Vibe coding works. I wish it didn't. But it does, well enough. And barring some revolution that overturns the new world disorder, machine learning cannot be undone.

Earlier this year, I surrendered, bought a $20/month Claude subscription, and vibe coded a web app for monitoring news feeds. It's been an enlightening but uncomfortable experience.

There are people who want to stop AI. I wish them well. But I remain convinced that the problem is not AI. It's the people who would use AI to profit while evading responsibility and liability. 

Every AI failure to date follows from some person's decision to implement an AI system without fully understanding what might happen (or they understood and knew they'd get away with it). We have only ourselves to blame for allowing software-driven cars, for accepting the legality of AI code laundering, and for allowing AI systems to dispense bad medical advice. Stopping AI in its present form begins at the ballot box, at least in the US.

Back in 2019, AI attracted attention for producing quirky, weird content. By 2022, it was producing occasionally passable code and lawsuits. By February 2025, AI researcher Andrej Karpathy coined the term "vibe coding," which for a while meant poorly written code coaxed from a machine learning model.

By late 2025, around the release of Anthropic's Opus 4.5 and OpenAI's Codex 5.2, the models had improved to the point that vibe coding was just coding. They produced code that was good enough – it wasn't perfect or optimized or clever or artful. But neither was it laughably bad.

Developers noticed, and the result – apart from a large growth in the number of commits to GitHub – was a slew of testimonials about what people managed to accomplish by using an AI model to work on a particular project.

Simon Willison, a veteran open source developer turned AI influencer, penned a fine example of the genre, "I vibe coded my dream macOS presentation app."

Security researcher Michael Taggart offered a more recent take, "I used AI. It worked. I hated it." Taggart's assessment is spot on, though I didn't end up hating the experience of working with AI. It's complicated.

I can understand how a professional programmer might resent AI's indifference to craft; I feel the same way about AI writing. 

At the same time, people who are not professional writers may be thrilled to have a tool that does something they don't enjoy. I can't condemn them for not sharing my views. 

I don't want to read essays written by AI and it's fair for skilled developers to dismiss apps coded by amateurs. I imagine talented musicians in the late 1970s – people who dedicated their lives to practicing their instrument – felt the same way about punk or rap music as those genres, where instrumental expertise wasn't the focus, became more popular.

Advances in AI won't do away with the need for engineering excellence. There will always be a place for people with high-level technical talent. Those making a living selling website templates and app design services on freelancing platforms won't be so lucky.

What I built and what I learned

I've been a hobbyist coder since the early 1980s when I learned BASIC. I got a bit more serious about programming when the iPhone came out. I wrote some iOS/Android games in a cross-platform Lua-based framework called Corona SDK (now Solar2D) and picked up bits of Python, JavaScript, TypeScript, Dart, and Flutter along the way.

About eight years ago, shortly after joining The Register, I started work on an Electron app called Vulture Feeds to track news stories via RSS/Atom. I ended up rewriting the app under the name RSSputin as I learned more about coding and have been using it since then.

Having covered AI for several years, I decided to see if I could create a hosted, improved version of RSSputin with Claude Code. The first commit landed on February 22, 2026, and seven weeks and 337 commits later, RSScal is now a functioning, commercial app.

The bulk of the code was emitted by Claude Code and manually committed after non-professional and AI review. If the code is truly terrible, it will not be long before the app is brought down. But I'm cautiously optimistic. I'm enjoying using RSScal and some of my colleagues have been banging it around. My sympathy for Linux admins has grown now that I am babysitting a server.

Whether or not RSScal is commercially viable, I can't say. There are older, more established RSS apps, both hosted and local, that compete for the limited set of people who care about site feeds. But the fact that I can spin up a competitor in a month or two at a cost of $40 in subscription fees or ~$200 if measured in token costs, plus $14/month for a small VPS instance for hosting, shows that online chatter about the SaaSpocalypse should be taken seriously. It's never been easier to create software.

Vibe coding hasn't worked out for everyone. Just recently, developer Jim Nielsen wrote up his account of being disappointed with his effort to vibe code his dream RSS app. He ended up with an Electron RSS app that he wasn't very happy with.

"The making from nothing isn't as hard anymore," he wrote. "But everything after that still is. Understanding it. Making it good. Distributing it. Supporting it. Maintaining it. All that stuff."

There's a lot of truth in that. But even with AI help, the coding can't just be assumed or dismissed. If I hadn't already built RSS apps by hand (in as much as using VSCode with autocompletion is "by hand"), I would have had a harder time prompting Claude to generate what I wanted.

Relying heavily on Claude Code is a dependency I'd rather not have. But it has enabled me to use a lot of technologies that I didn't know very well. RSScal runs in Docker containers. Its backend is Python (FastAPI), Celery, Redis, and PostgreSQL (Supabase). Its frontend is SvelteKit and Tailwind CSS. At some point, maybe I'll open source it, though AI has complicated the open source world significantly. In a way, the app already exists within Claude Code – anyone can just call it forth with the right spell.

One of the knocks against relying on AI is that you don't learn anything. But my level of comfort with Docker and Python and SvelteKit has improved significantly. AI absolutely will limit your learning and cause your skills to atrophy if you use it for everything and don't engage. But it can also be a tool that helps you overcome obstacles – I've found Claude far better at assembling complex command line strings than "googling Stack Overflow." 

Both capable and clueless

Working with an AI model like Claude Code is difficult because you have to hold two contradictory thoughts in your head – the model is both highly capable and utterly clueless.

There were situations where I committed a change and something broke and I asked Claude about it. The model would suggest a fix that didn't apply because Claude assumed I was working on the development build rather than the production build or that I was working directly with the database rather than doing so through Docker. Or Claude would implement some feature and fail to include basic security features like rate limiting.

At the same time, Claude sometimes added details or interface elements that I hadn't requested but turned out to be worthwhile. There were "creative" suggestions it made about web design that I kept.

The commodification of basic app creation has been underway for years. As soon as an app becomes popular, people create clones and offer them for sale through various markets like Flippa, Acquire, AppWill, and CodeCanyon. Or maybe they're selling entire e-commerce sites as turnkey businesses for six figures or more. AI will accelerate that commodification but writing code is only part of the picture.

Claude Code doesn't make you a great marketer or ensure that you're at the right place at the right time with the right idea. It doesn't build trust or develop the relationships that businesses depend on. It doesn't make your RSS app a good idea. But it may open doors you'd otherwise have passed by. ®