The original title is "How we built a voice-first, screen-aware product guide with the OpenAI Realtime API"

Short summary

The author built a voice-first product guide using OpenAI's Realtime API for sub-1s response latency, capturing screen context only on demand and using semantic selectors to point cursors at UI targets. The guide learns from existing product docs instead of scripted flows, ships as a single script tag, and is now live as Skilly. Key tradeoff: losing intermediate-text control for speed and naturalness.

• •Realtime API collapses STT→LLM→TTS into one bidirectional stream for sub-1-second latency
• •Screen capture gated on user intent, not continuous; semantic selectors ensure stable UI targeting across layouts
• •Grounding guide in product docs enables open-ended question handling vs scripted flows

Generated with AI, which can make mistakes.