ScenarioExplorations Eight ways to edit

A video style library · Eight reusable skills

Eight ways to edit.

One finished process, eight editing styles.
Same material, a different video every time.

Each style is a skill an AI agent follows to edit a video: its own tone, its own idea of how to explain, and its own editing rules for pace, cuts, camera, lettering, voice and music. We gave all eight the same finished work and the same facts. Pick a style, then play two side by side and watch the edit change the story.

01 · The library

Choose a style. Then pick a second.

Same input, same facts, about a minute each. What changes is everything an editor decides.

Calm

calm-minimal-video

Calm

    How it edits

    What drives the pace
    Cuts and camera
    Lettering
    Voice
    Music and sound
    Footage
    Ending

    Editing choices to notice · click to jump there

      02 · Get the skills

      Every style is a skill. Take the library.

      Each style on this page is an agent skill: a folder with the editing rules, a small render engine, its fonts and the tools that time the words and check the mix. An AI agent such as Claude Code reads it and turns your own finished process into a video in that style.

      Full library · ZIP · 1.8 MBAll ten skillsThe router that picks a style, the eight styles shown here and a ninth, ASMR sensory, that was not part of this test.Download the library

      Or one style at a time · each style works on its own

      What you need

      • An AI agent that runs skills, such as Claude Code.
      • Node 20 or newer, with ffmpeg and ffprobe on the PATH.
      • Google Chrome, or Playwright’s Chromium, to render the frames.
      • Python 3.10 or newer, run through uv (or pip with each skill’s requirements.txt).
      • A Scenario account with the Scenario MCP connected to your agent, for the voice, the music and the sounds (and the song and footage for hype). These steps use credits; the skills price each one first.

      How to install

      1. Unzip the library into your agent’s skills folder. For Claude Code that is ~/.claude/skills/, or .claude/skills/ inside a project. Keep the folders side by side.
      2. Connect the Scenario MCP to the agent.
      3. Ask for a video of something you finished, for example “make a calm video of this experiment”. The router asks what you made, gathers the proof and helps you pick a style.

      Tested on Windows. The skills are written for macOS too, but have not been tested there yet. The hype style is adapted from the music video skill of the Every Scenario exploration.

      03 · All at once

      Eight rhythms, one clock.

      All eight videos in one silent grid, started together. Without the sound you see the edit itself: how often each one cuts, where the camera moves and when the type lands.

      04 · Side by side

      Same story, different clocks.

      StyleSkillVoiceLengthWords a minuteVoice shareTempo
      Calmcalm-minimal-videoSulafat57.5 s7437%ambient bed
      Kinetickinetic-tech-showcasePuck60 s10752%120 BPM
      Documentarydocumentary-process-videoCallirrhoe61.6 s13059%90 BPM
      Trailercinematic-trailer-videoAlgenib60 s4320%120 BPM cue
      Tutorialtutorial-walkthrough-videoZubenelgenubi55.5 s12962%100 BPM
      Diagramdiagram-explainer-videoFenrir60 s13263%100 BPM
      Paper cut-outpaper-cutout-videoSulafat56.5 s12259%100 BPM
      Hypehype-music-videoSung50 s8874%127 BPM

      Measured on each final mix: words a minute and the share of the running time with a voice (sung, for the hype video). Tempo is the music bed as analyzed; the calm bed is ambient. Every mix lands between -14.8 and -14.1 LUFS.

      05 · What every style shares

      Different edits, one set of rules.

      The styles disagree on almost everything. Underneath, they run on the same spine.

      1. Natural narration. Text written as speech: short clauses, a breath at the start of each take and no stage directions. Two takes per section, chosen by measured pitch range and pace.
      2. Word-timed lettering. Each take is transcribed and matched to its script, so every word has an onset. Cards, rings, cut letters and paper pieces land on those onsets.
      3. Music fitted to the voice. In the narrated styles, a bed generated in sections that follow the script, then cut and placed around the narration. The hype video is the song itself.
      4. A measured mix. Loudness, true peak, voice over music, pace and intelligibility are measured on every final mix.
      5. Every claim on screen. Nothing is said that is not visible at that moment: the overlay under “ninety percent”, the closed eye under “closed”.
      6. The mono Scenario logo. Every video ends on the Scenario logo in one color, black on light pages and white on dark ones.

      06 · The test footage

      Same input for every style.

