agentleFS
vincentventalonvincentventalon/claude-code-video-editing-skill/skills/short-video/edit/SKILL.md

edit

Skill6 starsChanged 12 days agoMITView on GitHub

Edits a talking-head video from the raw clips to the video you post, in one command - joins the clips, cuts the silences, the false starts and the retakes (the last take wins, never inside a word), transcribes word by word (whisper.cpp on the computer, or ElevenLabs Scribe with the user's key), burns in subtitles word by word, renders 9:16 for TikTok and Instagram Reels in two views that take turns between sentences (full frame for the hook, then a split - a flat top zone over the camera in a rounded card - with the title and the boards on top and the subtitles on the seam), a title, boards on the words that say them (a figure, a list, a sentence), an end card. macOS first, Windows, Linux. Use when the user says "edit my video", "cut the silences", "remove my retakes", "add subtitles", "add captions", "make it a TikTok / a Reel / a Short", "add boards", "split screen", "monte ma vidéo", "coupe les blancs", "enlève mes reprises", "ajoute des sous-titres", "mets un titre sur ma vidéo", "mets des cartons", "fais-en un TikTok / un Reel", or drops raw clips to edit.

  • Reads credentials

18 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use daily-arxiv, exp-eval and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • LambenthanSkill

    daily-arxiv

    每日从 arXiv 拉取新论文,过滤相关性,auto-ingest 高优先级论文,检测 SOTA 更新

    Also in 5 projects that use this

  • LambenthanSkill

    exp-eval

    实验判决门:Review LLM 独立评判实验结果 → 4 种判决路径 → 自动更新 claims confidence、ideas status、graph edges

    Also in 5 projects that use this

  • skyllwtSkill

    exp-pilot-eval

    Pilot result evaluation — read pilot results, apply success criteria, update idea page (pilot_result, failure_reason if failed), generate PILOT_VERDICT_REPORT. Called by /ideate Phase 5 after /exp-pilot-run.

    Also in 4 projects that use this

  • skyllwtSkill

    exp-pilot-run

    Pilot experiment execution — read Pilot Spec YAML, write pilot code, run experiment(Confirm with the user before operation and require the applicant to conduct manual inspection), return results. Called by /ideate Phase 5. Does NOT modify wiki pages or judge pass/fail.

    Also in 4 projects that use this

  • phanghonghaoSkill

    poster

    HTML/CSS A4 海报生成。从沙龙/活动 Markdown 文件生成科技风 A4 海报 PDF/PNG。Use when user mentions: poster, 海报, 海报生成, make poster, generate poster, 活动海报, 沙龙海报.

    Also in 4 projects that use this

  • +6 more from Lambenthan
  • +1 more from skyllwt

The file

We split this skill into the rules it gives an agent

10 rules. 0 are common, 5 are distinctive, and 5 are very specific to this project.

  • 0

    Common

    Most of the 15 edited versions we read keep these.

  • 5

    Distinctive

    This author's own choices. Worth a look.

  • 5

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Write the file next to the video (<name>.boards.json), run again with --boards: it costs one render, not a new cut.

    line 114
  2. Distinctive:

    When one is missing, the script says it and does what it can:

    line 156
  3. Distinctive:

    no ffmpeg: it stops and says what to install.

    line 158
  4. Distinctive:

    no whisper-cli: silences only (no retake cut, no subtitles).

    line 162
  5. Very specific:

    no Node, or remotion off: ffmpeg draws the title, the boards and the end card (square boxes, a simple pop, no count-up).

    line 163
  6. Distinctive:

    Run the command. A 6 min rush takes about 2 to 3 minutes on a recent Mac.

    line 189
  7. Very specific:

    Read what it prints: duration before / after, the retakes it cut (each quotes the words), and the to check at 12.3s lines (a repeat it saw …

    line 190
  8. Very specific:

    Run edit_video.py again with the same clips and the same -o, and every option of the previous runs (--cut, --protect, --fix, --title, …

    line 219
  9. Distinctive:

    Keep the text away from the bottom fifth and the right edge of a 9:16: TikTok and Instagram put their buttons and the caption there.

    line 278
  10. Very specific:

    No subtitles on the video: --config says NO libass: the setup skill installs an ffmpeg with libass.

    line 298

“Common” is measured: kept by at least 6 in 10 of the 15 edited versions of this file we could read. “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

Edit a video

One command, one finished video:

python scripts/edit_video.py CLIP1.MOV CLIP2.MOV --lang fr --title "J'en ai *marre*"

