Scenario Video Assembly
Overview
Scenario has no editing tools on the MCP surface. Every compositor, concatenator, trimmer and
captioner is a deterministic model run through model_run. The compositor, concatenator and trimmer
ids are fixed and named below; Scenario ships two captioners, so that one is discovered with
recommend. Connection and the core loop: see the scenario skill; the clips themselves: see
scenario-video. If a sibling skill named here is missing from your available skills, load it with
skill_get (Scenario MCP) or ask the user to install it
(https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas and flag
the gap.
Quick reference: pick the backend first
| Need |
Model |
| Clips end to end, optional transitions |
model_scenario-video-concat |
| Anything overlapping: overlay, bed, titles |
model_scenario-compose-video |
| Still layout (thumbnails, sheets, key art) |
model_scenario-compose-image |
These ids are constants: each is Scenario's single deterministic tool for its operation, so
discovery would only re-derive it.
Concat is sequential: videos takes 2 to 50 files (never 1), preserveAudio (default true) carries
each clip's own track into the cut, and the optional transitions array's "length must be number of
videos - 1". Each entry is an object {type, duration}, never a bare string: type is one of 18
names read off the schema (fade, the default, is the plain crossfade; dissolve is a different
effect; none a hard cut) and duration runs 0.1 to 5 seconds, default 0.5. Omit the array for
hard cuts throughout, the right call whenever the brief does not ask one shot to flow into the next.
preserveAudio is not a guarantee of sync: a concatenated cut has come back with its audio drifting
against the picture, with and without transitions (a confirmed defect at authoring time). A concat
of clips carrying dialogue or effects is verified, never trusted: asset_get the output and compare
properties.duration with the inputs' sum (each transition shortens the cut by up to its
duration), then download it and check that the audio and video streams run the same length (a
local probe of the download is inspection, not editing) and that a visible cue (a cut, a mouth
opening, an impact) lands on its sound. When speech must stay locked to the picture, assemble in
Video Studio instead: each clip layer carries its own track at its startTime, so a drift
introduced by joining has nowhere to arise, at the price of hard cuts or per-layer fades in place of
the transition set. Concat stays the tool for silent clips and for cuts whose bed is laid
afterwards.
Video Studio is an absolute timeline: layers holds 1 to 50 image, video or audio sources placed by
startTime and stacked by zIndex (higher in front), with per-layer fadeIn/fadeOut rather than
a between-clip array, so a cross-fade between two clips belongs in concat. At least one layer must
be a video, so a music bed over a still is rejected.
The Video Studio contract
Per layer, three timing ideas have confusable names:
startTime and endTime place the layer on the timeline.
trimStart and trimEnd are amounts shaved off the head and tail of the source, not in and out
points: to play the first 20 seconds of a 90-second track, set trimEnd: 70.
duration overrides the layer's length. The top-level duration is a different field, setting
the whole composition, and applies only when durationMode is "custom".
Everything is in seconds at step: 0.1, so do not promise frame-accurate cuts. x and y are
strings taking pixels, percentages or words, and anchor says which point of the layer they
address: top-left, top-center, top-right, center-left, center, center-right,
bottom-left, bottom-center, bottom-right (the middle one is bare center), defaulting to
"top-left". A bottom-right logo is x: "right", y: "bottom", anchor: "bottom-right", flush to
the edge unless you inset it with a percentage.
canvasMode and durationMode both default to "auto", computed from the layers: a custom frame
is canvasMode: "custom" plus numeric canvasWidth and canvasHeight (there are no top-level
width/height, and unlike the layers' string width/height these are numbers), and a custom
length is durationMode: "custom" plus the top-level duration. fit only acts on a layer that
also sets the layer's own width and height, strings like x and y: to reframe a clip, set
both to the canvas size as bare pixel numbers in strings (width: "1920", height: "1080", the
form of the schema's own "0" default; a percent string also works) and pass "cover" (crops) or
"contain" (letterboxes), since the default "fill" stretches. Give an overlay layer an explicit
width and height even at its native size: leaving them empty is documented as preserving the
original but does not reliably, and a 560x80 strip came back at 1792x256 on a 1920x1080 canvas. The
output field is videoOutputFormat here, imageOutputFormat in Image Studio, and outputFormat in
concat and the trim utilities.
