The Age of Personal Software

Daniel and I have known each other for over fifteen years. In 2013, we co-founded a startup together — Gerabo. Months of unpaid work. It was fun, it was a great learning experience for my younger self, but I couldn’t pay the rent with fun. At some point I parted ways to earn a salary. The friendship survived. Daniel kept having ideas.

Every few months, a new concept — something in the food delivery space, a B2B tool, a consumer app. Some of them were genuinely good. My answer was always some version of “no.”

Not because Daniel was wrong about the market. Because the math never worked. I’d done the startup thing — I knew what “building something real” actually costs. Two to three months of heads-down development, probably a hired contractor or two, at minimum €20K if you were cutting corners. Full-time job, young family, a backlog of actual responsibilities — I was not about to gamble that on something with a 90% chance of failure again. The ideas stayed ideas.

That math has changed completely. And the change is bigger than I initially understood.

What The Old Economics Actually Looked Like

Twenty years of software engineering means I can build backends in my sleep. Event-driven systems, REST APIs, database schemas, CI/CD — none of it is hard for me anymore. But a working product is more than a backend.

Frontend design. Responsive CSS. Typography. Color systems. PageSpeed. SEO structure. Deployment pipelines for frameworks I haven’t touched in two years. These aren’t areas where I’m bad — they’re areas where I couldn’t meet my own standards without significant investment of time I didn’t have. A backend I could build in a weekend. A complete product I was proud of? Six to eight weeks minimum, and it would still have rough edges where my skills thinned out.

So the economics looked like this: weeks of my time (opportunity cost: real), technical gaps that either stay rough or get filled by paid specialists (cost: real), for an idea that probably fails (probability: high). That’s not a calculation you run and decide to proceed. That’s a calculation you run and order a coffee instead.

What Changed

Claude Code, specifically — not as a code generator but as something closer to a technical co-founder who never sleeps and doesn’t bill by the hour.

The gaps in my skillset stopped being gaps. Frontend design I wasn’t happy with? I’d describe what I wanted and iterate until it was right. Framework selection I wasn’t sure about? Real analysis of tradeoffs, current ecosystem state, with the right caveats about what the model doesn’t know. CSS that would have taken me forty minutes of Stack Overflow to write from memory? Done while I was thinking about the next thing.

The critical insight, and it took me a while to internalize it: the quality is high not because the AI is magic, but because I know what good looks like. Twenty years of experience didn’t stop mattering — it started mattering at a different leverage point. Not “write this code” but “that component structure will cause problems when we add pagination” and “this CSS approach will break on Safari” and “no, try again, the spacing is wrong.” The guidance loop I can run in real-time is what separates the output from slop.

The Evidence

CodeWithAgents.de — this site — is the most recent data point. 37 commits. One session. A bilingual website with a blog, pagination, per-language routing, SEO infrastructure including JSON-LD structured data and an auto-generated sitemap, and a custom markdown content system. The kind of thing that previously required a frontend developer, a designer, and someone who understood the operational side. now PageSpeed 100 after multiple migrations. Zero runtime JS bundles for content.

Before that: PowerBen, the presentation platform for my agentic engineering training. Empty repository to deployed mobile-first presentation in one session. Three context windows, animated slides, a custom subdomain, GitHub Actions deployment pipeline. The footer reads “Built in one session: human intent, AI hands, zero copy-paste.” That wasn’t branding. That was documentation.

And before that: a receipt scanner for a side project. Backend, frontend, OCR pipeline with image processing, GDPR pages, survey design. Three development phases, five critical bugs, one number at the end: 97.78% OCR confidence on a real receipt photograph taken under kitchen lighting. Solo. No contractor. No designer. No DevOps person.

None of these are “helped by AI.” They’re built with AI, directed by me.

The stoliar-immobilien Story

The most interesting data point isn’t any of those. It’s a friend — not a developer — who now maintains his own professional real estate website.

I set it up using KIRO (similar to Cursor). Got the base running, showed him how it worked, handed it over. He makes changes himself. Updates content, adjusts layouts, adds listings. No engineering background. No ongoing help from me. The tooling is approachable enough that someone who has never shipped production code can maintain a professional website because the AI handles the technical translation from “I want this to look like that” to actual working code.

If a non-developer can maintain a professional website independently, the implication for someone with twenty years of engineering experience is significant. The ceiling isn’t gone — it went up.

The New Math for Daniel

The reason Daniel’s ideas were always rejected wasn’t that they were bad. The cost of validating them was disproportionate to the risk. Building a prototype used to mean weeks and specialists. Now it means days of focused work with AI.

And there’s a feedback loop I didn’t expect: every side project makes me better at the day job. The 18-agent orchestration system I built at Kleinanzeigen shares architectural thinking with the side projects. The patterns I tested on personal projects, fast and cheap, became the judgment I applied to production systems with real consequences. The learning doesn’t stay siloed.

The risk calculation fundamentally changed. The downside of a failed prototype is now days of effort instead of months. The upside — validated idea, working product, or at minimum concrete lessons — stayed the same. That’s a different bet entirely.

The Three-Layer Reaction

When people see CodeWithAgents.de, they notice the design quality, the PageSpeed score, the multilingual blog. First reaction: impressed.

Then: “AI built this?”

The impression drops slightly for some people. Which is fine — I understand the reflex. Then a subset of those people try to replicate it. They run Claude Code or Cursor and produce something that… doesn’t look like this. The design is off. The architecture has problems. The code works but the result feels generic.

That’s the part that clarifies what’s actually happening. The quality isn’t coming from the AI. The AI is fast and capable and I couldn’t match its execution speed. But the quality comes from the feedback loop — from knowing that a component structure is wrong before it becomes a maintenance problem, from recognizing when a CSS approach will break, from insisting on a third iteration of an SVG because the proportions are still off. That judgment is what prevents ticking time bombs from shipping. It’s twenty years of accumulated pattern recognition applied at a different leverage point.

The AI is the execution engine. The experience is the steering.

One Experienced Developer Is Enough

The age of personal software is not coming. It’s here. Custom tools, exactly fitted to specific needs, at a fraction of what a team would cost. Products that used to require a frontend dev, a backend dev, a designer, and a DevOps engineer can now be built and maintained by one experienced developer who knows what good looks like.

That’s what I do — for my own projects, and increasingly for others. One person with the judgment to direct AI at speed, the experience to catch what the AI misses, and the standards to keep iterating until the result is right.

Daniel and I are building something. I’ll write about it when it ships.