I was reading this short piece by Niklas Gruhn and recognized myself in it immediately. Same situation, same feelings. I've been on both ends of it. I've sent the wall of text, and I've received it, and receiving it is worse. Something in your brain flags it before you finish the first line, and it makes sense. AI produces brilliant things sometimes, but most of the time it lands on the average of everything ever written about the topic, and we all have a pattern matcher that detects the average in milliseconds (no proof, but no doubts either).

I agree with Gruhn, and he's explicit about it himself: by all means, prompt AI. The problem was never the tool, it's stopping at iteration zero. Still, I have some differences with the fix he proposes. I don't think you necessarily have to write the thing in your own words. You have to read it, criticize it, guide it, jitter it, kick it once and again until it says something you would actually defend. The certificate isn't that every word came out of your own fingers, it's how many iterations and shake-ups it survived before you sent it.

I don't want to be a hypocrite. I use AI all the time, including in the writing of this piece. Same with code, where I barely write a line anymore, and I don't think I need to in order to ship something of quality. I guide it toward what I think good looks like, I question it, and I push it to question me back. What I try to avoid, whether writing or implementing code, is settling for the first thing that comes out.

There's something else in his code review example. It doesn't bother me exactly, it just happens, constantly, and to me it's a very clear signal. Here's the case: user A asks their AI to implement a feature and open the PR. User B, the reviewer, hands that PR to their AI and asks it to review and write feedback. B copies and pastes the feedback into the PR. User A copies the feedback from the PR and pastes it into the same session that wrote the code. This iterates until the meat proxy on the reviewing side (user B) writes "LGTM".

Now count how many humans added value in that loop. The answer is zero. So what's left is the ceremony: the handoffs, the notifications, the waiting, with a person standing at each end contributing nothing but copy-paste and latency. If both ends of the process are the same machine, the humans in the middle aren't reviewing, they're a transport layer.

So maybe being a meat proxy isn't just mildly annoying for everyone else, maybe it's a signal that something needs to be done at that point in the process. Just like a single manual step in a deployment pipeline used to be: something here should have been closed end to end and wasn't. When I catch myself being the meat proxy, the question isn't just should I have written this myself? It's why is there a human in this position at all?

There are two possible answers. Either that human is there for a reason: judgment, accountability, taste, the know-how and the context only that person has, among other things; to me all of those are very valid reasons and I'd defend them, but then they have to be exercising it, not just transporting. Or there is no reason, and then the honest move isn't to write a nicer message: it's to go build the link that finally connects both ends and takes the human out of the middle.