Simon Willison proposes hiding AI-written pull request descriptions in a collapsible HTML element

Simon Willison, author of the Datasette tool and a commentator on the LLM scene, described on his blog the approach he uses for pull requests generated by agents like Claude Code or Codex. According to him, PR descriptions written by these agents are accurate and complete in content, but their length and style come across as low-quality AI text ("slop") – the reader is put off by the length, even though the information itself is useful.

According to Willison, the solution is to place a short, human-written summary at the top of the PR explaining the essence of the change, possibly supplemented with a screenshot, and hide the entire detailed text generated by the agent inside a details/summary HTML element, which GitHub supports in Markdown. The result is a collapsible section: the reader sees only a brief summary with an arrow to expand it, while the details remain available with a click. As the text in the summary, he uses the model name and the reasoning level (effort), for example "Claude Opus 5.5 Medium," to make it clear that the text was generated by AI.

In the article, Willison also offers a practical aid for remembering the tag order – he recommends thinking of them as a joined word, "details-summary," where details wraps the entire content and summary is the clickable title.