AI-supported work as a TYPO3 integrator: What changes with new AI models

Yesterday, on 05.02.2026, Claude Opus 4.6 was released. I haven't tested it extensively yet, but the release got me thinking: What have AI tools actually changed in my everyday TYPO3 life in recent months?

The last few weeks have shown me that Opus actually feels like it delivers better results. For example, I almost always use Opus as a model in Notion AI and am extremely satisfied. If the new version is even better now, especially in terms of programming, it will be exciting.

But that's not what this article is about. It's about the bigger question: What do AI tools mean for us TYPO3 integrators? What works? What doesn't? And should you get involved now?

How I use AI in everyday TYPO3 life

I now use AI tools more and more for TYPO3-specific tasks. For example, creating new content blocks.

Since Claude Code has insight into my project, the tool knows exactly how I create content blocks and what the structure is like. Then all I have to do is say: I need a content block that can do this and that. Within a very short time, I get a basic structure that is already very functional and that I often only have to adapt slightly.

Or TCA extensions. Sometimes complicated manually. A matter of seconds with Claude Code.

For TYPO3 upgrades, I mainly use AI for frontend things: rewriting jQuery in modern vanilla JavaScript, modernizing CSS and things like that. I do the actual upgrade process without AI. I have so much experience there that an AI can't really help me much. And with tools like Rector and Fractor, you don't necessarily need AI.

Context is everything

Especially if you want to extend extensions or core functions, it is extremely helpful if the tool has insight into the original code. Or, of course, the code of your own extensions or site packages. If the overall context is understood, the result will be better. I am convinced of that.

That's a big difference to the past, when we only worked with isolated code snippets.

Where I let AI work independently

With TYPO3 code, I prefer to check step by step what is to be created and confirm every change. Or I take corrective action if I see that things are going in the wrong direction. This is where the issue of knowing what you're doing comes in again. At least to some extent, maybe not 100%, but 90%.

In other use cases, however, I have actually already created Claude agents that work independently in the background. For example, for SEO analyses of Screaming Frog exports. This is analyzed, summarized and formulated in a way that laypeople can understand. These reports are usually sent to end customers. At the end, the data is prepared in a nice HTML page, always with the focus on: How does the customer benefit?

This actually happens automatically. Start Screaming Frog, wait for analysis, export CSV files, start agent, done. My active part is limited to a few clicks. That used to take hours. Now it takes minutes.

Documentation and changelogs

I also have readme files created. I've even built myself a Claude code agent whose sole task is to record relevant changes in the documentation. For example, in readme files or changelog files.

I also have some training documents created with the help of AI, for example based on video transcripts or other sources.

Content production and time savings

The whole topic of content production has become much faster. When I have ideas for a live event or a webinar, everything is produced very quickly. Landing pages, email sequences, social media posts and so on.

I can't give you specific figures, but it's crazy how much more is possible.

Better quality through critical questioning

I have the feeling that my work has improved thanks to AI. Especially with texts, whether blog articles or anything else, I have the whole thing critically scrutinized again. Have I overlooked something? Have I forgotten something?

I usually work very interactively when creating blog articles, just like this one. This often helps me to see or notice things that I might otherwise have overlooked.

And when people say: Yes, but your texts all sound like AI now. I don't see it that way. The lyrics now sound the way I want them to sound. The way I perhaps wanted to write before, but couldn't manage it. Perhaps also due to time constraints.

In recent years, I've worked intensively on topics such as marketing, advertising psychology and copywriting. My texts have changed, of course. I'm no longer the same guy who started building up a small business 10 or 15 years ago. I have evolved. And I only see AI as a support.

The limits: TYPO3 as a niche product

Now comes the important part: what doesn't work?

We're back to the topic of TYPO3 as a niche product and possibly too little training data. If you want to have specific code created for TYPO3, whether Fluid, TypoScript or PHP, it often won't work. AI tools invent things that have never been done before.

A FluidViewHelper is suggested with conviction to solve a problem. And you know full well that this ViewHelper has never existed before.

That's still a problem. But I think it's getting better, especially if you provide context. If I now have a specific requirement, I would first send the link to the documentation or several links and say: Here, read through the documentation. Then I need this and that.

The danger for the inexperienced

I think it's dangerous, and I see it all the time: People just throw their TYPO3 problems or requirements into ChatGPT. As a rule, all that comes out is crap. Perhaps also because too little context was provided.

In general, ChatGPT would not be my first choice when it comes to programming.