A layer's type (image, video, audio, or mask) is optional and otherwise inferred from source.
There is no text layer, so titles, lower thirds and end cards arrive as rendered PNGs composited as
image layers; scenario-text-overlay renders them letter-perfect.
Worked example: three clips, a music bed, and captions
- For every source,
asset_get and read properties: the real duration (audio carries it too),
frameRate on video, and width/height on video and images alike. A model asked for eight
seconds does not always return exactly eight.
- Add them up to get each clip's
startTime, then model_schema_get on
model_scenario-compose-video.
model_run it with layers holding the three clips at their computed startTimes
(zIndex: 0), a logo PNG pinned by x/y/anchor at zIndex: 1, and the music as an audio
layer trimmed to the clips' total with trimEnd, since auto durationMode computes from every
layer, the bed included. volume (0 to 2) and mute are per-layer and apply to video layers
too, so balance the bed from either side. Set canvasMode: "custom" with canvasWidth and
canvasHeight; fps defaults to 30 and takes whole numbers from 1 to 120, so set it to the
clips' measured frameRate when they agree (rounded), which is what step 1 read it for.
jobs_wait, then caption the finished cut rather than each clip: model_scenario-caption-studio
(an authoring-time pick, since Scenario ships two captioners: re-discover with recommend) takes
it as video and transcribes whatever audio the master actually carries, so lower the bed's
volume rather than muting the dialogue. For Caption Studio, follow scenario-caption-studio
for corrected-SRT input and rendered-caption verification rather than inferring an upload route
from a file-typed schema: the exposed upload_asset kinds do not include text. It burns captions
in by default, returns a sidecar when outputSrt is true, and positions them with textPosition
(top, middle or bottom, default bottom, so move it off a bottom-edge overlay). Other
captioners use their own live input contracts.
asset_display to review, then asset_download with format left unset (it is an image
conversion target) for the file.
For a revision handoff, retain the clean master, source assets, actual submitted layer payload,
measured timings, narration text and voice settings, title PNGs with their editable layout settings,
and any corrected SRT and returned caption theme. Keep output ids linked to the recipe. Revise the
requested layer and recaption from the clean master; do not stack new captions over a burned export.
These records support reconstruction, not a promise of a native editor project or a tested replay in
a new environment.
Trim, split and resize have their own tool models. Note model_scenario-video-cut's
startTime/endTime are source in and out points, not timeline placement.
Common mistakes
- Reaching for an MCP tool like
video_compose, or for local ffmpeg to do the edit itself. The
editing surface is model_run on tool models; probing a download is fine.
- Handing concat one clip, or one transition per clip: the minimum is 2 videos, and transitions run
one fewer.
- Trusting
dry_run here as validation: it returns a cost estimate only, and assembly payloads are
the most structurally complex in the catalog.
- Captioning a concatenated master before checking its sync: the captioner times its lines to the
audio it hears, so drifted audio ships drifted captions on top of it.
