article --summary
The summary
The post walks through a talk about using AI in day-to-day open source maintenance, with the aim of sharing lessons that stay useful as specific tools and models change. The author describes two projects he built with AI help, one of which is a large C++ codebase written by someone who does not know C++. He also explains how AI handles routine GitHub tasks like keeping docs in sync and fixing bugs assigned to an automated bot.
The most practical section covers a prompt injection issue that could have exposed PHPStan's secrets, how it was fixed, and GitHub Actions security pitfalls that are unrelated to AI. Readers running AI agents in CI should consider this part first. The talk closes with the author's view that a well-written issue is preferable to a generated pull request, and that human judgment about taste and priorities remains essential.
Slides are available in English and Czech, with arrow-key navigation and a fullscreen option, and each slide ends with a short spoken summary so the recording is not required.
Editorial note: This is a summary of an article first published by PHPStan. The full piece is on their site.