editar
/editar
Montuje dopracowane wideo pionowe lub poziome z nagrania: analizuje klatki, pisze hak i karty słów kluczowych, a potem renderuje je w Remotion.
- Autor
- luisbuilds-dev ↗
- Repozytorium
- video-engineer ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 27.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 27.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Montuje dopracowane wideo pionowe lub poziome z nagrania: analizuje klatki, pisze hak i karty słów kluczowych, a potem renderuje je w Remotion.
Oryginalny opis (EN)
Edit a cleaned take into a finished video (vertical reel or horizontal). Use after /limpiar or when the user types /editar. Analyzes every frame (face and motion), writes the viral hook and keyword cards, composes so nothing covers the speaker, renders with Remotion and runs the quality check.
Jak działa
SKILL.md bez nagłówków drugiego poziomu · 42 linii, 1 351 słów
Podgląd SKILL.md
---
name: editar
description: Edit a cleaned take into a finished video (vertical reel or horizontal). Use after /limpiar or when the user types /editar. Analyzes every frame (face and motion), writes the viral hook and keyword cards, composes so nothing covers the speaker, renders with Remotion and runs the quality check.
---
# /editar
0. Read the user's living guide `estilos/<style>.guia.md` (if it exists) and follow it: it overrides the defaults below.
1. **See the video**: `npm run analizar -- --proyecto <id>`. It reads every frame (10 per second) locally: where the face is, where there is motion, and prints the rhythm per second (motion bar + words). Read `ritmo.json`: seconds with low motion and few words are where attention drops.
2. **Hook** — the first 3 seconds decide everything. Pick ONE angle that the content really supports (never clickbait):
1. *Counter-intuitive*: breaks a common belief.
2. *Curiosity gap*: an unusual result without the mechanism.
3. *Loss aversion*: a hidden cost or a mistake that costs time or money.
4. *Direct targeting*: filters the audience in the first line ("If you use X to do Y…").
5. *Open loop*: an unresolved situation that the video closes later.
Rules: the on-screen text and the spoken first line do DIFFERENT jobs (text = the *what*, short like a YouTube title with niche words; voice = the *why/how*). Max 7 words on screen; wrap 1-2 key words in `*asterisks*`. No greetings, no vague questions, no invented proof ("I tested", views), no ALL CAPS single words, no empty superlatives. Propose 2 options with their category; the user picks.
3. **Re-hooks and cards** (`tarjetas`, max 3 words each): one card between second 3 and 10 that re-hooks (a number, a tool, the reveal coming), then 1-3 per minute for what the viewer must remember. Time each card to the moment the word is said (`desdeMs` from `cortadas.json`), 1500-2500 ms long. Place cards in low-motion seconds from `ritmo.json` when you can: they bring the attention back.
4. **Moments** (graphics + sound). Mark what the content gives you; a script picks the animation and the sound effect:
- `{"tipo": "cifra", "desdeMs", "valor": 87, "sufijo": "%", "etiqueta": "..."}` — counts up, lands with an impact. ONLY numbers the user actually says.
- `{"tipo": "barras", "desdeMs", "titulo", "sufijo", "items": [{"etiqueta", "valor"}]}` — comparisons (before/after, A vs B).
- `{"tipo": "lista", "desdeMs", "titulo", "items": [{"texto", "desdeMs"}]}` — steps; each item ticks when it's said.
- `{"tipo": "sello", "desdeMs", "texto": "CONFIRMADO", "tono": "ok" | "alerta"}` — a stamp for a reveal or a twist.
- `{"tipo": "sonido", "desdeMs", "efecto": "caja" | "error" | "notificacion" | "tecleo" | "obturador" | "glitch" | "tictac" | "ding"}` — a specific sound (money, mistake, app, typing, screenshot, topic change, countdown, success).
- **Illustrations — show what the user TALKS ABOUT** (the most valuable moment type). Read the transcript and turn concrete objects and actions into animated scenes; 1-3 per minute, timed to the words:
- `{"tipo": "ilustracion", "escena": "icono", "desdeMs", "items": [{"icono": "printer3d", "etiqueta": "Piezas en 3D", "desdeMs"}]}` — an object they name, drawn stroke by stroke.
- `"escena": "proceso"` with 2-4 items (each with the `desdeMs` when it's said) — steps of a process joined by arrows.
- `"escena": "caos-orden"`, `"ordenMs"` (when the fix is said), `"etiquetaAntes"`, `"etiquetaDespues"`, `"items": [{"icono": "lock", "desdeMs": 0}]` — messy cables that untangle, get clamped and locked: for "organize", "clean up", "fix", "tidy".
- `"escena": "antes-despues"`, 2 items — the old way crossed out → the new way.
- `"escena": "ensamblar"`, 2-4 items — parts that fly in and snap together (building, components, a stack of tools).
- `{"tipo": "logos", "desdeMs", "titulo"?, "items": [{"logo": "bambulab", "etiqueta"?, "desdeMs"}]}` — real brand logos (in their color, with a soft halo) appearing as each brand is named. Find the slug with `npm run iconos -- --logo bambu youtube`. Only for brands the user actually mentions.
- `{"tipo": "diagrama", "disposicion": "flujo" | "radial" | "arbol", "desdeMs", "titulo"?, "nodos": [{"id", "etiqueta", "icono"? | "logo"?, "desdeMs"}], "enlaces"?: [{"de", "a"}]}` — nodes that appear when said and arrows that draw toward them. `flujo` = steps in order (links optional: they chain), `radial` = one concept and its parts (first node is the center), `arbol` = hierarchy or causes → effects (give the links). 2-6 nodes.
