Simon Willison proposes using HTML elements in GitHub PR descriptions to manage verbose AI-generated content from coding agents like Claude Code and Codex. While agent-written descriptions are accurate and comprehensive, their length discourages human review. Willison's solution places a concise, human-curated introduction—often with a screenshot—followed by the full agent output wrapped in , using the model name and effort level (e.g., "Claude Opus 5.5 Medium") as the summary label. This pattern preserves all technical detail while enabling quick scanning. The technique works natively in GitHub Markdown. Examples include datasette PR #3052 (redesigning row pages as vertical lists for wide tables), PR #3043 (adding pagination context), PR #3042 (lazy-loading edit-tools.js, ~220KB, only for editable users), datasette-agent PR #51 (new register_agent_reference_types plugin hook for chat composer # tagging), and llm PR #1838 (CLI commands for managing OpenAI-compatible endpoints in extra-openai-models.yaml). Willison offers mnemonics for tag order: summarizes the content, and the phrase "details-summary" (two then three syllables) aids recall. He suggests GitHub add a Markdown editor button for the pattern.