preloader
image

Two lines of code in a year – and still five projects shipped

I have been a software developer for more than 15 years. For a long time, writing code was what I did all day – and honestly, what I enjoyed most.

This year I wrote, roughly speaking, two lines of code myself. And only because I already had the changes in my head – the AI would actually have taken longer for those.

And yet I shipped more than I have in a long time: at least five projects, from my own website to small tools that save me work every day. AI did the rest – more precisely, I worked with Claude the way I would work with a very fast, very diligent colleague.

What I built this year

  • Immofolio – a dashboard for private landlords and property investors: returns and cash flow at a glance, AI-assisted utility statements, depreciation and tax exports, purchase analysis, tenant and document management, plus a mobile app.
  • Pachtbuch – an application for private lessors of farmland, grassland and forest: contracts, deadlines and lease payments in one place, plots on a map, uploaded contracts read automatically. The waiting list is just opening.
  • MyBicycle.parts – keeps track of the wear parts on your bikes, with Strava integration and a reminder when a part is due. I rebuilt the project completely this year.
  • This website – switched to German, tracking removed, build errors fixed – and a completely new website designed in parallel.
  • A land search pipeline – a tool that searches classified ads for suitable plots, rates them against my criteria and writes the matches to a Kanban board every morning.
  • Plus several client projects I unfortunately can’t talk about publicly.

What this looks like in practice

I describe what I want – the way I would explain it to a developer on my team. The AI looks at the existing code, asks questions, makes a plan, implements it and tests as far as possible. I follow along, decide and approve.

And I don’t even type anymore: I simply talk to Claude, the way I would talk to a colleague at the next desk. That is much faster than writing – I put my thoughts into words as they come, and the AI turns them into concrete tasks.

An example from this website: “German should be switched back on.” That turned into a new branch, fixed templates, translated content, a test build and a clean commit. My job was to look at the result and merge it.

So my work has shifted:

  • from typing to talking: if you can say clearly what you want, you get good results – quite literally.
  • from implementing to reviewing: I read diffs instead of staring at empty editors.
  • from how to what and why: the interesting questions are now architecture, priorities and product.

Example Pachtbuch: Claude plans, ralphex builds

With Pachtbuch I went one step further. First, I worked out a detailed plan with Claude: which features does the app need, what does the data model look like, how can it be broken down into tasks – and how do you check that each task is really done?

The implementation was then handled by ralphex, an open-source tool that works through exactly this kind of plan on its own. It takes on one task after another, each time in a fresh session so the AI stays focused. After each task the checks run, then it commits. At the end, several review agents look at the code: quality, implementation, tests.

The result: the complete app. My work went into the plan – and into reviewing the result. With my Claude Pro plan, the implementation took two days – mainly because my five-hour usage allowance was used up after about 30 minutes each time.

I did the same with MyBicycle.parts: plan with Claude, implementation with ralphex – and the entire project was rebuilt from scratch.

Where it still falls short

Sounds too good to be true? A few things don’t happen on their own – and that is exactly where experience matters:

  • The AI doesn’t know your environment. Our old website built perfectly locally, yet the deployment failed. The reason: the hosting service uses a years-old Hugo version that couldn’t read a newer file format. Found and fixed quickly – but only because someone knew where to look.
  • Small mistakes happen. Once, an image file slipped into a commit where it didn’t belong. Spotted, fixed, no harm done. But without a review, it would have ended up in the repository.
  • Not everything can be tested. When the AI can’t install packages in its environment, it writes code “blind”. The first real test then runs on my machine.
  • Responsibility stays with humans. Legal texts, prices, what is said publicly about clients – I decide that, not the AI.

My takeaway: AI doesn’t replace knowing what good software looks like. It replaces the typing.

What this means for you

This is exactly how I work for my clients today – whether it’s a website, automation or custom software. AI makes it fast, I make it good: every change crosses my desk before it goes live. For you, that means faster results and lower costs, without compromising on quality.

Want to know what this could do for your business? Just get in touch.