GitHub profile: @patbaumgartner
Your work involves coaching teams. What is your approach?
A lot of my work happens while the normal business still has to keep running. You cannot stop everything for six months to “transform.” We usually start by solving immediate problems together, understand where the team currently stands, define where we want to go and how we will know that we are improving, and then move there step by step.
Depending on the situation, that can mean Software Craft practices such as TDD, refactoring, continuous delivery and XP; architecture and modularisation with Spring Boot, Spring Modulith; Spring Security; optimising Java and Spring Boot applications for containers; or increasingly, agentic engineering.
The important part for me is that I work with the team. We make decisions together. I don’t want a team to become dependent on me. Ideally, after some time they no longer need me except perhaps as a sparring partner.
And maybe that explains why I find this new generation of AI tools so interesting: I like working with people, learning, experimenting, and solving problems. AI gives us another very powerful tool for doing exactly that.
Can you share a good story about how GitHub Copilot has helped you recently?
There are actually three very different examples that changed how I think about working with GitHub Copilot.
Let’s start with the first one.
One was a migration of around 900 teaching slides across 15 presentations from Reveal.js to Marp. These were not simple slides. Many contain code, animations, and even interactive examples where I can change JavaScript live while teaching. More importantly, the presentations had already been battle-tested with students. The flow worked. Losing one small piece could break the story I use in the classroom.
Doing that migration manually would probably have taken me a week or more—and realistically, I might simply have avoided doing it because it was such tedious work.
With Copilot, I approached it differently. We first analysed the existing material, defined the target, created a migration plan, established constraints, provided examples, and defined what had to remain intact. Then we worked from the big structural changes down into the details.
The really important part was verification.
We generated screenshots of the Reveal.js and Marp versions and compared them visually and algorithmically. We built a small application where I could run both slide decks next to each other to verify animations manually. We even compared the generated PDFs.
That meant the agent did not just say, “Migration complete.” We had evidence.
The entire migration took me roughly eight to ten hours. I still owned the content, the teaching intent, and every meaningful change. Copilot could fix a typo automatically, but if content, versions, dates or meaning changed, that was something we discussed.
For me, that experience reinforced an important principle: delegate execution, not responsibility. I own the result.
Can you share the second example?
A second example came from building a terminal UI application with Tamboui. It initially worked perfectly in Linux under WSL, and I had completely forgotten that somebody might run it directly in Windows CMD or PowerShell.
And then a real user did.
They discovered that the Enter key was simply ignored in some forms. That is another principle I strongly believe in: ship software early, because you learn from how people actually use it.
We used Copilot to turn that feedback into reproducible UI tests. On Linux, Copilot was navigating the application with tmux. Through working with it, I learned about psmux and could execute similar flows in PowerShell. We reproduced the problem, created a local fix, traced the real problem into Tamboui, and I submitted the upstream fix, which was merged.
During this work, the test descriptions themselves became too large for the available context. So I asked Copilot to help me redesign the way the tests were represented. Instead of verbose tables containing identifiers, names and repeated descriptions, it proposed a much more compact arrow notation: step → step → input, with additional information only where it was critical.
That surprised me. We were not just using AI to execute information; we were working together on how information should be represented for execution. It did not have to be maximally human-friendly. It had to preserve the semantics while being efficient for the agent.
How about the third example?
The third example is completely different again.
Someone close to me needed to practise mathematics, and I thought: why don’t we build a small learning app?
It covers addition, subtraction, multiplication, division and division with remainder for grades one through four. But I did not want to build another simple quiz application. Together with Copilot, I researched learning approaches such as spaced repetition, immediate feedback, adaptive difficulty, motivation, cooperative learning and gamification without dark patterns.
Interestingly, during that research Copilot challenged one of my original assumptions. I initially wanted the application to be generic for children everywhere. The research led us toward looking at the actual school curriculum instead. I realised that the order in which topics are introduced, the operations and number ranges differ enough that making it “universal” immediately was unnecessary.
So we focused on Switzerland first and aligned the progression accordingly. Because Switzerland is multilingual, the application is available in German, French, Italian and English.
It also has cooperative and competitive two-player modes, and when someone gets an answer wrong, the application explains one way to solve it and brings similar questions back through spaced repetition.
I deliberately chose an architecture without a backend. There is no reason for me to collect children’s personal data on a server, so the data stays on the device and the application can simply be hosted on GitHub Pages.
What started as a small experiment is now being used and shared among friends. It accidentally became a little product.
That experience showed me another side of Copilot: don’t only use it to implement your solution. Use it as a sparring partner to expand and challenge the solution space before you start building.
How do you typically use Copilot?
I actually don’t use the classic Copilot autocomplete experience anymore.
My work has moved almost completely toward agentic workflows.
I use VS Code when I need a tight visual feedback loop—when I want to look closely at documents, code, presentations or intermediate results and adapt the process as I go.
For larger execution-oriented tasks, I often work with the coding agent from the terminal. There I am more comfortable delegating a piece of work, letting it execute, and then inspecting the complete result. The interface I choose depends mostly on the feedback loop I need.
