The Past and Future of Vibe Coding

Thoughts from 2 years of yelling at digital gods

Before we dive in, a quick update that I’ve finished filming the complete overhaul to my extremely popular vibe coding course Build Your Own Apps, and the new videos will be dropping next week!

The term “vibe coding” wasn’t coined until February 2025, but I started in January 2024. Cursor had launched a few months earlier, and tipped off by Dan Shipper and a couple other friends I decided to give it a go.

It was slow, clunky, prone to errors, but even then the promise was clear. As I reviewed and implemented suggestions from Cursor’s chat over the following months, I managed to build a mostly-usable “Strava for writing” mobile app, while writing less than 20% of the code myself.

As someone who has dabbled in coding on and off over the last decade, it felt revolutionary. It was the missing piece for anyone like me who never really learned how to build software but could kinda hack things together given enough time and patience. But whereas in the past I’d often hit some gnarly issue I couldn’t resolve by digging through StackOverflow, suddenly I had an infinitely patient tutor and assistant.

Even so, it was slow going. I wrote the first lines of code for the app on January 3rd, but didn’t release it to the public until June. There were weeks where I got stuck in debugging loops that Cursor couldn’t resolve, and often had to throw out entire days of work because the code was a mess.

Despite eventually getting the app live, I ended up abandoning it, partially because vibe coding just wasn’t good enough to really do the heavy lifting needed to keep building a production app and because I wanted to focus on writing Husk.

I suspected at the time, though, that if I was patient, the tools would get good enough to let me do the kind of work they weren’t yet ready for.

November 2024

About six months later I dived in again. Claude’s Sonnet 3.5 was live, Bolt and Lovable and Replit were blowing up, and Cursor had added “agent mode” that could implement the code changes for you and run autonomously for far longer than before. I started hacking on some tools to help with my book editing process, having already found a lot of value in using AI as a first-round editor, and was shocked by how much easier it was to build software compared to the start of the year.

By January 2025 I was hooked and could feel how much potential there was, which led to creating Build Your Own Apps. I’m still somewhat shocked by how strong the response was to this course, and it was amazing to see people who had never coded before build their own tools and push them live.

But the truth is vibe coding still wasn’t quite where I wanted it to be yet in January. It got stuck. It created spaghetti. Building a production-ready app still required a decent amount of technical knowledge, especially once you got to making it secure and performant.

Even though I had spent so much time mastering vibe coding, I still wasn’t using it to create software for anyone besides myself. The personal tools I could hack together were useful. But I was running into limits in what the tools could do that I didn’t have the bandwidth or patience to push through.

So once again, I stepped back for a few months, more confident than before that the tools would soon get good enough to let me do the kind of work I aspired towards. I published Husk, wrote The Birth of Paradise (publishing ~December), got to work on Husk’s sequel, and went back to waiting.

June 2025

When my third daughter was born in June, I took a couple months off writing to be home with her and Cosette, and used the bit of free time I had during naps to get back up to speed on AI tools and vibe coding.

And as excited as I was in January 2024 and January 2025, this summer is when I really felt “oh yeah this is it.”

Whereas January 2024 was akin to hiring an off-shore $5/hr developer to try to hack something together for you, and January 2025 was like having a decent junior developer, now it feels like you actually have a team of rather capable engineers who can build whatever you want for you. If you know how to work with them.

I started hacking on Covici for myself. I wanted a book writing tool that checked all the boxes I had been looking for but couldn’t find, and the coding tools just… kept working. I didn’t expect it to turn into something I’d actually launch for other people to use, but it reached that quality level after a couple months of hacking and I was having so much fun building it I figured hey why not release it?

But the real “whoa” moment came when I started recording the update to Build Your Own Apps a couple weeks ago. Aspects of the course that took a few hours across multiple videos in January got done in minutes in a single video.

More complicated integrations like billing customers with Stripe, which I had saved for the very end of the course last time, were suddenly painless so I taught them much sooner. What took 10-20 hours of instruction before could be covered in 5, leaving time to go much deeper.

Whereas vibe coding at the start of this year could build you a workable but minimal viable product, now it can actually take you all the way to a production-ready app. I didn’t expect the tech to get this good this fast, but it has. Even OpenAI says that their own tool, Codex, is writing 70% of their code now.

There are still challenges, of course. You cannot one-shot anything workable. You have to learn how to plan, debug, ask good questions, which tools to use, etc. But the dream of having a professional development team available 24/7 for 1/100th the cost is basically here. And it’s only getting better and cheaper.

The Future & What’s Worth Learning

I’m skeptical that we’ll ever have a world of “personal software” where you tell an agent to whip you up a SaaS for your business and it just seamlessly works the way you want it forever.

Not because that won’t be possible, the tools can already probably build you an MVP, but you won’t want to maintain it. Good software requires maintenance and ongoing attention, which is why most people will still prefer to pay to use something someone else is maintaining.

There will be exceptions, of course. I suspect per-seat pricing for a tool like Slack is going to be hard to defend. But most writers will not make their own writing app. Most video editors will not make their own video editing app. Most entrepreneurs will not roll their own tax app.

The question is no longer what do you want to build but rather what are you willing to maintain?

With enough patience and a few thousand dollars of AI credits you can build probably 90% of software you might want right now. But are you willing to fix bugs in the middle of the night? Resolve Stripe disputes?

Sure some of that will be taken over by AI agents but there will always be some degree of maintenance you need to do, simply because your attention and energy are the only counterforce you have against the natural entropy that decays anything manmade.

But along with the question of what you’re willing to maintain, as the barriers to creation drop the value of taste will only increase. It’s always been important, but your ability to differentiate from all of the other AI-generated software will be your ability to bring your unique taste and vision to it.

Someday AI’s obsession with purple gradients will be an amusing memory, instead of an annoying habit you have to constantly fight against

I’ve started learning design and Figma primarily because I can sense that my ability to design is now a bigger bottleneck than my ability to code. You can get pretty far telling AI what you want a site or app to look like, but you are still bottlenecked by its limited creativity.

If you do have any dreams of creating software, this is an incredible time to dive in. Even if you were frustrated and roadblocked by the vibe coding tools a few months ago, you’ll be surprised by how much easier it is today. Even I was surprised by how much it’s changed, and I’m doing this every day.

It might not feel worth it to learn considering it’s going to keep getting better and better. But the sooner you learn how to use the tools and find their limits and work around them, the more prepared you’ll be for the next improvement. And I think to some extent, agent orchestration is what a lot of digital knowledge work is going to look like in the next five to ten years.

We’re also at an interesting point in time where if you start building something now, by the time you get to a level of complexity where you’re hitting the limits of the tools there will already be new ones out. You won’t have to take six month breaks like I did. I’ve already experienced this by having bugs in Covici from August solved by Codex and Sonnet 4.5 in September.

Obviously Build Your Own Apps is a great way to get started, especially with the new material coming out next week. But even short of that just download Cursor and start trying to build something.

You might be surprised by how far you can get.