Agreed. The value of PR descriptions is often in things that are NOT directly inferable from the changes:
- What original problem led to this PR?
- Are there related issues / prior discussions?
- What alternatives have been tried? Does this supersede previous attempts?
- Why are certain seemingly unreasonable choices made? What are the tradeoffs?
AI that generates summary based on the changes of an isolated PR is not really helpful in a meaningful way. Good PR descriptions need not only a holistic understanding of the entire codebase, but even the contribution history of the project. That’s of course harder to do, but I think it’s a possible direction to make it actually useful.
Thank you GitHub, but that’s not how we build software. AI should help us focus more on reasoning, design, communication.
Filter spam and unnecessary noise, not helping creating them.