Bun serves docs to Claude Code as Markdown instead of HTML (twitter.com)

Bun — the fast JavaScript runtime and web server — appears to be serving its documentation to Anthropic’s Claude Code as raw Markdown rather than rendered HTML. The change (or misconfiguration) means Claude Code receives content in a simpler, text-first format that preserves code fences, headings and inline formatting directly, instead of an HTML DOM that requires stripping tags or rendering. That matters because LLMs and code-focused assistants often prefer semantically clear Markdown for retrieval and context windows: it reduces noisy tokens from HTML, keeps code blocks intact, and can improve prompt-conditioning and generation quality when the model ingests docs. Technically, this behavior likely involves content negotiation (Accept headers), user-agent sniffing, or a default Content-Type of text/markdown being returned for doc routes. For the AI/ML community the implications are mixed: Markdown can make RAG pipelines and prompt prep simpler and more token-efficient, but it sacrifices structured metadata available in HTML (microdata, canonical links, link graph, schema.org) that downstream tools and crawlers rely on. Web developers should verify server responses, Content-Type headers and Accept-header handling to ensure clients get the intended format; LLM toolchains should be robust to either format and consider normalizing inputs or leveraging Markdown when code fidelity is a priority.