Making the Most of AI in Open Source
November 20, 2026 · 2 min read
Today I gave a talk at TechMeetup Ostrava about how AI changed the way I work on PHPStan. AI tools and trends change from week to week, so instead of tricks that expire with the next model, I tried to share lessons that should last.
The talk was in Czech, but I translated the slides to English. Every slide also ends with a short summary of what I said, so you don’t need the recording to follow along:
- What I built with AI that wouldn’t exist without it: documentation for all 1100+ error identifiers, and PHPStan Turbo, about 200 thousand lines of C++, even though I don’t know C++.
- AI in everyday work on GitHub: keeping the documentation in sync with the code, and
phpstan-botfixing bugs I assign to it. - One big disaster that almost happened: how a prompt injection in an issue could have stolen PHPStan’s secrets, how I fixed it, and the GitHub Actions security traps that have nothing to do with AI.
- Mob Programming 2.0: why I’d rather get a well-described issue than a generated pull request.
- Where humans are still irreplaceable: taste, priorities, and the idea an agent doesn’t come up with on its own.
- Take care of yourselves.
Click into the slides and use the arrow keys, or the controls in their bottom left corner, which also go fullscreen. They’re also available on their own page, and the Czech original too.
If you’re running AI agents in GitHub Actions, the third part is the one I’d read first.