Scenario / Explorations Auto Edit

One selfie take. One Claude skill.

Auto Edit.

Film it raw.
Get it back edited.

Opus 5.5 in Claude Code cut it, framed it and set the type. Through the Scenario MCP, Seedance 2.5 rebuilt the room with the lips still in sync, and ElevenLabs made the sound and the score.
Play the edit, then switch to the raw take.

Raw
AI Edit

AI Edit Type behind the head

The skill

Edit yours.

The skill hands your agent the whole pipeline: the cut, the cutout, the camera, the type, the Scenario MCP calls for the world, the sound and the score, and the finish. Point it at your own talking-head clip.

Download the skillscenario-auto-edit · ZIP Make yoursStart creating on Scenario ↗
  1. 1

    Connect the Scenario MCP

    Add it to Claude Code once. It runs the video, sound and music models for the skill. mcp.scenario.com ↗

  2. 2

    Unzip the skill

    Put the folder in your skills directory.

    ~/.claude/skills/scenario-auto-edit/
  3. 3

    Run it on your clip

    Open Claude Code next to the video and ask for the edit.

    /scenario-auto-edit raw.mov

Needs A Mac with Apple Silicon, Python 3 and ffmpeg. The cutout uses Apple Vision on your machine; the world, sound and music run through the Scenario MCP.

How it was made

One take, one skill, no timeline.

Made with Claude Code and the Scenario MCP. The input is a single 16-second selfie. Opus 5.5 writes and runs the edit as code: the cut, the cutout, the camera, the type and the finish. The Scenario MCP runs the generative parts: Seedance 2.5 rebuilds the world around the speaker, and ElevenLabs makes the sound and the score. The skill packs that pipeline so your agent can run it on your own clip.

  1. 1

    Cut to the words

    Every word is timed, then each cut point is moved to where the voice really starts. Every pause goes; no word is clipped. The cut is transcribed again to prove it.

    With Whisper · ffmpeg · Opus 5.5

  2. 2

    Lift the speaker off the room

    A person matte for every frame, made on the machine. It lets the type sit behind the head and the world change around the speaker.

    With Apple Vision

  3. 3

    Camera and type

    Sub-pixel punch-ins hide each jump cut, one smooth move per shot. Headlines set behind the head, captions word by word, motion blur at 24 frames.

    With Python · Opus 5.5

  4. 4

    Rebuild the world

    The cut clip goes to Seedance 2.5 as a reference video. It replaces the room with a studio set, a monitor wall and a blue-hour world, and keeps the lips in sync.

    With Seedance 2.5 via the Scenario MCP

  5. 5

    One character, every format

    On “everything”, one original character arrives as a sketch, key art, a sprite, a poster and a 3D model, as cards that fly in around the speaker.

    With GPT Image 2.5 Sunburst · Meshy 7.1 · Blender

  6. 6

    Score it, drop on the key word

    A score built in sections, its drop measured and placed on “Scenario”. Sound effects on the cuts and the reveals. The music ducks well under the voice.

    With ElevenLabs Music Advanced v2.5 · ElevenLabs Sound Effects 2 via the Scenario MCP

  7. 7

    Show the receipt

    On “Come”, the frame opens into an editor that shows this edit's real tracks and the one command that made it, then the logo holds.

    With Python · Opus 5.5

  8. 8

    Finish like a film

    Bloom, grain and a light chromatic pulse on the hits. The voice is cleaned and the master is levelled for social feeds.

    With Python · ffmpeg

What did not work

  • Sound effects synthesised in code sounded cheap. Generated ones replaced them.
  • The phone's HDR video, read as standard video, blew out the skin. Converting it first fixed the colour.
  • A montage of random generated worlds felt generic. One character in every format replaced it.
  • The music sat too loud for the voice, and the logo left too soon. Both were fixed from one note each.

Notes, not keyframes

Every change was a sentence: “the logo stays there too short”, “the music is a bit too loud”, “the cut before Come bothers me”.

Tip

Measure the music's drop first and put it on the word you want remembered. Then cut the rest to it.

Agent skill · ZIP

scenario-auto-edit

Download the skill

Scenario

Make yours

Start creating