I turned Markdown into a protocol for generative UI
8 hours ago
- #AI
- #UI
- #Markdown
- Markdown is proposed as a protocol for agentic UI, combining text, executable code, and data in a single stream.
- Three core ideas underpin the prototype: Markdown as protocol, streaming execution, and a `mount()` primitive for reactive UIs.
- The protocol uses three block types: Text (plain markdown), Code fence (executable code), and Data fence (streams data into UI).
- Feedback loop involves LLM generating markdown with code blocks, executing code incrementally, and feeding output back to the LLM.
- Streaming execution allows statements to execute as they are generated, improving responsiveness.
- Four data flow patterns are identified: Client → Server (forms), Server → Client (live updates), LLM → Client (streaming), and Client → Server (callbacks).
- A slot mechanism enables complex UIs by mounting skeleton interfaces first and filling sections later.
- Security considerations are noted, with the assumption that security is reasonably solved for the prototype's purposes.
- The prototype leverages existing LLM knowledge of markdown, TypeScript, and React, optimizing for LLM ergonomics.