AGENT SKILL

trove-broll-ingest

Find the usable moments in raw footage, cut each one, grade it, and add it to a b-roll library. Use when a raw clip or a folder of raws needs selecting, grading, and cataloguing. Not for a talking-head script and not for a finished edit.

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-broll-ingest
license
MIT
metadata
{ "inspired-by": "BuildLoop video skills", "trove-category": "designs-media", "trove-category-label": "Designs & media" }
description
Find the usable moments in raw footage, cut each one, grade it, and add it to a b-roll library. Use when a raw clip or a folder of raws needs selecting, grading, and cataloguing. Not for a talking-head script and not for a finished edit.
when_to_use
find the usable moments in this clip; ingest this footage into my b-roll library; grade and catalog this raw clip; pull the good shots out of this footage
user-invocable
true

02 资源文件 8

03 SKILL.md 正文

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

B-roll ingest

A raw clip becomes a handful of graded selects in a catalog, each one a moment an edit can drop in. Cutting a talking-head script belongs to trove-take-cut. Finishing an edit belongs to trove-video-edit.

The library root comes from the user. There is no default path. If they did not name one, ask, and do not create a library anywhere they did not point at.

Grade from the transfer, not from habit

Probe the clip with trove-ffmpeg's probe.py and read the transfer characteristics before choosing a LUT. S-Log3 footage graded with a Rec.709 LUT comes out flat and wrong, and the mistake is invisible until the shot sits next to a correctly graded one.

Probed transfer LUT
Rec.709, BT.709 assets/luts/cinematic-rec709.cube
S-Log3, S-Gamut3 assets/luts/cinematic-slog3.cube
Anything else Stop and ask. Do not guess a grade.

Each has a -soft variant for a lighter grade. The user can pass their own .cube, which overrides the choice entirely. Apply it with color.py --lut.

Workflow

Read references/runtime-spec.md in trove-ffmpeg before the first command. Never write a raw ffmpeg command. The source clip is never modified.

  1. Probe. Transfer characteristics decide the grade. Note the frame rate and duration for the cuts.

  2. See the whole clip. Build a timestamped contact sheet with look.py and read every frame of it. For a clip where the action is easy to miss, build a second sheet over just that span. The sheet is how you find moments. Scene detection and the filename are not.

  3. Name the moments. Read selection-rules.md and write one sentence per select naming what happens and at what second. Most clips yield none to four. Show the list to the user before cutting.

  4. Cut and grade. cut.py --accurate for the span, then color.py --lut with the LUT chosen above. Exposure or white balance problems get color.py --correct before the grade, not instead of it. Name the file <category>_<moment>_<source>.mp4 inside the library.

  5. Check the opening frame. Extract the first frame with look.py --at and look at it. If it shows anything other than the action already happening, the span starts too early. Re-cut.

  6. Shake, only when it helps. Run scripts/shake.py on a shot that looks unsteady. Stabilize a locked shot through trove-ffmpeg, score the result, and keep it only when the score dropped. Never stabilize a pan, tilt, or any move the camera made on purpose. Record the decision either way.

  7. Catalog. Append the select with scripts/catalog.py, pointing --library at the root the user named. The entry matches catalog.schema.json.

The LUTs

The four .cube files under assets/luts/ are a cinematic grade carried over from the BuildLoop video skills, kept byte for byte and renamed. They are a starting grade, not a house look. A project with its own grade should pass it.

04 同仓库更多 Skills