---
name: mc-voice-refs
description: >
  Fill captures/<slug>/voice.yaml with real reference audio/video links and
  clean-segment timestamps for voice cloning, or a stylized fallback spec.
  CHEAP-MODEL WORK — search and listening, not synthesis; never spend
  frontier capacity here. Use when the user says "find voice refs for
  <name>" or "fill in the voice segments".
---

# MentorCore Voice References (cheap link-and-segment pass)

Fill in `captures/<slug>/voice.yaml` per `docs/PERSONA-CAPTURE.md` §voice.yaml.
The capture's `pacing_notes` (frontier-authored) describe the target delivery —
your job is to find audio that matches it.

## Steps

1. Read the capture's `voice.yaml` (policy + `pacing_notes`) and `sources.md`
   (often already names interviews/podcasts).
2. **Cloned tiers** (`cloned-personal` / `cloned-licensed`):
   - Find 2–4 candidate recordings (podcasts, interviews, talks). Verify each
     URL actually resolves — never list a link you didn't check.
   - Select `clean_segments`: 5–15s spans of SOLO speech — no music, no
     crosstalk, no audience. Selection criteria, in order: (a) pacing matches
     `pacing_notes` — F5-TTS clones the segment's cadence, so an
     unrepresentative segment poisons the voice; (b) audio cleanliness;
     (c) emotional register typical for the persona.
   - To check segments: `yt-dlp -x` the audio and listen/inspect locally, or
     use timestamped transcripts. Record why each segment was chosen in
     `quality_notes`.
   - F5-TTS output is `cloned-personal` ONLY (CC-BY-NC) — never suggest
     cloned output for a public tier.
3. **Generated/stylized tiers** (no usable recordings, e.g. long-dead
   figures): flesh out `stylized_fallback.description` (age, register, tempo,
   accent era) from the capture's style.md; leave `kokoro_voicepack_hint`
   null unless you can audition voicepacks locally.
4. Commit: `Voice refs: <slug> (<n> segments)`.

Hard rules: no invented URLs, no invented timestamps. An empty `references`
list with an honest note beats plausible-looking fabrication — the compile
stage will fail loudly on a dead link, but a wrong-cadence clone fails
silently.