How you recognize incorrect output

I recognize it through my experience and because I am always very close to the developments due to my work on the Education Committee. I actively follow the change logs, read through all the things that are relevant for integrators and sometimes even test them myself. The whole thing is not just a job, but also a hobby for me.

It really is difficult for people with less experience. I would really recommend it: If something like a ViewHelper is suggested, don't just copy-paste it, but look directly at the documentation. In the ViewHelper reference.

Does this ViewHelper exist? And if so, are the parameters suggested by the AI correct? That may also be the case: The ViewHelper itself may exist, but completely wrong, non-existent parameters are suggested.

You have to actively look into the documentation yourself and try to find out: Is what is being recommended to me even correct?

Data protection and API keys

As a rule, no personal or private data ends up in the AI tools in my projects. This is not even necessary in normal TYPO3 projects because you work in the code. There is no personal data to be found there.

The tools also don't need to access a project's database, where personal data might be stored. You have to be careful that you don't just blindly give access to everything.

The dilemma with API keys: If you give AI tools access to your codebase, they can also read your .env files with API keys. And if you want to test API connections, the real credentials often have to be there.

What I do: I am aware of the risk and weigh it up. I'm more cautious with sensitive customer projects with critical API access. For my own projects or non-critical development environments, I take the risk.

A possible alternative: locally installed LLMs on your own computer or server. I personally have no recent experience with this. My early attempts some time ago were not convincing because the computer performance and model quality were not yet sufficient. Things could be different now. If data protection is particularly important to you, this could be an option that you should look into.

My advice: For every project, consider how sensitive the data is that the AI could see. You have to decide for yourself where your limits lie and which solution is right for you.

You have to learn to prompt

Of course, you also have to learn how to prompt, although it's getting easier all the time. At the very beginning, when ChatGPT and co. were relatively new, it was very important to prompt correctly. Role assignment, context, task and so on.

I have the feeling that this is no longer quite so necessary nowadays because the language models are getting better and better. Nevertheless, I think it's better to give too much information than too little.

Shit in, shit out. If you give little information, you won't get much good output.

My clear opinion: AI expertise is a must right now

The topic won't be important in a year or two, but now. I am convinced that those who refuse to use AI will eventually fall by the wayside. They will be overtaken by those who use it.

The problem is that it may then be people with less experience who deliver poorer quality. Experienced developers who use AI will always have an advantage over less experienced people. I am convinced of that.

Because you as a developer, as an integrator, are still better able to judge whether the code is really good than perhaps an AI. And code quality will remain important, especially in the long term.

Of course you can code something together with AI now. But whether this is sustainable and will still work in a year or two, especially if functions are to be expanded, is another matter.

I think AI expertise is a must these days. Whether it's coding, content production, whatever. Especially for freelancers who work alone.

Where you should start

It's not difficult to get started, but it depends a bit on what you feel comfortable with.

Personally, I now use Claude Code on the command line. I feel comfortable there. It works. That shouldn't be a problem for developers.

I know there are people who don't like working on the command line. You can also use Claude Code in the desktop app, for example, at least on macOS. I don't know about Windows.

Of course, there are also other agents or programs such as Cursor or the AI integrated in PHPStorm, the AI Assistant, or Junie as an agent. You have to test it out. I've also tested a few things and so far I'm most satisfied with Claude Code because the results are the best for me.

Use cases: Start simple.

Start by perhaps having jQuery rewritten in vanilla JavaScript when you upgrade. That works great. Or start by having modern CSS written for any effects. Little things. That's how I started.

Then at some point maybe TCA or maybe a content block.

You have to think about it: Could using AI deliver better results? Could it make me more productive? Could it be faster?

It doesn't always have to be. If I invest two hours in prompting for a task that I would otherwise have completed in 30 minutes, I'm missing the point.

You have to experiment and simply test things. Then you realize: Does it work for me or no, that was kind of stupid. You have to get to grips with it.

My conclusion

AI tools are already changing the work of TYPO3 integrators. Not in the distant future. Now.

They make you faster. They help you deliver better quality. They give you the ability to do more, even if you work alone.

But they don't replace know-how. Without TYPO3 experience, you won't know if the output is right. Especially with a niche product like TYPO3 with little training data, you need to know what you're doing.

If you deal with it now, you have a head start. If you ignore it, you'll be overtaken at some point. It's as simple as that.

Start small. Test. Experiment. And learn to use the tools for your use cases.