(scripts/ is in this skill's folder; python3 on macOS / Linux if python is missing; on Windows py if python is missing or opens the Microsoft Store. On Windows the commands run in Git Bash or PowerShell: write paths with /, C:/Users/me/Downloads/clip.mp4, quoted when they have spaces.) It joins the clips in the order given, cuts what is not speech, the false starts and the retakes, transcribes, and renders, next to -o (default: ~/Videos/edited/<first clip name>.mp4, C:\Users\<you>\Videos\edited on Windows):

filewhat it is
<name>.mp4the video to post: 1080x1920 for TikTok and Reels, full frame and split taking turns (below), subtitles word by word, the title on top, the boards, the end card on the last seconds, the voice evened out to about -14 LUFS (the loudness the platforms play at)
<name>.cut.mp4the cut alone, as filmed, no text (for someone who finishes in their own editor)
<name>.srt, <name>.txtthe transcript of the video (the .srt can be uploaded as captions)
<name>.assthe subtitles as burned in, plain text, with the title, the boards and the end card when ffmpeg draws them (not Remotion)
<name>.takes.tsvevery island of sound, kept or cut and why, with its words
<name>.timeline.jsonthe kept passages and the words, with their times in the edited video; the views (views) and the boards placed (boards)
<name>.state.jsonthis video's state: the clips, the exact command of the last run (command), the steps done (steps.edit, then steps.publish written by the publish skill). The short-video skill reads it to pick the video up where it was left

*word* in the title = highlighted, " | " = a new line. An iPhone clip comes out upright, its HDR brought back to normal colours.

The two views (the scene)

The 9:16 has two views, and by default (--scene auto) they take turns:

  • full frame: the camera fills the screen (filmed sideways: in the middle on a blurred copy of itself; --fit fill crops it to fill). The title sits on top in a white box for its first seconds, the subtitles at the bottom of the middle third.
  • split: a flat top zone (the style's background, anthracite by default) over the camera in a card with rounded corners, on the lower 58 %. The title and the boards are in the top zone, the subtitles on the seam. Filmed upright, the card shows the slice of the clip that starts 7 % down (the head of a selfie, chin included); filmed sideways, the whole clip at the card's width, not cropped. This geometry is fixed.

The rule of auto, the one the script applies (plan_views in scripts/finish.py), and the one to explain when asked:

  1. The hook is full frame: from 0 to the end of the first sentence that ends after 3 s (in practice 3 to 5 s).
  2. A sentence with a board on it is split, the board in the top zone; a sentence with nothing to show is full frame.
  3. Never much more than 8 s without a change: a full-frame run past 8 s turns split at the next sentence (the title at the top); after 8 s of split, the next board comes over the camera, in a box, in full frame.
  4. Without boards, the two views take turns every 4 to 8 s, the split showing the title at the top. With neither title nor boards, full frame only (an empty top zone is not shown).
  5. Views only change between two sentences (after a . ! ? or a pause; inside a sentence longer than 8 s, after a comma or where a pause was cut), never under a subtitle line, never once the end card is up, and a view lasts 2 s at least.

--scene split or --scene full keeps one view for the whole video (split: the title stays at the top between the boards).

Boards

A board is a card drawn on the word that says it, in the top zone of the split (over the camera, in a box, in full frame). Three kinds, nothing more:

kindtextshows
stat"150 min | = 3 h of work"the figure, big, in the accent colour (it counts up when it starts with a number), its caption under it
list"Cut | Subtitles | Publish"a numbered list, one item per part, the items coming in one by one
text"What I love: *teaching*"a sentence, *word* highlighted, " | " = a new line

They go in a JSON file passed with --boards boards.json:

[
  {"say": "90 days", "text": "90 days | one video a day", "kind": "stat"},
  {"at": 21.7, "text": "Ten videos | 2.5 to 3 hours | Sunday, 5 pm", "kind": "list"}
]

say = the words that say it, as in <name>.txt (the board comes up on the first of them, the first time they are said after the previous board); at = the second in the edited video (<name>.timeline.json, words). A board stays to the end of its sentence (2.5 s at least), never past the next one; "until": 30.5 holds it longer. examples/boards.json is a real one, for the video whose transcript is examples/jen-ai-marre.txt.

When the user says "add boards" / "mets des cartons", or when a first render has a title but nothing else to show: read <name>.txt and propose the boards before rendering, as a short list (the words that trigger each, the kind, the text):

  • a stat for every figure that matters (a duration, a price, a count), a list when they enumerate (steps, tools, reasons), a text for the one sentence they would want read; their words, shortened, never a fact they did not say;
  • the first one on the first idea after the hook, then one every 5 to 10 s at most, 4 to 10 for a minute of video; a board that says the same as the subtitle under it is useless;
  • say with 2 to 4 words copied from <name>.txt exactly (spelling included), so the board lands on the word.

