Of course instead of using proper Accept/Content-Type headers this whole thing will be vibe-coded on top of user agent sniffing

After Bun reported a 10x token reduction by serving markdown docs to Claude Code, I implemented the same approach for our docs site.