Show human effort

The message was about 1,400 words, it arrived in a channel I’m in on Tuesday afternoon, and it started with “Here’s a summary of the investigation”. It was an agent’s write-up of a slow query, pasted in whole. Somewhere in the middle of it, under a heading about “potential contributing factors”, was the actual cause, an index that had been dropped in a migration two weeks earlier. I found it on the second read. The engineer who pasted it, when I asked, hadn’t.

Two days later Tom Bedor published If You are Asking for Human Attention, Demonstrate Human Effort, and I’ve sent it to more people this week than anything else I’ve read this year. It’s short. His example is a teammate who sent him an AI critique of his design, with a note saying they hadn’t read it themselves and it might not be accurate. His reaction was that if it wasn’t worth the sender’s time to read, it wasn’t worth his. He still forwards AI output himself when it’s useful, but he labels it and adds what he thinks, and he reviews AI-written code himself before he asks anyone else to.

That’s the rule in the title, and it’s the same idea as Clayton Ramsey’s I’d rather read the prompt, which I wrote about last May. Ramsey’s complaint was about generated text pretending to be the writer’s own. Bedor’s is about the more honest version, where everyone knows a model wrote it and the sender just passes it along, and in a way it’s worse, because it’s so easy and feels so harmless. The cost lands on the reader every time, and on a team the reader is usually someone more senior with less time.

What I’ve asked the team for since Thursday is small. If you paste an agent’s output into a channel or a doc, put two lines of your own above it: what you think it found, and whether you checked. If you didn’t check, say so, and say why you’re sending it anyway. “I haven’t verified this but the index theory looks right to me” is fine. A wall of text with no sentence of yours in it isn’t.

The effort doesn’t have to be writing. Most of what I forward these days is agent output too. Before I pass on a list of an agent’s findings, I’ve crossed out the ones I don’t agree with, with a reason, and that selection is the human part. It takes a few minutes, and it’s the only part of the message the reader can’t get by asking a model themselves.

The engineer with the slow query rewrote his message as three sentences and the agent’s write-up as a thread reply. Two people thanked him for it, which I think surprised him.