Write the file next to the video (<name>.boards.json), run again with --boards: it costs one render, not a new cut. The run lists the boards it could not place (words not found).

Ask the user — these three things, nothing else

  1. Where the clips are (often the newest videos in ~/Downloads; an AirDrop from an iPhone lands there), and their order if it is not the order of their names.
  2. The language they speak, unless --config already shows lang.
  3. A title on top: yes or no. Yes: their words, or offer one from what they say once the first run has written <name>.txt (a new title costs one render, not a new cut).

An audit of their account (audit-*/AUDIT.md in the current folder, or one they give, written by the audit skill): read its "For the edit" block ("Pour le montage") first, to offer a title shaped like hook_style / hook_example (unless title_on_screen says no) and to give the edited length next to target_duration_s. Nothing else changes.

The rest has a default: the end card says "Follow for more" in their language ("Abonne-toi pour la suite" in French); change it only if they ask (--cta "...", --no-cta).

First time

python scripts/edit_video.py --config

prints every setting, where it comes from (command line, then ~/.config/video-editing/config.json, then the default) and what is installed. A tool missing (ffmpeg, libass, whisper-cli, the ElevenLabs key with stt elevenlabs) or no config file yet ((none: defaults)): follow this line, the one the setup skill (installed next to this one) waits for (it diagnoses, shows the commands before installing anything, writes the settings):

Il manque un outil, le modèle whisper, la clé ElevenLabs ou un réglage ? Suis la skill setup, installée à côté de celle-ci (../setup/SKILL.md, son diagnostic : python3 ../setup/scripts/setup.py --json, py sur Windows), puis reprends le montage là où il s'est arrêté.

When one is missing, the script says it and does what it can:

  • no ffmpeg: it stops and says what to install;
  • ffmpeg without libass: the video comes out without subtitles burned in (the .srt is written all the same), the title, the boards and the end card drawn by Remotion if Node is there, else none;
  • no whisper-cli: silences only (no retake cut, no subtitles);
  • no Node, or remotion off: ffmpeg draws the title, the boards and the end card (square boxes, a simple pop, no count-up); the video comes out all the same (likewise when Remotion cannot render: Windows on ARM has no Remotion build);
  • stt = elevenlabs without a key: it stops before doing anything.

The two transcription engines

Same output for both (words with their start and end), the cut and the subtitles do not know which one spoke.

  • whisper (default): whisper.cpp on this computer, free, nothing leaves it. Each island of sound is transcribed on its own, then the kept ones are heard again for the timing of each word (the subtitles light up the word being said).
  • elevenlabs: ElevenLabs Scribe, paid on the user's ElevenLabs account, better on names and on the retakes whisper smooths over. One call for the whole video, kept in a cache (running again never pays twice). --stt elevenlabs, or "stt": "elevenlabs" in the config. The key is ELEVENLABS_API_KEY, in the environment or in ~/.config/video-editing/env (a line ELEVENLABS_API_KEY=..., chmod 600): the setup skill writes it, or the user does. Never ask for the key in the chat, never print it, never put it in config.json.

What to do

  1. Run the command. A 6 min rush takes about 2 to 3 minutes on a recent Mac.
  2. Read what it prints: duration before / after, the retakes it cut (each quotes the words), and the to check at 12.3s lines (a repeat it saw but could not cut: no silence to cut in).
  3. Read the takes, like an editor. Open <name>.takes.tsv and read the kept lines in order, as a script. The script only catches a retake that starts like the next one. Look for what it cannot see: a sentence abandoned and said again in other words just after ("let me show you the 5 steps" ... "in fact there are 5 ways, let me show you"), a lone fragment ("So..."). If you find one, run again with --cut A:B (the times of the abandoned lines, straight from the file) and the same -o. In doubt, keep: a repeat left costs a second pass, a sentence lost costs the video. Never cut something because it is off topic, a joke or an aside.
  4. Read <name>.txt for misheard words (names, brands, jargon): run again with --fix "heard=meant" (--fix "clode code=Claude Code"); it fixes the subtitles, the .srt and the .txt.
  5. Look before handing it back: stills of <name>.mp4 (ffmpeg -ss 1 -i <name>.mp4 -frames:v 1 a.png, with the ffmpeg path --config prints when ffmpeg alone is not found), one in each view (views in <name>.timeline.json: the hook, a split, the last second), then read them: the title, a board, a subtitle on the seam, the end card, nothing over the face.
  6. Hand back the path of the video, what you cut and why, the to check times if any, the views (how many turns), and the next step: boards if there are none yet ("mets des cartons"), then publishing with the publish skill (with Nadia, or a ready-to-post kit).

