I tried to vibe-code my way to a SaaS… and failed

I completely relate to that feeling, because I was exactly where you are a year ago.

I’m a non-technical founder too (I run my family's grocery business and come from an aspiring pro-photography background). When I started more than a year ago, I hit that exact same wall: endless loops and architectural roadblocks. The "vibe-coding" magic faded because it simply doesn't scale.

But instead of hiring a developer, I changed my entire approach in January when I started building Luminens—a 16-bit browser RAW photo editor running on WebGPU and libraw-wasm. Talk about a nightmare of structural complexity.

I moved away from automated builders (I've tried all of them), switched to VS Code, and started using ChatGPT 5.4 xHigh, which I found incredibly reliable and capable, yes better than Claude. My entire method shifted: every single time I hit a wall, I didn't just ask the AI to "fix it." I asked it to explain exactly what was happening under the hood and then propose a solution.

By doing this repeatedly, day after day, you naturally start to understand the architecture, the backend, and the logic. My real strength became my obsession: if I hit a wall, I simply refuse to stop until I break through it.

The result? The app now smoothly processes a 200MB Hasselblad RAW file in real-time on a modest Ryzen 4000 with integrated graphics. To test it, I opened the same file on Photopea on the exact same machine, and Photopea completely crashed the whole PC.

You are 100% right: AI doesn't replace structural engineering. But if you are obsessive enough to treat the AI as a master professor to learn the architecture yourself rather than just a code generator, a non-developer can genuinely out-optimize major industry software.

Never give up on your capacity to learn. Wish you the best of luck with the rebuild of Xolora! 🙌

Just launched Ligara — a creator platform I built entirely with Lovable in 4 days with zero coding experience. Lovable has been incredible for going from idea to live product fast. Happy to share my experience if anyone's curious.

Wow, thank you for sharing this experience! And loved that you pivoted and it was a happy ending post.

I think there’s some framing that’s needed here. If you started your project 6 months ago that’s very different from the model/orchestration/tool landscape of 2 months ago. So 6 months ago that wall would have been much harder to “scale” 🤷‍♂️🤣

Another fundamental paradox is that agentic infra development is changing so quickly that it’s potentially advantageous to iterate on architecture, “ontology”, schema, and more abstract aspects than go straight to coding, ie it might be better to delay some key decisions than to rush into building. This can set you up to spend less time chasing your tail. For example, I’ve been spending a lot of time thinking about observability and governance tools and that landscape is pretty immature relative to what agents can do. So achieving “scale” depends on exactly what you’re doing. That thing that saves you a month of work might release next month so the entire month you spent sorting out problem X might have been better spent on Y etc. Architecture work and development roadmap refinement can result in serendipitous delays, dead ends that save you from coding useless components, endless debugging loops, etc. Maybe the problem you’re trying to solve improves with the next model upgrade. Maybe you needed to switch models or orchestration platforms. Developers won’t magically know how this looks.

You could need a team of engineers to deploy something “at scale” but depending on who you talk to and their particular expertise you’re going to get varying opinions so it’s still valuable to research with AI and constantly seek more information. I wouldn’t be so quick to defer to a developer as a default solution that can solve all of your problems as complex tools usually require teams, and a single developer isn’t a silver bullet. Learning to use agents with access to calling tools like Jira/ClickUp and GitHub/Bitbucket has been huge for me. I’ve also added Slack to my workflow and these are all necessary prerequisites to managing complex tasks between people and AI.

This may be a stupid question, but what exactly is vibe coding?

Not to discredit your experience, what you hit is real, and a lot of non-technical founders run into it. But I'm a non-technical founder, and I think the tool you used was wrong for the complex application you were trying to build. I used Replit Agent to vibe-code my product (ARKLIS: An AI CFO for SMBs) over the last 8 months, but the unlock wasn't the coding tool itself. It was using Claude alongside it to engineer the prompts I was feeding in. Plain-English idea → Claude turns it into a precise, structured prompt → Replit builds it without spiraling. That second layer saved me from the debugging loops you described, even when the architecture got genuinely complex. The takeaway being in the age of AI, the better prompter will become KING!!

Thank you so much for sharing this. I had big ambitions while building Good Husband as well, but the reality was that I had to scale things down and launch in stages in order to execute sustainably.

I'm not a developer by training, although I have some development knowledge. However, I founded a company in 2014 and hired several developers to create a solution that's still on the market (even though it has evolved, of course). This experience taught me that software isn't just what you see. Business rules, data volume, and consequently, architecture are of paramount importance.

Since then, I've developed a new solution, 100% using Vibe coding, which I developed entirely on my own. We'll be deploying it for two major clients in September.

To avoid taking risks with the architecture, here's the method I used.

To begin, I used the BMAD method, which, while time-consuming in terms of responses and interaction with an AI, allows you to ask the right questions and define the right direction and architecture.

Then, I coded with Claude Opus and, at regular intervals, had the generated code audited by another AI. I started with Gemini and now I'm doing this with GPT5.5.

So, BMAD is my architect, Claude my developer, and GPT5.5 my QA, and so far, it's working.

If this can help, feel free to contact me. I'll be happy to answer any questions.

The architecture complexity wall is exactly it. AI gets you to 70% shockingly fast, then the last 30% is where all the real engineering actually lives: data model decisions, deployment pipelines, auth flows.

The specific area where I see AI-generated code cut corners most dangerously is anything involving third-party auth or credentials. The moment you need to store API keys properly, handle OAuth, or integrate with services that require verified identity, the AI will produce code that looks correct but has real security or reliability gaps. You need to actually understand what's happening there — it's not a place to delegate blindly.

Respect for pivoting to rebuild it properly instead of shipping the fragile version. Where did the complexity start unraveling first — was it the data model, auth, or the deployment infra?

the failure pattern that shows up most often is the language one. you can vibecode the product but you cannot vibecode the explanation. the first version of TAM Network felt similar. it took me 18 months between when i could explain it to myself and when i could explain it to a stranger. once the receipt word landed, the shipping rhythm came back. it was a vocabulary problem all along.