a friend's site in Ardèche where I mostly carried things and watched. Long stretches without laying anything. I never quit. I just couldn't start.
Then I bought the land and started.
Two years on site. 1,880 bales between the load-bearing walls and the partitions. The first bale went in crooked and I had to redo it twice. The fiftieth was square the first try. Around the six-hundredth I noticed I had stopped sweating during the plaster pass. Around the twelve-hundredth my hands knew the cut angle without looking. The 1,880th bale went in the way the first one should have, except by then I wasn't even thinking about it.
The method I used at bale 1 and the method I used at bale 1,880 was the same method. The book I read in 2018 didn't change. The video I watched in 2019 didn't change. What changed: my hands had done it 1,880 times on the same house.
This is the part nobody who tells you vibe coding is dead has ever lived through.
Your first feature is crooked. Your fiftieth is square. The first time the model generates code you don't immediately want to throw away, you've already shipped more than you remember. The first time you stop second-guessing the architecture, you've stopped counting. The first time you ship a feature in two hours that used to take two days, you don't notice it happening. Somebody else points it out. 😅
The method does not change between rep 1 and rep 1,880. Your hands change.
But nobody has five years to ship a SaaS. That's the actual problem.
How to Do It in 12 Reps Instead of 1,880
I wrote a book to compress that curve.
Not a theory book. There's already a good one for that. Gene Kim and Steve Yegge published Vibe Coding: Building Production-Grade Software this year and it's the right reference if you are a senior dev who wants the patterns formalized. Read it after.
This one is different. It walks the reader through 12 reps on the exact same project. A small CRM for tradespeople, plumbers, electricians, carpenters. Not 12 different tutorials on 12 unrelated subjects. Twelve passes on the same codebase. Each chapter takes the CRM further. Auth in chapter one. CRUD in two. Search in three. Notifications in four. And so on, in the same sequence you'd build a house: foundation, frame, roof, walls, plaster, finishings.
The method itself is in the book, the eight-step Blueprint Method, the same one I run on every project I ship. It fits in maybe twenty pages. The other 270 pages are reps. Because the method without reps is the piece of paper I just talked about.
Caveat I'm not going to soften: this only works if you do all twelve reps. Doing four chapters and quitting will not ship anything. You'll have learned something, but you won't have built the muscle. (Same as laying fifty bales and stopping. The house is still a hole in the ground.)
There's a private companion repo for readers, where the CRM state is committed at the end of every chapter. If you skip a chapter or get stuck, you clone the snapshot and keep going. I learned this trick from straw too: every workshop I ever attended ended a phase with a wall everybody could touch. You don't move on from theory. You move on from a wall.
If you're already shipping, you don't need this book. Go further with the prompt contracts framework I built after enough disasters. That's the next layer. Vibe Coding, For Real: From Demo to Live App is for the foundation. Prompt Contracts is for the upper floors.
The book is on Amazon: https://amzn.eu/d/04X9k88d
Most people who tell you vibe coding doesn't work wrote one feature, watched Lovable spit out broken JSX, and closed the tab. One bale. One wall. Lazy conclusion.
With method, you build straw houses. With method, you build vibe-coded apps. Solid. Comfortable. Built to last. 🏠
Sources
- Vibe Coding: Building Production-Grade Software by Gene Kim and Steve Yegge (IT Revolution, 2026)
- Maison Feuillette, Montargis, France (1920)
- Vibe Coding, For Real: From Demo to Live App (April 2026): https://amzn.eu/d/04X9k88d
Top comments (0)