AGENT SKILL

trove-video-edit

Edit a pre-cut talking-head video: captions, overlays, zooms, and a music bed. Use when the cut is already made and the edit is what remains. Not for raw multi-take footage and not for finding b-roll in a clip.

License 许可
MIT
所属仓库
bravew/trove
最近更新
2026-11-12
Stars
★ 10
Forks
1

01 前言 Front Matter

name
trove-video-edit
license
MIT
metadata
{ "inspired-by": "BuildLoop video skills", "trove-category": "designs-media", "trove-category-label": "Designs & media" }
description
Edit a pre-cut talking-head video: captions, overlays, zooms, and a music bed. Use when the cut is already made and the edit is what remains. Not for raw multi-take footage and not for finding b-roll in a clip.
when_to_use
edit this pre-cut video; add captions and overlays to this cut; one-shot edit this intro; add music and zooms to this talking head
user-invocable
true

02 资源文件 10

03 SKILL.md 正文

<!-- AUTO-GENERATED from SKILL.md.tmpl — do not edit directly --> <!-- Regenerate: bun run build:skills -->

Video edit

A pre-cut talking-head clip gets captions, overlays, punch-in zooms, and a music bed. The cut already exists. Choosing takes belongs to trove-take-cut, finding b-roll in raw footage belongs to trove-broll-ingest, and checking the finished piece belongs to trove-video-review.

Everything on screen comes from one file, edit_plan.json. If it is not in the plan it does not get rendered. That is what keeps the edit reviewable: the plan is the edit, and the render is a projection of it.

Modes

Probe with trove-ffmpeg's probe.py and pick from the geometry.

Shorts Intro Longform
Shape Portrait, 9:16 Landscape, up to 30 seconds Landscape, longer
Density A few beats, hand placed A few beats One every half minute or so, per section

Read references/runtime-spec.md in trove-ffmpeg before the first command. Never write a raw ffmpeg command.

The plan

  1. Transcribe to word level and write words.json in the shape trove-take-cut defines. The transcript is the timing source for every beat.

  2. Correct it yourself. Read the transcript against the script when the user supplied one, and against what you can hear otherwise. Fix names and terms the recognizer got wrong and write each correction down. There is no second model and no API call for this. You are the pass.

  3. Author the beats. One per moment that earns it. Each names the words it lands on (anchor, copied from the transcript) and what the viewer understands from it that the words alone do not carry (reason). A reason of "this adds energy" is not a reason. Drop that beat. The kinds and when to use each are in overlays.md. The file matches edit-plan.schema.json.

  4. Lint before anything renders. Run scripts/lint_plan.py with the probed duration, and with --fps when the composition renders at a rate other than its default of 30. The plan carries no frame rate. A failing plan does not render. Fix the plan.

Music, sound effects, and images come from directories the user supplies. This skill ships none, fetches none, and generates none. A moment with no asset gets no image beat.

Preview, then final

Render a preview first: a short, low-resolution pass that is cheap to throw away. Show the user the file and stop. The final render happens only after they approve the preview in this session. Re-renders after feedback stay at preview until they approve again.

State once, before the first render, that Remotion is free for individuals and small teams, that companies may need a paid license, and link the Remotion license. Render SDR finals with --color-space bt709. Remotion then converts the pixels to limited range and tags the matrix, primaries, and transfer as BT.709. The composition lives in assets/remotion/ and reads the plan. It renders the speaker and the overlays as separate layers, so a zoom on the speaker never crops an overlay. Copy the clip into that directory's public/ folder and name it in plan.source as a bare filename. The composition resolves it with staticFile, which is the only path form Remotion's renderer can fetch.

Audio

Score the approved render, not the preview. Duck the music under the voice and normalize loudness with trove-ffmpeg's audio.py and loudness.py. When the user supplied no music, the cut ships silent rather than with a bed they did not choose. Then hand the result to trove-video-review.

What this skill does not decide

Colors, type, and pacing beyond the rules in overlays.md come from the user's project. There is no default brand, no default music, and no default call to action. Ask when a choice would change the result, and leave the moment unadorned when there is nothing to ask.

04 同仓库更多 Skills