Video generation gets the attention, and it is the least interesting part of the workflow. A model produces a shot; a plan produces a video. The plan is a shot list, a caption track, and a crop specification — three artefacts that can be produced before anything is generated.
This page covers the three skills in the directory that do that work — Video Shot List, Subtitle Timing, and Vertical Crop — and where they fit against the models.
The short answer
Write the shot list first, then generate or film one asset per shot, then caption, then crop. Doing it in that order means each step has a checkable output; doing it in reverse produces footage nobody can assemble.
Why a shot list is the first skill
A script is written to be read; a shot list is written to be produced. The conversion is mechanical once the rule is stated: one shot, one asset, one action, and a duration that fits the format.
Video Shot List enforces the one-asset rule because the alternative is a shot that needs three generations to look right — the first thing every video project discovers too late.
Captions are a timing problem
A transcript with no timecodes is not a caption track. Subtitle Timing produces cue-level timings from a transcript, with lines short enough to read and cues that start before the audio and end after it. The check is mechanical, which is why the skill is worth having: you can verify every cue in a script.
The failure mode it prevents is drift — subtitles that are right for thirty seconds and half a second behind for the rest of the clip.
The vertical crop
Horizontal footage does not become vertical by cropping the middle. Vertical Crop plans a crop per shot: where the subject sits, what the safe area is, and where to re-frame when two people are at opposite edges.
It produces a specification rather than a render, which is the honest boundary — a bad crop is obvious to a human in one second, and the spec is faster to fix than a re-encode.
The comparison that matters
| Step | Skill output | Model output |
|---|---|---|
| Shot list | Numbered shots, one asset each | — |
| Generation | — | Any single shot it can describe |
| Captions | Cues with timings | — |
| Crop | Per-shot spec | — |
| Assembly | — | — (an editor) |
Read the table as a division of labour. The models own one row; skills own three; the editor owns the last one and nobody automates it.
A worked brief
The brief: a thirty-second product clip, six shots, captions, vertical version for social.
The shot list gives six numbered shots with durations summing to thirty seconds. Generation produces six clips — expect a few retries on hands and text, and plan the retry budget. Captions are timed against the voiceover. The crop plan marks where the product sits in each shot, and the vertical version re-frames around it.
The order is the point. Skip the shot list and the retries are unstructured; skip the timing pass and the captions drift; skip the crop plan and the vertical version is unusable.
Where the skills stop
None of them judge whether the video is good. They produce production inputs that a person can accept or reject quickly, which is the useful split: taste stays human, structure gets automated.
Frequently asked questions
Can the shot list be written by a general assistant? Yes — the skill is the rule, not the model. The point is the one-asset constraint and the fixed output shape, which a prompt can also provide if you write it down.
Do I need these for a single clip? No. One shot is one thing; the overhead appears at six or more, which is when continuity starts to fail.
What about audio? Voiceover is a generation step, not these skills. See ElevenLabs MCP and video tools.
Last checked: 11 November 2026.
Related: video generation after the quality race, briefing a video model, and video tools.
