Ask HN: Are PMs and colleagues sending you untested AI PRs that do not work?
5 comments
Do you have automated test for the PR?
> So now I have to review broken PRs at breakneck speed because everyone has gotten so productive that I've become a bottleneck.
These PRs seem like they would take mere minutes to identify as being slop, no? I guess if you're getting dozens of them, that can be a time-sink.
In any case, it's good that you're a bottleneck; you're more than that, you're a sieve, preventing this crap from getting into the codebase.
I, personally, would be up front: "This was written by AI, with zero tests and zero thought. It doesn't follow best practices, and will not be merged as is. I will be rejecting all PRs of this nature in the future." Then close the PR, and copy-and-paste as needed; after the third one, contact their manager.
These PRs seem like they would take mere minutes to identify as being slop, no? I guess if you're getting dozens of them, that can be a time-sink.
In any case, it's good that you're a bottleneck; you're more than that, you're a sieve, preventing this crap from getting into the codebase.
I, personally, would be up front: "This was written by AI, with zero tests and zero thought. It doesn't follow best practices, and will not be merged as is. I will be rejecting all PRs of this nature in the future." Then close the PR, and copy-and-paste as needed; after the third one, contact their manager.
You don't even need to object to the genAI aspect of it. Submitting untested code as a PR is unacceptable, period, regardless of whether or not a human wrote it.
So now I have to review broken PRs at breakneck speed because everyone has gotten so productive that I've become a bottleneck.
Usually I get 2 scenarios:
1. Some engineer that consumes the API I built, finds an issue and immediately assumes its an API bug and tells the AI to go fix it, so I get a PR and a message saying "I fixed it, can you approve?". I look at the PR, and not only does it not come remotely close to fixing anything, it does not even build, bc the AI just invented a problem, created some solution and came back with a report on how the job was done. In the end the issue ends up being on the client all along, but zero time was spent investigating that first.
2. PM wants new feature, make a really nice, detailed ticket, tells AI to go at it, AI builds it, completely breaking architecture, just stacking things on top of each other, adding endpoint in weird places, not respecting domain boundaries, etc. Reports back as job done.
I have seen a, maybe 5% success rate on these scenarios, mostly when the changes are closer to one liners.
Has anyone been having this productivity boost experience? How have you handled it?