Instagram Trend Music Video Maker · skill pública de Open Skill
Instagram Trend Music Video Maker
Instagram Trend Music Video Maker es una Open Skill pública en Digen. Ábrela para ver qué crea este flujo.
Film
7 1 6
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3872514/2026-10-20/thtyku7q.jpg
---
allowed-tools:
- search_web
- generate_music
- generate_image
- generate_video
- recognize_image
- recognize_video
- probe_media
- extract_clip
- extract_frame
- mix_clips
---
# Instagram Trend Music Video Maker (v2)
Vertical 9:16 Reels dance MV, 10–20 s, hook in the first second, original audio in the trending style.
## When to use
- The user asks for a trending Reels / TikTok dance video, an "AI music video for Instagram", or a trend-based MV.
- The user wants to redo, tweak, or continue an MV previously built with this skill.
## Inputs
- **Character** — default: an original art-directable dancer (state the default, keep it swappable).
- **Vibe** — fast/energetic (default) or calmer.
- **Length** — default 12–14 s.
- **Cover** — optional uploaded poster/photo for the Reels cover.
- **Song/trend** — optional; match genre + tempo, never the master recording.
Ask only for what is missing and blocking; otherwise default and proceed.
## Hard limits (verified, do not guess)
- `ai2v` (image + driving audio): max **14 s**, output **704×1280** 9:16, **24 fps**. Only channel that keeps dance moves locked to the beat.
- `i2v` / `ref2v` / `t2v` / `fl2v`: 10 s max, ≤480×864 — too soft for Reels.
- `generate_music`: returns ~20–23 s → must be trimmed.
- `extract_clip(accurate=true)`: frame-accurate; use for every surgical cut.
- `mix_clips`: **appends a ~2 s platform end card** (DIGEN/AGENT) — always trim before delivery.
## Steps
First use `write_todos` to list the stages, then update the list as you progress.
1. **Trend research** — `search_web` for "trending Instagram Reels songs this week". Pick a choreography-friendly one (chant hook, clear beat, gesture rhythm). Note title, artist, genre, BPM, and the gesture the trend is built on. **Copyright:** never reproduce the track. Make an *original* song in that style, with original lyrics and an original title. Never imply it is the licensed trend track.
2. **Track** — `generate_music(model='music', entity_type='music', name=<original title>)`. Prompt and lyrics per `references/prompt-templates.md`. Then `probe_media` for the real duration.
3. **Character bible (clean plate)** — `generate_image(model='t2i', orientation='portrait', entity_type='characters', name='Character Plate')`. Must include: `Absolutely NO text, NO letters, NO signage, NO writing, NO logos, NO brand marks anywhere in frame.` Lock age, skin tone, hair, outfit pieces, palette, lens, lighting direction.
4. **Angle ladder** — 2 stills with `generate_image(model='i2i', input_urls=[base_still], orientation='portrait', entity_type='characters')`: medium close-up and low-angle full body. Prefix "Same exact character, same outfit, location and lighting, but reframe as…". QC each with `recognize_image` (identity, text, anatomy).
5. **Takes** — `generate_video(model='ai2v', input_urls=[<angle still>], audio_url=<track>, duration=14, resolution='720P', orientation='portrait', entity_type='mv_clips')`. Prompt must say: `ONE single continuous unbroken shot, no cuts, no scene changes`, choreography in beat units, the camera move, and `no text, no logos`. **One take at a time** — parallel batches exhaust credits and half-fail.
6. **QC** — `recognize_video` asking for: identity stability, weighted vs floaty motion, **ghosted/double-exposure window**, hidden cuts, on-screen text, first second, final pose. Demand timestamps; confirm any suspect window twice with a tight `extract_clip` probe (±1 s disagreement is normal). Use `extract_frame` to pull stills from the suspect window if the text report is ambiguous.
7. **Seam surgery** — remove the ghost with `extract_clip(accurate=true)`: `[{start:0,end:ghost_start},{start:ghost_end,end:total}]`. Prefer boundaries on whole beats (112 BPM = 0.536 s/beat, 4 beats = 2.143 s). A hard cut on a beat is invisible; a ghosted blend reads as an AI defect.
8. **Assemble + de-brand** — `mix_clips(orientation='portrait', fps=30, crf=18, preset='slow', entity_type='mv_clip