1---2name: scenario-video-assembly3description: "Use when generated clips must become a finished video on Scenario via MCP: cutting a shot list together, laying a timeline, concatenating with transitions, overlaying a logo or title card, adding a music bed or voiceover, burning in or exporting captions and subtitles, trimming and splitting, or producing vertical and square variants of one master. Keywords: edit, assemble, stitch, concat, compositor, timeline, layers, overlay, captions, SRT, ad, UGC, social cut."4license: MIT5metadata:6 source: https://skillmd.com/skills/scenario-labs/scenario-video-assembly7---89# Scenario Video Assembly1011## Overview1213Scenario has no editing tools on the MCP surface. Every compositor, concatenator, trimmer and14captioner is a deterministic model run through `model_run`. The compositor, concatenator and trimmer15ids are fixed and named below; Scenario ships two captioners, so that one is discovered with16`recommend`. Connection and the core loop: see the `scenario` skill; the clips themselves: see17`scenario-video`. If a sibling skill named here is missing from your available skills, load it with18`skill_get` (Scenario MCP) or ask the user to install it19(https://github.com/scenario-labs/skills#install); unattended, proceed from tool schemas and flag20the gap.2122## Quick reference: pick the backend first2324| Need | Model |25| ------------------------------------------ | ------------------------------ |26| Clips end to end, optional transitions | `model_scenario-video-concat` |27| Anything overlapping: overlay, bed, titles | `model_scenario-compose-video` |28| Still layout (thumbnails, sheets, key art) | `model_scenario-compose-image` |2930These ids are constants: each is Scenario's single deterministic tool for its operation, so31discovery would only re-derive it.3233Concat is sequential: `videos` takes 2 to 50 files (never 1), `preserveAudio` (default true) carries34each clip's own track into the cut, and the optional `transitions` array's "length must be number of35videos - 1". Each entry is an object `{type, duration}`, never a bare string: `type` is one of 1836names read off the schema (`fade`, the default, is the plain crossfade; `dissolve` is a different37effect; `none` a hard cut) and `duration` runs 0.1 to 5 seconds, default 0.5. Omit the array for38hard cuts throughout, the right call whenever the brief does not ask one shot to flow into the next.3940`preserveAudio` is not a guarantee of sync: a concatenated cut has come back with its audio drifting41against the picture, with and without transitions (a confirmed defect at authoring time). A concat42of clips carrying dialogue or effects is verified, never trusted: `asset_get` the output and compare43`properties.duration` with the inputs' sum (each transition shortens the cut by up to its44`duration`), then download it and check that the audio and video streams run the same length (a45local probe of the download is inspection, not editing) and that a visible cue (a cut, a mouth46opening, an impact) lands on its sound. When speech must stay locked to the picture, assemble in47Video Studio instead: each clip layer carries its own track at its `startTime`, so a drift48introduced by joining has nowhere to arise, at the price of hard cuts or per-layer fades in place of49the transition set. Concat stays the tool for silent clips and for cuts whose bed is laid50afterwards.5152Video Studio is an absolute timeline: `layers` holds 1 to 50 image, video or audio sources placed by53`startTime` and stacked by `zIndex` (higher in front), with per-layer `fadeIn`/`fadeOut` rather than54a between-clip array, so a cross-fade between two clips belongs in concat. At least one layer must55be a video, so a music bed over a still is rejected.5657## The Video Studio contract5859Per layer, three timing ideas have confusable names:6061- `startTime` and `endTime` place the layer on the timeline.62- `trimStart` and `trimEnd` are amounts shaved off the head and tail of the source, not in and out63 points: to play the first 20 seconds of a 90-second track, set `trimEnd: 70`.64- `duration` overrides the layer's length. The top-level `duration` is a different field, setting65 the whole composition, and applies only when `durationMode` is `"custom"`.6667Everything is in seconds at `step: 0.1`, so do not promise frame-accurate cuts. `x` and `y` are68strings taking pixels, percentages or words, and `anchor` says which point of the layer they69address: `top-left`, `top-center`, `top-right`, `center-left`, `center`, `center-right`,70`bottom-left`, `bottom-center`, `bottom-right` (the middle one is bare `center`), defaulting to71`"top-left"`. A bottom-right logo is `x: "right"`, `y: "bottom"`, `anchor: "bottom-right"`, flush to72the edge unless you inset it with