gioandthemachin

agreed, much better when docs and PR descriptions are written by humans. Humans show doubt, when reading you can see which part they're confident with and which one they aren't. That's a useful signal that is lost with AI-written PR notes. AI sounds always confident, whether it's right or wrong.

word count and format parsing measure the shape of the narrative though, not whether it matches the diff. I think the question is not if it's well formed, but if it makes sense and has all the information, and I don't think you can do that with heuristics. Do you do anything in addition to that, like for example, an extra agent session where given the narrative and diff, the agent lists the claims it cannot verify from the code, or something along those lines?

annoying AF, but maybe we'll get a free reset out of it