Lately, I've been thinking about how much time AI actually saves and noticed something worrying.
I particularly took a note in how PR descriptions and user stories generated by AI through a prompt of somebody else negatively impact my own performance.
For PR description, the AI usually sums up what was changed in the code, instead of what was changed in the product. And it does that in an exceedingly verbose way. This is useless. I see what changed in the code - that is the whole point of the PR. But I would like to know what changed in the product behavior: a) because it updates my understanding of the product and b) because it is then much easier to look for contradictions in the code.
Also, since these summaries are quite verbose, they can easily hit the character limit for description length. Sure, Claude can detect this and condense the text, but it can't take into account that when I run the deployment pipeline, that pipeline can't update the PR description with the link to the deployed environment, as it has no characters to spare.
The issue is even worse with user stories. If you give AI access to the code, it has the bad habit of presuming the implementation and suggesting that in the US description. It tells you how the code should change instead of what needs to change in the product behavior. It will link the follow-up stories and elaborate what is out of the scope of the current story. Sure, it might be handy sometimes, but that immense wall of text requires much more cognitive energy to parse, understand and then to reason about what is actually missing.
In more complex products, split across multiple repos, you can easily have a lot of dependencies and processes that might not yet make it into the documentation. Claude doesn't know about them. I do and I do check for these when refining the story. But if I have to read through unnecessary mumbo jumbo, it wastes not only my time, but also time of all the people on the meeting.
So next time you'll be prompting Claude think about this - it might save you time, but it might waste somebody else's time in the process.
P.S.: No tokens were burned writing this post.