What the user asks next

Run edit_video.py again with the same clips and the same -o, and every option of the previous runs (--cut, --protect, --fix, --title, --cta, --noise, --boards, --scene...: each run starts from the clips). The last command is in <name>.state.json (command): start from it, add the new option. It re-transcribes nothing, and the cut is not rendered again when it did not change.

they sayyou pass
"cut the part about the price"--cut A:B (times from <name>.takes.tsv)
"keep the pause after the joke"--protect A:B
"cut harder, there is background noise", nothing got cut--noise 25 (the islands change: the new ones are transcribed)
"in the subtitles it's cloud, not clown"--fix "clown=cloud"
"another title", "no title"--title "...", --title ""
"another end card", "no end card"--cta "...", --no-cta
"add boards", "mets des cartons", "show the figures"--boards <name>.boards.json (section Boards)
"full frame only", "no split", "only the split"--scene full, --scene split
"no subtitles"--no-subtitles
"it's slow", a PC without a graphics card--model small (or large-v3-turbo-q5_0); for every video, the setup skill sets whisper_model
"just the cut, I finish it myself"--rough (-o is then the cut)
"bigger subtitles", "another colour", "higher", "one word at a time", "another font"the style (below)

Say what you changed for each request; one you cannot place in the takes, ask, do not guess.

The look: style.json

style.json (next to scripts/) holds the defaults: Inter ExtraBold (shipped, SIL OFL), white subtitles with the said word in mint green, an anthracite top zone. The font and the colour of the said word are chosen at setup (the setup skill asks for them and writes them); the rest is set in the JSON. To change the look, write only the keys that change in ~/.config/video-editing/style.json (kept when the skill updates), or in a file passed with --style FILE:

{"captions": {"size": 130, "highlight": "#7CFF6B", "y": 0.62, "max_words": 1}}
keydefaultwhat
font, font_fileInter ExtraBold (shipped; Anton too)a font installed on the computer by its name, or font_file = the path of a .ttf / .otf and font = its family name
captions.size86px on the 1080x1920 frame
captions.color, .highlightwhite, mint #3DDC97the line, the word being said
captions.outline, .outline_width, .shadownear black, 6, 2
captions.max_words, .max_chars3, 18words per line (1 = one word at a time)
captions.uppercase, .punctuation, .popfalse, false, true
captions.y / .y_wide0.64 / 0.85full frame: height of the text centre, 0 = top, 1 = bottom (9:16 / 16:9); in the split they sit on the seam
title.size, .color, .box, .highlight92, near black on white, green *word*full frame
title.duration, .y3.5 s, 0.17
end_card.text"" = "Follow for more" in the language
end_card.size, .color, .box, .duration, .y80, near black on mint, 3 s, 0.2
split.backgroundanthracite #17181Cthe top zone and around the card
split.text, .highlight, .title_sizeoff-white, mint, 120the title and the boards on the top zone
boards.size, .muted104, greya board's size, the caption of a stat
wide_scale0.72sizes in 16:9
loudness-14LUFS aimed at, one even gain that never pushes a peak past -1.5 dBTP; null keeps the sound as recorded

Keep the text away from the bottom fifth and the right edge of a 9:16: TikTok and Instagram put their buttons and the caption there.

Cut settings

settingdefaultwhen to touch it
--min-sil0.22 sthe edit feels breathless: 0.35; pauses left between sentences: 0.18
--noise32 dBbackground noise (street, fan) and nothing gets cut: 25
--pad0.05 sa word's ending sounds clipped: 0.08

--keep-retakes keeps every take (a deliberate repetition, a chorus), --keep-fillers keeps the lone "um"s, --no-transcript skips the transcription (silences only, no subtitles).

When it fails

  • A 0 s video or a concat error: the clips do not share the same format (an iPhone clip and another camera's): edit each clip alone.
  • nothing kept: a very quiet recording, --noise 40 (25 is for background noise).
  • No subtitles on the video: --config says NO libass: the setup skill installs an ffmpeg with libass.
  • ElevenLabs refused the key / no credits left: the key or the account; --stt whisper works meanwhile, free.

Not in this skill

The review page in the browser with the audio check, the split fitted to your own framing and the scenes measured on what works, the motion design, choosing your style (seven looks or your own, set on your own take): the paid version, https://pifpafcut.com/skill.

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.