Icons are names in English: find them with `npm run iconos -- printer cable screw` (Lucide + icons saved earlier). If none fits, the same command shows free-license matches from Iconify; save the best one with `npm run iconos -- --guardar collection:name local-name` (license and author go to `iconos/CREDITOS.md`; never use one the command rejects). Compose stops if a name doesn't exist. For a scene that doesn't exist yet (a specific object animation), use `/ampliar`.
- `{"tipo": "camara", "desdeMs", "efecto": "sacudida" | "flash" | "glitch" | "zoom-suave" | "zoom-salida" | "blanco-negro" | "vineta" | "latido" | "desenfoque" | "barrido", "duracionMs"?, "intensidad"?}` — camera effects: shake on a hit, flash on a cut, glitch on a topic change, slow zoom while explaining, zoom-out to reveal, black & white for a flashback, vignette to focus, pulse on a punchline, blur-in, whip-pan transition between sections.
If the user wants something not listed, use `/ampliar`.
Whooshes on cards and zooms, pops on list items and the hook impact are added automatically by the style's `efectos.densidad`.
5. **Music** (optional): pick ONE track from `sonidos/musica.json` whose mood fits the content (`animo`, `uso`; energy 4-5 for fast montages, 2 for calm explanations). If `npm run musica` shows it's not downloaded, ask before downloading. The script picks the most energetic part, ducks it under the voice and fades it. Tell the user to paste `proyectos/<id>/credito.txt` in the post description (CC BY 4.0 requires it). Trending commercial songs can't be used here: they're added inside Instagram/TikTok.
5b. **Screen recordings** (optional, mostly horizontal): for each one run `npm run pantalla -- --proyecto <id> --video <path> --nombre p1` and look at the single grid image it makes to pick zoom zones (`{"desdeMs", "x", "y", "w", "h"}`, fractions). Then ALWAYS run `npm run sensibles -- --proyecto <id>`: OCR finds emails, phones, IDs, cards, bank accounts, API keys, IPs, whatever follows "password:" and the user's private phrases (`config.json` → `privado`). It prints only the types (and two letters) and leaves `sensibles.jpg` with each item boxed in red. **Ask the user** (one question): "I see some data you may not want to show: <types>. They are marked in `proyectos/<id>/sensibles.jpg`. Blur them all, none, or all except some?" Don't open that image yourself and never repeat the data. Save the answer with `npm run sensibles -- --proyecto <id> --difuminar si [--excepto ip,telefono]` or `--difuminar no`; `componer` waits for it. Blurred items follow scrolls and zooms. Add `"pantallas": [{"nombre": "p1", "desdeMs", "hastaMs", "inicioS", "zooms": [...]}]` to the decisions; you appear in a bubble meanwhile.
5c. **Sections** (optional): if the style has `cabecera.barra`, add `"secciones": [{"titulo": "Intro", "desdeMs": 0}, …]` (2-5 short titles, 1-2 words) and a bar at the top marks where the viewer is and fills up like a timer. Without them, the chapters are used. The voice wave (`cabecera.onda`) needs nothing.
**Format**: ask the user (unless they already said it) whether this video is for Reels/TikTok/Shorts (vertical) or YouTube (horizontal). Never assume it from the shape of the recording: people record vertical and publish on YouTube.
6. Write `proyectos/<id>/decisiones.json`: `{"formato": "vertical" | "horizontal", "gancho": {"texto": "...", "categoria": "..."}, "tarjetas": [...], "momentos": [...], "musica": {"pista": "wallpaper"} | null}`.
7. **Compose**: `npm run componer -- --proyecto <id>`. It places, in this order, the hook, the graphics, the cards and every caption page so they never overlap each other nor cover the face or the moving hands (with inertia so captions don't jump), and adds punch-in zooms toward the face where the shot stays still longer than the style's `maxPlanoMs`. Don't edit positions by hand.
8. **Render**: `npm run render -- --proyecto <id>`, then `/revisar`. To check one detail without rendering everything: `npm run render -- --proyecto <id> --capturas 120,340` (frame numbers) writes `captura-<n>.png`. If the review reports a sensitive datum still readable, don't deliver: re-run `npm run sensibles`, compose and render again (if it persists, add the phrase to `privado` or cut that part of the recording, and tell the user).
**Framing.** If the source shape differs from the output (vertical take → YouTube, horizontal → Reel), `componer` reframes automatically so the whole face stays in frame, with a blurred fill on the sides when needed. Nothing to decide here.
Wersje
W bazie nie ma innych skilli o nazwie editar.
Migawka bazy z 13.11.2026 · wersja paczki b2cdbbd4a078 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100frame-inspect
Wyodrębnia klatki z wideo i obrazów, renderuje Remotion, sprawdza specyfikację (rozdzielczość, fps, długość) i wspiera wizualną kontrolę jakości przez ffmpeg, ffprobe i skrypt video-tool.sh. - 100remotion-video-check
Sprawdza wizualnie każde wideo wyrenderowane w Remotion: pauzy, czarne klatki, szczyty dźwięku i nieaktualne renderowania, a potem poprawia błędy. - 100motion-video
Pisze kompozycję Remotion, renderuje ją do MP4 lub PNG lokalnie i sprawdza klatki, bez płatnego API. - 100qa
Kontroluje jakość renderów wideo w Remotion: kod mierzy każdą klatkę, a model Sage ocenia obraz i wybrane klatki. - 100remotion-motion
Buduje, podgląda i renderuje wideo w Remotion: konfiguracja projektu z szablonu, klatki stylu, animatik i finalne rendery przez kolejkę renderowania w studio. - 100promo-reel
Zamienia krótki brief w brandowe pionowe wideo promocyjne 15 lub 30 sekund w formacie 1080x1920 na szablonie Remotion.