Do we still need code generators when AI is programming?

Table of Contents

• Two Ways to Generate Code

• An Example

• 1. Reliability

• 2. Speed

• 3. Verifiability

• And What About Vibe-Coding?

• Final Words

"Do you actually still need your code generators now that AI is programming?" I hear this question regularly: from former colleagues, in job interviews, in customer meetings.

It's obvious. We have been building code generators since the founding of wunschlösung because we don't want to waste time on things that can be automated. Today, everyone generates code automatically, just differently than we did back then.
By the end of 2022, I described my first attempts with ChatGPT here. At that time, we were considering how we could enhance our generators with AI. It turned out differently: Today, AI uses our generators. And it works so well that my answer is now: Yes, we need them. Maybe they are more useful than ever.

Two Ways to Generate Code

Briefly to the classification, because "generate code" now means two very different things:
Our classic generators read a "model": a compact description of what exists in an application. For example, products, dealers, and orders. From this, they generate large parts of the application: data management, interfaces, the framework for business logic, and the back-office interfaces.

To distinguish, we call them "static" generators. They are not called static because they do not evolve—we are constantly working on them. Rather, their result is fixed: same model in, same code out. Every time.

AI works differently. It can basically build anything you can describe to it. But it also builds it a little differently each time.
Today we work with coding agents that run in isolated sandbox environments and operate our generators within them. The agent modifies the model, starts the generator, and only writes by hand what the generator cannot.

An Example

Let's take a marketplace. In the future, buyers should be able to rate sellers, but only if they have actually bought something there. A coding agent gets precisely this request as an assignment. What happens next:

The agent adds the rating to the model: stars, text, buyer, seller. That's a few lines.

He starts the generator. Seconds later, the database, interfaces, and the administration in the back office are ready.

The rule "only after a real purchase" is written by the agent himself in a place reserved exactly for such extensions. Likewise, the display in the shop.

In the review, we look at the model, the rule, and the display. We already know the rest because it looks the same in each of our projects.
A few sentences in, a finished feature out, and we know exactly which places we need to check. We don't know a better way to build platforms. This has three reasons:

1. Reliability

AI-generated code is maximally flexible. You can wish for anything and get a solution for it. Most of the time, it's good, but not always.

Many people don't even look at the code anymore and only check if the result does what it's supposed to. For a prototype, that's perfectly fine. But for a platform that goes live for real customers, that's not enough. Then there needs to be good reasons to trust the code. And creating those reasons is the real effort: What takes time today is quality assurance, not implementation anymore.

Statically generated code brings reliability and security with it. It is created according to fixed rules that we have developed test-driven and tested in many projects over the years. If a bug does appear, we fix it once in the generator, and it’s gone in all projects.
The combination is the real trick: a foundation that we know inside and out, and on top of that, full flexibility. The generator can't implement your feature? Then let the AI build it. But it doesn't start from scratch, rather within a codebase where the same patterns apply everywhere.

2. Speed

AI writes code much faster than humans. Larger features in large projects still take hours, and with testing, feedback, and corrections, sometimes days.

Our generators need only seconds for the basic framework of a complete application, even with large models. After all these years, I still enjoy watching them. By the way, they don't consume any tokens.

To be honest: Building and maintaining the generators also takes time. But it pays off for us because we've been building platforms and marketplaces for years, and every improvement benefits all projects.

When implementation is no longer the bottleneck, another question comes to the forefront: What functionalities does your platform actually need? This question decides more than any individual line of code. That's why we invest the time we've gained right here.

3. Verifiability

For laypeople, AI-generated code is a black box. Professionals can read it, but they are overwhelmed by the volume. Often the result appears to work, but is it safe and scalable? Can I go live with it? The real difficulty is figuring out where to look and what to test.
Here, separation helps: We know for every file whether it comes from the generator. We don't need to review this part line by line; instead, we check the model from which it originates. What’s left is the handwritten part, whether by human or AI. That's what we focus on.

How much does this matter? In our largest marketplace project, around 90% of the lines of code in the backend and back office are generated. What is developed individually is mainly what buyers see. That’s where a marketplace should not look like every other.

We know where we need to look. We continuously adjust how exactly, because what AI can reliably do is shifting month by month. We keep a particularly close eye on the critical points, and at each, we can intervene ourselves. You get the speed of AI and, in addition, what no vibe-coding tool provides: a team that really knows your software and ensures that it works and scales.

And what about Vibe-Coding?

Does this mean you should stay away from Vibe-Coding? On the contrary. Feel free to build your own prototypes! There is no faster way to find out if an idea works.
It gets exciting when the prototype shows that your idea has potential. Then it is the best starting point we could wish for the project: It shows better than any concept paper what you intend to do. Are you interested in a prototype but don't have the option or time to prepare one? No problem. We also create prototypes for you.

From here, we take you by the hand. We establish with you what the platform needs for real operation, build it so that we can take responsibility for it, and are happy to take care of operation and scaling afterward. The goal is always the same: We make your platform successful together with you.

Do you have such a prototype and wonder how it becomes a platform? Get in touch. And if you don't have one yet, we are happy to build it together.

Final Words

Generators and AI are complementary tools. One delivers the same thing every time, while the other delivers something new each time. Combined, we can launch platforms more affordably and quickly, without sacrificing quality.

In 2022, I always politely asked ChatGPT for everything, hoping that the AI remains nice to us later. As of today, it works with our tools and according to our rules. I continue to say please and thank you, just to be safe.