agentleFS
FluffyattackFluffyattack/edit-bay/skills/editing-videos/SKILL.md

editing-videos

Skill0 starsChanged 20 days agoMITView on GitHub

Edits videos end to end with ffmpeg and Python instead of an editing app: local word-level transcription, cut lists, frame-accurate rough cuts, word-by-word captions, overlays, B-roll, sound effects, music, previews and final export, in gated stages (brief, rough cut, story lock, design pass, preview, picture lock, finish). Keeps a persistent edit bay of reference notes, playbooks and LEARNINGS.md so every video applies what earlier notes taught. Use when the user wants to cut, trim, caption, subtitle, add graphics or B-roll to, or render a video, reel, TikTok, YouTube Short, vlog or talking-head recording; remove pauses, filler words or retakes; give timestamped notes on an edit; or collect style references for future videos.

  • Installs packages

2 other projects use a version of this.

Adopted by

Most recently taken up by

  • Other versions2 projects use a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 2

    1. kmilo93sd/agent-video-editor★ 4same name
    2. F0rger123/claude-video-editing-god-mode★ 1same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

The file

We split this skill into the rules it gives an agent

11 rules. About 0 are common, 5 are distinctive, and 6 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 6

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Notes become rules: Every correction that should hold in future is logged, so the next video starts ahead.

    line 21
  2. Distinctive:

    All projects live in one folder, the edit bay.

    line 30
  3. Very specific:

    If an edit bay exists, read its CLAUDE.md, LEARNINGS.md, the matching playbook and the reference notes named in the brief before cutting …

    line 37
  4. Distinctive:

    Never skip a lock: everything later is built on it.

    line 56
  5. Very specific:

    Brief: Ask before cutting: the one message, audience, platform and aspect ratio, length, which references apply (and for which aspect), …

    line 59
  6. Very specific:

    If it reports problems, fix the plan and run it again.

    line 78
  7. Very specific:

    When a note is vague ("make it punchier"), don't guess.

    line 84
  8. Distinctive:

    After each round of notes, sort them:

    line 91
  9. Distinctive:

    If it's unclear whether a note is universal, ask — a wrong rule damages every future video.

    line 97
  10. Very specific:

    When someone shares a video they like, write a reference note (guides/edit-bay.md): download a 720p copy with yt-dlp -f …

    line 112
  11. Very specific:

    Read it before writing overlays.json or editing a cut list by hand.

    line 121

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Editing videos

Claude edits by writing plans and running scripts: transcript with word timings → cut list → ffmpeg. The person decides; Claude plans, runs the tools and checks the results. Two habits keep the results personal instead of generic:

  1. Personal style is written down in reference notes that say what exactly works in a video someone likes.
  2. Notes become rules. Every correction that should hold in future is logged, so the next video starts ahead.

SKILL_DIR below means this skill's directory. Scripts need Python 3.10+ and ffmpeg (with libass and libx264); transcribe.py also needs pip install faster-whisper, graphics usually pip install pillow. Check what is missing first (ffmpeg -version, python -c "import faster_whisper") and tell the person what to install — don't install system packages unasked.

The edit bay

All projects live in one folder, the edit bay. If none exists, ask where it should go, then run:

python SKILL_DIR/scripts/create_edit_bay.py ~/edit-bay
python SKILL_DIR/scripts/new_project.py ~/edit-bay "Desk setup reel"

If an edit bay exists, read its CLAUDE.md, LEARNINGS.md, the matching playbook and the reference notes named in the brief before cutting anything. Layout, reference notes and how rules mature: guides/edit-bay.md.

Workflow

Copy this checklist into the conversation and tick it off as you go:

Edit progress:
- [ ] 1. Brief agreed (brief.md)
- [ ] 2. Footage reviewed (transcripts, contact sheets)
- [ ] 3. Rough cut rendered → STORY LOCK from the person
- [ ] 4. Design pass: storyboard → stills → test clips
- [ ] 5. Preview rendered, notes applied → PICTURE LOCK from the person
- [ ] 6. Final exported and checked
- [ ] 7. Learnings proposed and saved

Each stage is cheaper to change than the next, so fix things early and render full resolution once. Never skip a lock: everything later is built on it. Going back to an earlier stage is always fine. Commands, options and checks for every stage: guides/stages.md.

  1. Brief. Ask before cutting: the one message, audience, platform and aspect ratio, length, which references apply (and for which aspect), scripted or unscripted. Write brief.md.
  2. Footage review. transcribe.py for each recording; contact_sheet.py to see the picture (Claude can't play video but can read images). Scripted: check every line was said and find duplicate takes. Unscripted: find the strongest moments and propose an order. Fix misheard names in the transcript before going on.
  3. Rough cut. suggest_cuts.py removes pauses and fillers and flags retakes. Decide which clips stay and in what order (usually the last complete take wins) and say why. check_plan.py, then render_rough_cut.py. The person reviews content only — message, order, pace, gaps. Story first, design later.
  4. Design pass, in this order, each with a round of notes: storyboard.md (text only) → stills (render.py --stills) → short test clips (render.py --from --to) for anything that moves. Look at every still yourself before showing it. How captions, graphics, B-roll and sound are made: guides/design-with-code.md.
  5. Preview. Whole video at 720p, version in the file name (preview-v2.mp4) so every timecode in a note is unambiguous. Suggest one uninterrupted viewing first, then a second pass noting timecodes.
  6. Finish. Re-render the rough cut with --full-resolution, then render.py --final with the same captions and overlays. Verify with ffprobe and a few stills from the final file.
  7. Learnings. See below.

Validate before every render: python SKILL_DIR/scripts/check_plan.py <cut_list.json | overlays.json>. If it reports problems, fix the plan and run it again; render only when it prints OK.

Notes that work

A useful note has three parts: where (version + timecode), what you see, what you want — for example: "preview v2 at 0:12 — the price card hides the product — move it to the top-left." When a note is vague ("make it punchier"), don't guess. Ask where, or offer two or three concrete readings (faster cuts? punch-in zooms on key words? a sound on each card?). Log every note in the project's notes.md.

Timecodes refer to the named preview version; *.words.json → clips maps rough-cut time back to source time.

Turning notes into rules

After each round of notes, sort them:

  • This video only ("move the card at 0:12") → notes.md, nothing else.
  • Every video ("cards never cover the product", "max three words per caption") → an entry in LEARNINGS.md (date, rule, why, origin); proven rules move into the matching playbook.

If it's unclear whether a note is universal, ask — a wrong rule damages every future video. Write rules so they can be checked (numbers, positions, durations). If a new note contradicts an existing rule, ask whether the rule should change instead of silently making an exception.

At the end of a project, list the learnings found in this chat, ask which to save, save them, and tick "Learnings saved" in brief.md.

Render budget

Rendering costs time (and, for every round, tokens). Previews at 720p, test clips instead of whole videos, full resolution exactly once. On slow machines use the small Whisper model or smaller and run long jobs in the background with progress updates; faster-whisper uses a CUDA GPU automatically when present.

References

When someone shares a video they like, write a reference note (guides/edit-bay.md): download a 720p copy with yt-dlp -f "bv*[height<=720]+ba/b[height<=720]" -o "references/<name>/video-720p.%(ext)s" <url>, make a contact sheet, and agree together on what works, area by area, with timecodes. Other people's footage is for analysis only and never goes into the person's videos. Mix single elements from several references so the result looks like the person, not like one particular creator.

File formats

Transcript, cut list, words file and overlays plan are JSON: guides/file-formats.md. Read it before writing overlays.json or editing a cut list by hand.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.