Turning LLM output streams into live UIs: Conduit skips app-building with Claude Code
cinooo1 · reddit · 2026-11-18
- The author asked Claude to explain 7/8 of 1000 to his daughter; Claude spent minutes building an interactive app as an artifact, which sparked the question: why generate an app at all?
- Core idea: treat the LLM's entire output stream as an instruction language for another system (e.g. a browser) instead of text for humans; the receiving system keeps its own state the AI can inspect and build on, interpreting instructions into visual, interactive experiences as they stream.
- Experiments with interactive math lessons, visualizations and prototyping look promising; it runs on an existing Claude Code subscription (also tested on Codex) with no extra API cost, and accepts non-text input like drawings and gestures.
- Findings: exact-coordinate rendering needs a smart model; Opus 5.5 medium works reasonably, Fable medium is better for complex cases, and model choice matters a lot.
- Code is MIT-licensed on GitHub (conduit) with a full write-up. As models get smarter and cheaper, generating experiences directly could unlock a richer suite of bespoke UIs.
More from coding & agent
- Google donates Agent Substrate to CNCF: 500ms restores, 10x density — krishnan · 2026-11-18
- Law school AI course wraps: 6 student teams ship working legal AI prototypes — jkubicki · 2026-11-18
- n8n's open-source agent Nod: secure remote access via private reverse proxy — victorialslocum · 2026-11-18
- Floot Design launches to kill AI slop: pick your UI before any code ships — heyshrutimishra · 2026-11-18
- A 6-stage roadmap for learning AI-powered system design — goyalshaliniuk · 2026-11-18
- FeinNote: a portable local Whisper transcription app for Windows, no accounts needed — _fortexe · 2026-11-18