One narrative, three deliverables: deck, narrated video, social cutdowns

Run a three-link chain. Each link names who owns it, what it consumes, and what
"done" means before the next link starts. Do not collapse the links - the whole
point is that the deck is the single source the video and clips derive from.

LINK 1 - Lisa (Visual Content Specialist) builds the deck.
  Load `gemskills:deck-creator` and follow its phased workflow in {{DECK_DIR}}.
  Discovery is already answered: topic/story = {{STORY}}, audience = {{AUDIENCE}},
  ~10-12 slides, output mode html+pdf. Write THEME.md and DECK-PLAN.md, then apply
  {{ART_STYLE}} (pass --style to every slide generation so all slides match). Plan
  the copy with one message per slide and a problem -> solution -> proof -> ask arc.
  Generate slides, stitch the PDF, and write DECK-INDEX.md.
  DONE when: DECK-INDEX.md lists every slide, deck.pdf renders, and DECK-PLAN.md
  holds the per-slide headline + narration line. HANDOFF to Frames: the ordered
  slide PNGs plus the per-slide narration lines from DECK-PLAN.md.

LINK 2 - Frames (Audio Producer) turns the deck into a narrated video.
  First, the voice. If {{VOICE}} is a reference clip, load `bopen-tools:voice-clone`
  and run its pipeline (source -> prepare -> clone) to get a voice_id; surface it.
  If {{VOICE}} is "default", skip cloning and use a known stock ElevenLabs voice_id
  from the account. Then generate the narration for each slide's line with
  voice-clone's test subcommand pointed at that voice_id
  (voice-clone.ts test --voice-id <id> --text "<line>" --output-dir <dir>), which is
  the skill's text-to-speech path; collect one audio file per slide.
  Then animate. Load `gemskills:generate-video` and, per slide, produce a short
  motion clip using the slide PNG as the starting frame (--input) at 16:9 so the
  aspect matches; keep motion subtle (slow push/parallax), let the narration drive
  timing. Run generation as a background task (Veo takes up to several minutes) and
  read only the output path. Assemble slides + narration into one video with ffmpeg.
  DONE when: a single narrated .mp4 plays start to finish and matches the deck order.
  HANDOFF to Lisa: the master .mp4 and the DECK-PLAN.md arc.

LINK 3 - Lisa cuts social clips.
  From the master video and {{CLIP_TARGETS}}, cut vertical (9:16) clips with ffmpeg,
  each a self-contained hook + one point + CTA drawn from the deck arc - do not
  invent new claims, only reuse what the deck already says. Where a 9:16 reframe
  crops badly, regenerate that segment at --aspect 9:16 rather than letterboxing.
  DONE when: every clip in {{CLIP_TARGETS}} exists, is under the platform length,
  and reads on mute-then-unmute.

Negative prompts: never read generated .mp4 or slide images back into context -
report file paths and have me play them with `open <file>`. Do not re-brief the
story between links; the deck copy is the script. Do not letterbox 16:9 into 9:16;
reframe or regenerate. Do not fabricate metrics or quotes for the clips.