Cuento Reel

Cuento Reel (narrated multi-character story)

Films a cuento, not a marketing reel. A voice-only Narrator (cuentacuentos) holds the thread in third person; characters are invented avatars that ACT. Two visual building blocks:

• Narration beats (no lip-sync) → SILENT animated clips built with broll-story (ONE multi-panel storyboard sheet → seedance, muted), with the Narrator's voice-over on top. Never one-still-per-clip, never a centered talking-head.
• Dialogue / action-with-speech beats → REAL lip-synced clips of one character at a time (avatar-talking-video / p-video-avatar), woven in as guest scenes that keep their OWN voice. Use this whenever a shot needs ACTION and LIP-SYNC at the same time.

This is the generalized pipeline behind narnia/ (see that folder + its serie-narnia.pauta.md / *.script.md for a full worked example).

Folder layout (a "cuento" project)

<serie>/
  serie-<x>.pauta.md                 # series bible (templates/serie.pauta.md)
  cast/                              # REUSABLE avatars, made ONCE per series
    narrador/                        # voice-only storyteller (older cuentacuentos)
      voice_brief.json  voices/...   # MiniMax voice_id (no face needed)
    <personaje>/                     # face + voice (avatar-invent)
      scene.json  refs/  angles/  voices/  talking_profile.json
  <libro>/<NN_slug>/                 # ONE episode = one reel
    <NN_slug>.script.md              # beat sheet (templates/episodio.script.md)
    dialog/script.json               # character dialogue (audio-theater)
    broll/anim/*.mp4                 # silent animated narration-beat clips
    plan.json                        # master-narration plan (assemble_narration)
    storyboard.json                  # scenes (compose_reel)
    narration.mp3  final.mp4         # outputs

Hard rules (learned the hard way — do not break)

1. NO Ken Burns, ever. Every scene motion: "none". Movement must be REAL (animate the still into video); never a static frame with a push/zoom, and never emphasis on a guest scene (it would re-introduce a push-in).
2. gap: 0 in the plan. Guest (lip-sync) clips have no trailing pad; any gap>0 makes each guest fall short and the picture drifts ahead of the audio. compose_reel now warns if it sees gap>0 with guest scenes.
3. 1:1 scene ↔ plan-segment correspondence, same order. The composer pins every scene boundary to assemble_narration's exact offsets (frame-exact A/V). Keep storyboard.scenes in the SAME order as plan.segments.
4. Narrator is voice-only and OLD. A warm, grandfatherly cuentacuentos timbre (older than a generic narrator). Voice-only — it never needs a face on screen.
5. Original text for copyrighted sources. Copyright protects expression, not ideas. If the source tale is under copyright (e.g. Narnia), write 100% ORIGINAL narration + dialogue inspired by the plot; never reproduce or closely paraphrase the prose. Public-domain tales (Verne, Andersen, Grimm, Quiroga…) are free, but their modern TRANSLATIONS may not be — adapt in your own words. Note personal use.
6. Animation tool by shot type — NON-NEGOTIABLE.
   • No lip-sync during the animation (interludios / b-roll) → broll-story: author ONE multi-panel storyboard sheet and animate it in a SINGLE seedance pass (cheaper + more coherent). Do NOT animate a separate still per beat.
   • Action + lip-sync at the same time (a character moving/acting AND speaking on camera) → p-video-avatar (avatar-talking-video). That's the rule: silent motion = broll-story; speaking-on-camera = p-video-avatar.
7. Series compile. Join episodes with ami/compile_series.py (or the series' equivalent): video dips to/from black at each chapter, audio runs straight through (never fade the voice), a ≥0.5s black pause between chapters and ~1s of black at the end so it doesn't cut abruptly. The intro is silent; music starts on the first episode.

