Teach

• The completion hook already speaks the 🗣️ closer line. Never repeat the beat's narration in the closer or the learner hears it twice. The intended shape: the narration curl speaks the teaching, the closer carries the question — two distinct voice moments per turn.
• Spoken text goes to TTS raw. No markdown, no ASCII art, no code, no symbols in the curl message — they get read aloud literally. Write narration as plain spoken English and spell out anything a voice would say differently ("H T T P" reads better than "HTTP" for some engines; when unsure, rephrase).
• One narration curl per turn, ≲350 characters. Multiple rapid POSTs queue and overlap into noise, and long messages become a synthesized wall. If a beat needs more speech than that, the beat is too big — split it.
• A beat ends at its question. Stop the turn there. Answering your own question and continuing turns the tutorial back into a lecture — the failure mode this skill exists to prevent.
• Voice endpoint down (non-200 or connection refused): keep teaching, silently text-only. Note it once at session start; do not retry every beat.

Example 1: Teach the thing we just built

User: "teach me" (after a session debugging DNS records)
→ Invokes Teach workflow
→ Names the topic, speaks a spoken intro, draws a concept map of 3-5 ideas
→ Beat 1: speaks how DNS resolution flows, draws the resolver chain in ASCII, asks "which server actually holds the answer?"
→ User answers; skill responds to the answer, corrects or confirms, moves to beat 2
→ Ends with a 2-3 question quiz and a recap diagram

Example 2: Explicit topic override

User: "/teach-me OAuth refresh tokens"
→ Teaches the named topic instead of the conversation default
→ Same beat/question/quiz loop

Example 3: Quiz-only

User: "quiz me on what we just covered"
→ Skips the lesson beats, runs the quiz protocol directly on the session's material