Hiding AI-Generated PR Details with <details><summary> on GitHub

The readability paradox of AI-written PRs

Willison notes that modern coding agents such as Claude Code and Codex produce PR descriptions that are “accurate and comprehensive,” covering problem statements, solution details, and verification steps. The very completeness that makes them useful also turns them into “sloppy” blocks of text that reviewers skim past, regardless of the factual value. His frustration stems from an instinctive reaction to “not reading all of that,” even when the author is himself.

Collapsing the bulk with native Markdown

The core of his proposal is to wrap the agent-generated essay in HTML’s <details><summary> tags, which GitHub renders as a collapsed section with an expandable arrow. The summary line—often the model name and effort level, e.g., “Claude Opus 5.5 Medium”—serves as a high-level cue, while the opening paragraph outside the tags provides a human-curated overview. Willison demonstrates the markup:

<details>
  <summary>Claude Opus 5.5</summary>
  <h2>The hook</h2>
  Plugins implement <code>register_agent_reference_types(datasette)</code>...
</details>

Because GitHub Markdown supports these tags directly, no additional tooling is required.

Real-world examples from the Datasette project

Willison lists five recent pull requests where the pattern was applied, including PR #3052 (redesigning a row page) and PR #3042 (loading edit-tools.js on demand). Each entry shows a succinct opener, sometimes accompanied by a screenshot, followed by the full AI-generated text hidden inside the collapsible block.

Remembering the tag order

To avoid mixing up the element order, Willison suggests the mnemonic “details-summary,” emphasizing that the <summary> tag contains the clickable line and must be nested inside <details>. He also wishes GitHub added a Markdown button for the tags, which would simplify mobile entry.

Our read: Willison’s approach reframes the PR description from a monolithic document into a tiered presentation that respects both human attention limits and the completeness of AI output. By leveraging a built-in GitHub feature, the method sidesteps the need for custom scripts or third-party plugins, making it instantly deployable across repositories that already accept HTML in Markdown.