a percentage.7374`canvasMode` and `durationMode` both default to `"auto"`, computed from the layers: a custom frame75is `canvasMode: "custom"` plus numeric `canvasWidth` and `canvasHeight` (there are no top-level76`width`/`height`, and unlike the layers' string `width`/`height` these are numbers), and a custom77length is `durationMode: "custom"` plus the top-level `duration`. `fit` only acts on a layer that78also sets the layer's own `width` and `height`, strings like `x` and `y`: to reframe a clip, set79both to the canvas size as bare pixel numbers in strings (`width: "1920"`, `height: "1080"`, the80form of the schema's own `"0"` default; a percent string also works) and pass `"cover"` (crops) or81`"contain"` (letterboxes), since the default `"fill"` stretches. Give an overlay layer an explicit82`width` and `height` even at its native size: leaving them empty is documented as preserving the83original but does not reliably, and a 560x80 strip came back at 1792x256 on a 1920x1080 canvas. The84output field is `videoOutputFormat` here, `imageOutputFormat` in Image Studio, and `outputFormat` in85concat and the trim utilities.8687A layer's `type` (image, video, audio, or mask) is optional and otherwise inferred from `source`.88There is no text layer, so titles, lower thirds and end cards arrive as rendered PNGs composited as89image layers; `scenario-text-overlay` renders them letter-perfect.9091## Worked example: three clips, a music bed, and captions92931. For every source, `asset_get` and read `properties`: the real `duration` (audio carries it too),94 `frameRate` on video, and `width`/`height` on video and images alike. A model asked for eight95 seconds does not always return exactly eight.962. Add them up to get each clip's `startTime`, then `model_schema_get` on97 `model_scenario-compose-video`.983. `model_run` it with `layers` holding the three clips at their computed `startTime`s99 (`zIndex: 0`), a logo PNG pinned by `x`/`y`/`anchor` at `zIndex: 1`, and the music as an audio100 layer trimmed to the clips' total with `trimEnd`, since auto `durationMode` computes from every101 layer, the bed included. `volume` (0 to 2) and `mute` are per-layer and apply to video layers102 too, so balance the bed from either side. Set `canvasMode: "custom"` with `canvasWidth` and103 `canvasHeight`; `fps` defaults to 30 and takes whole numbers from 1 to 120, so set it to the104 clips' measured `frameRate` when they agree (rounded), which is what step 1 read it for.1054. `jobs_wait`, then caption the finished cut rather than each clip: `model_scenario-caption-studio`106 (an authoring-time pick, since Scenario ships two captioners: re-discover with `recommend`) takes107 it as `video` and transcribes whatever audio the master actually carries, so lower the bed's108 `volume` rather than muting the dialogue. For Caption Studio, follow `scenario-caption-studio`109 for corrected-SRT input and rendered-caption verification rather than inferring an upload route110 from a file-typed schema: the exposed `upload_asset` kinds do not include text. It burns captions111 in by default, returns a sidecar when `outputSrt` is true, and positions them with `textPosition`112 (`top`, `middle` or `bottom`, default `bottom`, so move it off a bottom-edge overlay). Other113 captioners use their own live input contracts.1145. `asset_display` to review, then `asset_download` with `format` left unset (it is an image115 conversion target) for the file.116117For a revision handoff, retain the clean master, source assets, actual submitted layer payload,118measured timings, narration text and voice settings, title PNGs with their editable layout settings,119and any corrected SRT and returned caption theme. Keep output ids linked to the recipe. Revise the120requested layer and recaption from the clean master; do not stack new captions over a burned export.121These records support reconstruction, not a promise of a native editor project or a tested replay in122a new environment.123124Trim, split and resize have their own tool models. Note `model_scenario-video-cut`'s125`startTime`/`endTime` are source in and out points, not timeline placement.126127## Common mistakes128129- Reaching for an MCP tool like `video_compose`, or for local ffmpeg to do the edit itself. The130 editing surface is `model_run` on tool models; probing a download is fine.131- Handing concat one clip, or one transition per clip: the minimum is 2 videos, and transitions run132 one fewer.133- Trusting `dry_run` here as validation: it returns a cost estimate only, and assembly payloads are134 the most structurally complex in the catalog.135- Captioning a concatenated master before checking its sync: the captioner times its lines to the136 audio it hears, so drifted audio ships drifted captions on top of it.