Simon Willison, an independent open source developer who maintains Datasette and the LLM command-line tool, has a complaint about agent-written paperwork. A pull request (PR) is a proposed code change that someone reviews before it is merged, and coding agents now write the description that comes with it.
“Today’s coding agents are great at writing PR descriptions,” he writes. “The problem is that they are too good.” In his words, “An agent-written PR description is accurate and comprehensive.” His reaction on seeing one is “I’m not reading all of that.”
A human opener, with the agent’s text folded underneath
His fix is small. He writes a short introduction himself, then pastes the agent’s full description inside an HTML <details><summary> block. GitHub renders that as a collapsed line that expands when clicked. Willison labels it with the model name and often the effort level, such as “Claude Opus 5.5 Medium”, “to make it abundantly clear what is going on.”
What he wants from the opener: “I want human editorial judgement.” He links five of his Datasette and LLM PRs that use the pattern.
The summary has to carry the warnings
Folding hides text; it does not shorten it. The agent’s account of how it checked its work is still there, just out of sight. A reviewer who trusts the opener and never expands the section may merge a change without seeing that the tests only covered part of it.
Willison is a maintainer reading PRs on his own projects, so his opener comes from someone who has read the diff. If you are the author, put what a reviewer must not miss in that opener: what you actually ran, what you did not test, and anything the agent flagged as uncertain. Whatever goes only in the folded section is what a skimming reviewer will never read.
A teammate opening your PR then sees one paragraph from a person who stands behind it, with the agent’s full log a click away if the change looks risky.
End of article