Where the Neos community draws its lines when AI codes
In the first Neos Community Crosspollination Call, we had just been talking about metadata for assets and the next Neos release when Maya brought a question from her day-to-day PM work into the round. One client had wanted to know in a pitch how much of the code in their project came from the agency and how much from AI. Another client had asked for guidelines, and whether her data ends up in some model’s training. Elsewhere, a client simply wanted to know whether things would get cheaper now. That turned into a question for all of us, thirteen people on a Friday afternoon: how we handle this, and how much of our code we still type ourselves.
The shortest answer: “I generate nothing.” The answer at the other end described a project where almost nothing is written by hand anymore: only specifications get written, and the generated code is strictly reviewed. It was exactly this reviewing that prompted the line “that’s roughly how I imagine hell”. In between lay local models for everything that touches client data, and lines of code that need to be understood before they’re allowed into the repository.
The spectrum in our community is extremely broad. Gerhard and I kept talking after the call, and I think it’s the same for all of us: what AI can do deeply affects our role as software developers. My writing career, on the other hand, is benefiting.
From “I generate nothing” to a project with almost no code written by hand
The no came from Marc Henry, without hesitation. The economics and the politics behind the big models don’t square with his own ethics. Ethically trained models would be interesting, but they don’t exist in any practical form yet.
Next to that stood careful partial use: type most of it yourself, use AI only in small steps. Especially on Neos topics, we heard, quality fluctuates; the model even made up Neos policies that don’t exist. Others generate considerably more, but everyone stressed how closely they review what comes out.
One agency uses local models on their own laptops for client data, GDPR topics, and pitches. Another makes AI use dependent on the client’s signature; only one has signed so far, so their Neos projects remain handcrafted. The project where only specifications get written runs with exactly this kind of consent.
Sebastian uses Claude partly “locked up”: their open-source wrapper claude-safe puts the coding agent in a sandbox so prompt injections can’t reach sensitive files. “That’s part of our policy.”
And elsewhere again, AI has long since established itself in the programming department: Fusion code from the agent, they said, is sometimes indistinguishable from hand-written code to the naked eye. The same person spoke openly about the crisis of meaning this triggers, and about the worry for juniors who no longer get to make their own mistakes.
An hour on AI, barely a model name
Listening, I noticed what almost never came up in this round: model names. The conversation wasn’t about tools but about lines: who gets to see what, what needs sign-off, what stays manual, and what people still learn themselves. Even the person who generates nothing can explain where their line is. Even the agency that generates almost everything has put rules, reviews, and a quality bar right next to it.
My line isn’t the code
At Code Q we work intensively with AI on Neos projects. We now get Fusion and YAML generated reliably, but that didn’t work out of the box. We first had to teach the tool how Neos works and how Neos works at our place. That means: internal docs and reusable skills (instruction files for the tool) with guidance on node presets, naming conventions, constraint logic, and our typical solution patterns. Since then, structural tasks run very well, like retrofitting a description field for every image. Visual component building is where the tools still struggle more: when we have a slider built, it just ends up looking kind of mediocre.
The harder part for me isn’t code generation, it’s access. Security is a super complicated problem, one we’ve only partially solved ourselves. For data access, I’m currently building my own super MCP server through which agents reach systems like our project management tool, time tracking, logs, and Sentry. It’s not finished yet.
The common ground we could still share
In my view, the community doesn’t need a single shared way of using AI. Between deliberate abstinence and almost full generation lie clients, values, and teams so different that one common line wouldn’t be honest.
The economic pressure behind the question of whether an agency uses AI is real all the same, and it can’t be moralized away. In the call, someone put it this way: without AI, you could then “take up carpentry”. Just as true is the environmental question: according to the International Energy Agency, AI’s electricity demand is growing considerably, even if it can hardly be attributed to any single request. Using AI costs something, and going without costs something too.
An excellent task for AI
At the very end of the call, someone remarked that transcription is “an excellent task for AI”, because I had an AI transcript running and had promised to capture the results of the round. This text delivers on that: built from the transcript, with AI support and my own work.
The next call is already scheduled for September, and the different ways we handle a change this big to our work remain fascinating.
) thing from the start.