CrissAlvarezH/persona-hablando

Persona hablando + animaciones

The creator recorded himself talking. His recording is the narration: there is no generated voice (don't run tools/tts.py or standin_words.py). Everything else follows the reel-studio skill and the video's style skill: the style says how the animations look; this format says how his footage is cut and where the animations go. formats/persona-hablando/format.json lists the rules he sees on the Formatos page — follow them.

Files

• footage/ — his clips, in the order they play (names start 01-, 02-…). He uploads them in the studio (Nuevo reel, or the video's page → Grabación). Not versioned; never edit or delete them.
• audio/takes.json — each clip's words with times (from footage.py transcribe).
• audio/edit.json — the cut: the pieces that stay, in order. Yours to edit, with his approval.
• public/footage.mp4, public/vo.wav — the cut and its sound (from footage.py cut). vo.wav is the narration: audio/mix.py mixes it like a generated voice.
• src/words.json, src/vo_env.json, src/footage.json, audio/vo.txt — written by the cut: the words on the cut's clock (so cue(), ACTS and captions work as always), and where each piece starts (JUMPS).
• src/talk/Footage.tsx — <Footage />, the cut as a muted video layer; cutZoom(t) alternates zoom per piece.

The tool runs with the venv (it uses faster-whisper: Ajustes → Requisitos → Transcripción; if it's missing, ask him to install it there): ../../../../.venv/bin/python ../../../../tools/footage.py … from the video's folder.

Steps

1. Transcribe — footage.py transcribe --lang <the profile's language code, e.g. es>. It prints the pieces of a first cut (speech split at pauses > 0.4 s), numbered, with their text. New clips added later are transcribed and appended at the end, without touching the pieces already there.
2. Propose the cuts — read the pieces as an editor: false starts, flubs, "eh"/"este" fillers that drag, a sentence said twice (keep the last good take, unless the first one is clearly better), tangents that don't serve the video's idea, a weak opening when a stronger hook is said later. Show him a short numbered list — piece, what's said, what you'd do and why — and the resulting length against the profile's target. One message, with your recommendation, so "sí" is a full answer. Never cut without his approval. footage.py show N prints piece N's words with times when a cut falls inside a piece: split it in two in edit.json around the flub (cut ~0.05 s before a word starts and ~0.15 s after one ends, never mid-word).
3. Cut — edit audio/edit.json (delete, trim, split or reorder pieces), then footage.py cut. Check the printed length, look at a few frames (./snap.sh out/cut <sec> … once <Footage /> is in the film) and ask him to watch the cut in the studio's player before animating: "Guion" and "Cortes" both count from here. If he asks for changes, edit edit.json and cut again: the words, the captions' text and the timings follow. audio/vo.txt is rewritten by every cut: it says what he said, so don't edit it to change the script — change the cut. Tag audio/beats.json against its lines as usual.
4. Script check — the video-script checklist still applies to what he said: hook in the first seconds, one idea, an answer that isn't given away early, payoff, CTA. If the hook is weak, propose a reorder (a piece from later) or a line he could re-record, rather than animating around it.
5. Animations — brainstorm with the style's visual language a picture per idea (reel-studio → Pictures). Three ways to place them, chosen per beat:
   • Beside him (most of the time): he stays full frame; the animation lives in the free part of the frame (usually the top third or the bottom third, away from his face) or as a card. Numbers, lists, terms, comparisons.
   • Full screen (a few seconds): the animation covers the frame and his voice keeps going. For what is better seen than said (a process, a map, a chart). Come back to his face before he makes the next point.
   • None: when he is making an emotional or personal point, his face is the picture. Find his face's position in the frame first (a still of the cut) and keep it clear: never cover his eyes or mouth.
6. Build — src/film/Film.tsx: <Footage zoom={cutZoom(t)} /> as the bottom layer (instead of the style's background), the scenes on top, captions over both, <Audio src={staticFile('mix.wav')} /> as always. The style's backgrounds and paper textures belong to full-screen beats only. Acts in src/lib.ts from cue() as usual; the template's TOTAL holds the last frame for the closing second. Each jump between pieces (JUMPS) needs a cover: cutZoom alternates the framing, or an animation entering on that cut.
7. Board, sound, mute test, render — as in reel-studio. In the mute test, the frames must make sense without sound: if long stretches are only his face, say so and propose where an animation would carry the idea.

Delivery

Report the cut's length, how many pieces were removed and why (in a line), what is animated where, and anything synthetic. Record in LESSONS.md (general part) what cost a revision of the cut.