      To make the comparison fair, all eight videos tell one real, finished process with one set of records: four pencil sketches turned into finished renders in Scenario, each sketch the reference image for one detailed prompt, as in the Help Center’s Sketch-to-Render. The harbor is on GPT Image 2.5 Sunburst (OpenAI); the inventor, dragon and scooter are on Gemini 3.1 Flash (Google). Drag across it.

      The harbor render: a stone fishermen’s street at blue hour running down to a quay, two moored white boats and a lit lighthouse
      The harbor pencil sketch: the same street in graphite lines on white paper
      Pencil sketch Render

      90% of the sketch’s lines land within 2 px of an edge in the render

      By chance, against renders of other subjects: 8 to 14%. GPT Image 2.5 Sunburst (OpenAI), with a strict line-for-line prompt. The first harbor, a looser prompt on Gemini 3.1 Flash, scored 71%.

      The harbor three times with the sketch’s lines in cyan on top: the pencil sketch, the first render on Gemini 3.1 Flash with the boats and lighthouse off the lines, and the GPT Image 2.5 Sunburst render with them on the lines
      Pencil sketchGemini 3.1 Flash, first prompt: 71% within 2 pxGPT Image 2.5 Sunburst, stricter prompt: 90%

      The first harbor let the boats and the lighthouse drift off the pencil lines. A stricter, line-for-line prompt on another model put them back, so that render became the test footage. The prompts differ and each ran once: two real attempts, not a ranking of models.

      The dragon’s eye three times: half open in the pencil sketch, fully closed in the first render, half open again in the second render
      Pencil sketch: half openFirst try: closedSecond try: half open

      The one real fix, in every video

      The first dragon prompt mentioned the eye only in passing, at the end of a list of colors and materials, and the eye came back closed. The second prompt gave it its own instruction and it came back half open. Each style tells this same failure and fix in its own way.

      The second prompt’s new instruction, word for word:

      “Keep the dragon's expression from the sketch: its visible eye is HALF OPEN, the upper eyelid lowered halfway over a glowing golden eye with a slit pupil, gazing sleepily toward the viewer. The eye must not be closed.”

      Models

      Who made what, all run on Scenario.

      Have a finished process? Generate the voice, the music and the footage on Scenario, then edit it your way.

      Create yours

      How it was made

      One library, eight editors.

      Made with Claude Code working through the Scenario MCP. Each style is a written skill: its dials, its grammar and a worked example. An agent following one skill turns a finished process into one video, built frame by frame in code around real footage.

      1. 1

        One evidence pack

        A real process with its records: four pencil sketches, seven renders with the sketch as the reference image, one same-prompt model comparison, five clips. Every prompt and job time was kept.

        Footage Gemini 3.1 Flash (Google) · GPT Image 2.5 Sunburst (OpenAI) · GPT Image 2 (OpenAI) · Seedance 2.0 Mini (ByteDance)

      2. 2

        One brief for eight styles

        The same subject, facts and length for all. Only what the pack shows may be said, and every claim must be on screen when it is said.

      3. 3

        A script per style

        Each skill decides how to explain: a quiet recap, proof on the beat, a diary, a trailer, a walkthrough, a diagram, paper pieces or a song.

      4. 4

        Voice and music

        Narration cast per style, two takes per section; word onsets measured from the chosen take; a bed generated to fit the script; a sound kit per style.

        Voice Gemini 3.8 Flash TTS · Music ElevenLabs Music v2.5, MiniMax Music 3.0 · Sound Sonilo

      5. 5

        The edit, in code

        Scenes rendered frame by frame in a browser, lettering placed on the word onsets, mixed to a measured loudness, ending on the mono logo.

        Masks SAM 3.1 Image (Meta) for the paper cut-out

      6. 6

        Check before shipping

        A word sheet for every claim, a loudness and intelligibility report, a freeze check and a contact sheet every half second.

      Good to know

      • The styles are skills for an AI agent, downloadable from Get the skills. They were tested on Windows; macOS is not tested yet.
      • The tutorial and kinetic videos use the Help Center’s own screen recording of Edit with Prompts, with the author’s email blurred.
      • The test sketches were generated, not hand drawn.

      Tip

      Pick the style by the goal: kinetic for a launch, calm for a portfolio, documentary for a postmortem, the trailer for a reveal and a tutorial for a tool.

      The library

      8 styles

      Calm, kinetic, documentary, trailer, tutorial, diagram explainer, paper cut-out and hype music video.

      PDF · Illustrated

      The making-of

      Read the making-of

      Help Center · The test case

      Sketch-to-Render

      Read the article

      Scenario

      Make yours

      Start creating