I also had a fascinating experience using the coding-agent workflow through GitHub while travelling to JavaLand. I had started exploring an idea for a Maven plugin that optimises which Spring Boot auto-configurations are loaded. During roughly a four-hour journey, I was steering the agent from my phone: giving it a task, reviewing what it had done, thinking about the next step, and sending the next instruction.
By the time I arrived, I had a working Maven plugin, a CI pipeline, performance tests against a sample application, and reports that could demonstrate whether the optimisation actually worked.
That was the first time I had worked that way, and I was amazed that it actually worked. I wasn’t trying to write Java on my phone. I was reviewing, deciding and steering while the agent executed.
I also use MCP servers when they give me useful capabilities or context. For example, I have used Context7 for documentation, integrations with my Logseq knowledge base, and I have experimented with an MCP integration for Adobe Premiere Pro.
But I don’t use MCP just because I can. If a good CLI exists, I often prefer the CLI because it feels more direct and faster to me.
The agent is also not a “junior developer” in my mental model. I think of it much more as a sparring partner. But just like anyone joining a project, it first needs to be brought up to speed.
Context is king.
Can you share some tips and best practices?
The biggest thing I would recommend is: stop one-shot prompting.
For anything non-trivial, I want to understand the problem before execution starts. Quite often I explicitly ask the agent to interview me first. The purpose of that interview is to uncover assumptions and expectations that are still implicit.
A good agent can execute the wrong idea incredibly quickly. That does not make the detour cheaper.
I experienced this myself while working on my Spring Boot auto-configuration optimiser. Initially I wanted to support both Spring Boot 3.5 and 4.x. There were architectural differences, so we implemented both approaches. Later I realised I no longer needed the 3.5 path and removing it was painful.
The lesson was simple: detours are expensive. Get the requirements right before you accelerate.
Once the intent is clear, I want a plan. When I review that plan, four things matter especially to me: are assumptions visible, are verification gates defined, does the approach fit the architecture, and does it avoid unnecessary work?
The second principle is to define task-specific verification gates.
“Done” should not mean that the model says it is done.
For a refactoring, the gates might include passing tests, proving that all intended files were handled, updating documentation and leaving no TODOs behind. For the slide migration, the gates included screenshots, visual diffs, animations, PDF output and content completeness.
The gates depend on the task. The principle does not: the agent should be able to provide evidence that the work satisfies the definition of done.
I recently came across a skill called unlazy that follows a similar idea: create explicit gates, work through them deliberately, and prove that they have been completed rather than taking shortcuts.
The third principle is to turn your lessons into reusable skills.
I have started writing and maintaining my own Agent Skills. When I learn something from an execution—perhaps the agent misunderstood a convention, a useful check was missing, or I found a better way to approach a task—I update the skill.
The next time I perform that task, the lesson is already there.
One example is a skill I call maintainer-audit, inspired by work in the open agent skills community. I use it across my own repositories. It checks and, where appropriate, fixes the things I expect a maintained project to have: repository documentation, CI pipelines, Dependabot configuration, CodeQL, consistent cleanup and other project standards.
The goal is not to generate another audit report for me to read. It brings the repository back toward the baseline I want automatically. That makes my projects more consistent and easier to maintain.
I think of these skills as living engineering knowledge rather than static prompts.
The fourth principle is to put deterministic engineering tools around the probabilistic model.
I have my own quality harness that wraps tools such as Checkstyle, PMD and Spring Modulith architecture checks and produces a consolidated report that the agent can consume. I can run that from hooks and let the agent iterate:
change → execute checks → read findings → fix → execute again.
It continues until there are no findings left.
That does not mean I stop looking at the software. I still run the application, interact with it and decide whether it behaves and looks the way I expect. Sometimes I deliberately implement feature by feature because the feedback loop is much tighter.
I learned that asking for everything at once because “the AI can do it” is usually a bad idea. Large, generic specifications create large, generic implementations—and cleaning them up afterwards can be harder than building smaller increments correctly.
More broadly, I think we developers need to talk much more about how we use these tools. Share the experiments that worked. Share the ones that didn’t. Borrow ideas from each other and adapt them to your own context instead of looking for one universal AI development process.
That is also why I like the idea of AI as a sparring partner. On a Friday night, when I want to experiment with something and no human pair is available, I suddenly have somebody—or something—I can challenge an idea with, review an approach with, or use as a very sophisticated rubber duck.
But I still take those ideas back to real people. Human feedback, product understanding and real-world context remain essential.
How do you think AI is changing our profession?
I don’t think software engineering students should primarily aspire to become code-producing machines. They need to become strong problem solvers. They need to understand users and stakeholders, discover the real problem behind a requirement, challenge assumptions, understand products and markets, make architectural trade-offs, think critically, work in a structured way and transfer useful solutions from one context into another.
The model is extraordinarily good at generating possibilities and executing things. But as a human, I usually know much more of the actual context—including the things nobody ever documented.
So I want to remain in charge of shaping the solution and let the agent take over more of the execution.
In that sense, I think AI is putting the engineering back into software engineering.
And I like that.