Workflow

A) Once per SERIES

1. Write the bible. Copy templates/serie.pauta.md → <serie>/serie-<x>.pauta.md: north star, visual DNA (style/era/palette), language, cast table, per-episode map.
2. Create the Narrator voice (voice-only). Copy templates/narrador.voice_brief.json → <serie>/cast/narrador/voice_brief.json (older cuentacuentos timbre), then design+clone it:

   python3 .cursor/skills/avatar-invent/scripts/design_voice.py \
       --avatar-dir <serie>/cast/narrador --name narrador \
       --voice-brief <serie>/cast/narrador/voice_brief.json
   # -> <serie>/cast/narrador/voices/narrador.json (voice_id)
   

3. Create each CHARACTER avatar (face + voice) with avatar-invent (pauses once for the casting review — make the SUBJECT a vivid, concrete face so it stays consistent across clips):

   python3 .cursor/skills/avatar-invent/scripts/invent_avatar.py <serie>/cast/<char> \
       --description "<vivid age/face/hair/wardrobe, era-accurate>" --setting <fit> --language es
   # review cast/<char>/scene.json + voice_brief.json, then re-run to generate.
   

   Anchor each character's refs/<char>_hero_master.png and repeat its DNA verbatim in every broll sheet so the face never drifts.

B) Per EPISODE

4. Write the beat sheet — copy templates/episodio.script.md. Mark each beat N (narration), D (character dialogue), VIS (what we see). Then derive:
   • the running Narrator script (only the N lines, third person), and
   • the dialogue list (the D lines, per character).
5. Generate dialogue lip-sync clips (one character per clip). Two paths:
   • Default (character's own MiniMax voice): let assemble_narration do it — add a kind:"guest" segment (it calls avatar-talking-video and uses the clip).
   • Specific non-MiniMax voice (e.g. youthful/child voices that ElevenLabs blocks): synthesize the line with audio-theater (Gemini voices, see dialog/script.json), lip-sync it onto the face, then reference the clip:

     python3 .cursor/skills/avatar-talking-video/scripts/generate_video.py \
         --audio <line>.wav --avatar-dir <serie>/cast/<char> \
         --image <serie>/cast/<char>/refs/<char>_hero_master.png --out-name <dlg_id>
     

     Use the clip as a kind:"audio" plan segment (file = the .mp4) AND a storyboard guest scene (broll_clip = that .mp4).
6. Generate narration-beat visuals (SILENT, real motion) — always broll-story (see Hard Rule 6): one multi-panel storyboard sheet per beat (or per contiguous group of beats) → animate in a single seedance pass, muted. Do NOT animate one still per beat. Save to <libro>/<NN_slug>/broll/story/*.mp4 (or broll/anim/). All will be motion:"none" scenes. Only use p-video-avatar here if the shot also needs on-camera lip-sync.
7. Assemble the master narration (interleave narrator + dialogue). Copy templates/plan.example.json → plan.json (gap:0, segments in story order: tts = narrator, audio/guest = dialogue clips), then:

   python3 .cursor/skills/avatar-reel-composer/scripts/assemble_narration.py \
       <libro>/<NN_slug>/plan.json --base-dir .
   # -> narration.mp3 + narration.align.json + assemble_narration.out.json
   

8. Write the storyboard — copy templates/storyboard.example.json. One scene per plan segment, SAME order: guest for dialogue clips, broll (broll_source:"existing", broll_clip) for narration beats; EVERY scene motion:"none"; a finish block (subtitles + cinematic music, max_words ~6). The script = the full running narrator text. Then compose (reuses the pre-built narration when --out-dir points at the episode folder):

   python3 .cursor/skills/avatar-reel-composer/scripts/compose_reel.py \
       <libro>/<NN_slug>/storyboard.json --base-dir . \
       --out-dir <libro>/<NN_slug> --finish
   

9. QA lip-sync. Confirm each character cut lands where its line ends (extract frames at the boundary and compare to the subtitle/audio). See REFERENCE.md for the exact A/V-sync check.

Templates

• templates/serie.pauta.md — series bible.
• templates/episodio.script.md — beat sheet (N / D / VIS).
• templates/narrador.voice_brief.json — older cuentacuentos narrator voice.
• templates/dialog.script.json — audio-theater dialogue (character voices).
• templates/plan.example.json — master-narration plan (gap:0, interleaved).
• templates/storyboard.example.json — scenes (motion none, guest+broll, finish).

Additional resources

• Schemas (plan/storyboard/scene), troubleshooting, A/V-sync check: REFERENCE.md
• Orchestrated skills: avatar-invent, broll-story, audio-theater, avatar-talking-video, avatar-reel-composer (assemble_narration + compose_reel).
• Worked example: the narnia/ folder and its serie-narnia.pauta.md.