Why I do this
It started in May 2025 with two tasks I would not have trusted myself to do alone: player software for embedded devices in the stadium, in a language I had never used before, and a custom anti-aliasing path in C so that the preview of an LED board does not flicker. Both worked in the end. Both took weeks that would not have been necessary with clearer briefs.
After that I used the agent for everything: for servers, for databases, for deployments, for migrations, for the bookkeeping. That worked astonishingly well and got me into trouble astonishingly often.
At some point I started writing down every correction so I would not have to explain the same thing three times. After a year there were around 160 notes. When I read them through in one go for the first time, it was clear there was a course in there. It is ten patterns, always the same ones, across all projects.
The first version still sits in the stadium repo today and is called WORKING_STYLE.MD. It dates from September 2025 and already contains half of what the ten rules are today, down to the sentence that the agent is one of many versions and the work of the previous ones may be right even if it does not understand it.
The honest part: for a long time I thought everyone knew this anyway. That is not true. If a model of this calibre needs 160 written rules to work with me on production systems, then a team that has been at it for two months needs them all the more. A workshop rarely sells novelty. It sells time: one day instead of three months of trial and error.
The twenty years before
For a workshop about agents in a team it is not only what I can do with the tool that counts. It counts that I know the sides on which the introduction fails: requirements, product, code and the people who are supposed to do this on the side.
Technical project lead in an agency group
Projects between client, creative and development. That is where you learn to write a requirement down so that the right thing gets built in the end, and to notice when a description is not yet enough. The same work goes into every usable brief for an agent today.
Web analyst at a large e-commerce provider
Preparing and interpreting numbers. What was measured and what it means are two different statements. Anyone who has had to keep those apart professionally notices faster when an agent delivers an interpretation and owes the measurement behind it.
Senior product manager at Germany's largest out-of-home advertising company
Product responsibility in a large corporation, with all the alignment that comes with it. Technology, sales and legal rarely sit in one room and still want the same decision.
Product manager, developer and systems architect
For several clients at the same time. In these projects I lead developer teams. The question of how a team works with agents without harming itself is not a theoretical one for me.
What I work on
Clients and system names stay out. What counts is the kind of systems, because that is where the examples come from.
Stadium perimeter advertising: player and preview
Two halves of one project. The player software runs on embedded devices in the stadium, plays several LED boards in sync and has 40 MB of RAM for it: two video players alternating, preloading instead of buffering, restart before the memory fills up. The second half calculates for customers how their design looks from the camera. Templates over 11,000 pixels wide, projected in perspective onto a stadium photo, downscaled up to about 30 to 1. That needed a custom anti-aliasing path in C, because at such heavy downscaling the usual filters read four of every thousand source pixels and the rest ends up as flicker on the board.
Dating app with end-to-end encryption
Go backend on three servers, highly available Postgres, load balancer pair with failover, iOS and Android. Specially protected data under Article 9, hence a correspondingly paranoid baseline: own VPN backbone, encrypted disks, backups in three tiers down to cold storage, web application firewall, central logs. Almost all of it built and run through agents.
Playout cluster for sports advertising
The infrastructure behind the stadiums: playout cluster with failover, storage synchronisation across three providers, central logs, delivery of half a million ad insertions. This is where you learn what fault tolerance means when the fault is live in the stadium.
Content platforms and data pipelines
Several sites of my own with AI-supported content, product data from partner networks and editorial guardrails. Ground rule: invent nothing, every rating is traceable when in doubt. With language models that is more work than it sounds.
Local models in the home lab
Ollama, vLLM and MLX on Mac minis, one graphics card and a single-board computer. Comparing vision models, fine-tuning models, embeddings and vector search, a recommendation pipeline with implicit user signals. Including the expensive lesson that fine-tuning can make a model worse at following instructions.
My own workshop room in Berlin
Renting out two rooms, with everything around it: booking flow, invoices, payment provider, door codes, a tablet for room control, an advertising screen in the shop window. My own company as a test bed for what agents are good for outside of code.
What I am not
So nobody books with the wrong expectations. For these subjects there are better people than me. If I know one, I am happy to pass them on.
- No machine learning background. I do not train models from scratch and I do not explain transformer architecture.
- No formal evaluation methods. My prompt work is practice and measuring the result, not eval methodology with metrics.
- No AI strategy for large corporations. I talk about tools in the hands of teams, not about transformation programmes.
- No vendor comparison. I work with Claude Code and do not know the alternatives deeply enough for a fair verdict.
- No certificates. You get a certificate of attendance if you need one, but no recognised qualification.
- No trainer training. Twenty years of projects and teams yes, teaching background no. That is why the first dates are pilot dates and priced accordingly.
The room
The open dates run in my own workshop room in Berlin Prenzlauer Berg. The room has 4.5 square metres of whiteboard, a 75-inch screen, its own kitchen and a lounge for the breaks. It belongs to me, which is why the groups are small and the day is not run on a stopwatch.
Address
Grellstraße 37, 10409 Berlin
Getting there
Greifswalder Straße S-Bahn station, four minutes on foot
Capacity
Up to 10 people in the workshop room, lounge for breaks
Questions before registering are expressly welcome.
Write to me about what you have in mind. If another format fits better or you do not need the topic at all, I will say so.