claude-video-editing-skills

skill
Security Audit
Fail
Health Pass
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 41 GitHub stars
Code Fail
  • rm -rf — Recursive force deletion command in install.sh
Permissions Pass
  • Permissions — No dangerous permissions requested

No AI report is available for this listing yet.

SUMMARY

Free, local AI video editor for Claude Code: 95 skills turn raw footage into finished Reels, Shorts and YouTube edits - auto cuts, captions (Hindi/Devanagari too), motion graphics, scene sequences, colour, sound mix, QC.

README.md

🎬 Claude Video Editing Skills

A complete, free, local video editor for Claude Code

Raw clip in → cut, graded, mixed, captioned and delivered edit out.

No account, no watermark, no credits, no cloud. Hindi / Devanagari ready.

Stars
License: MIT
Claude Code
Skills
Version

Windows · macOS · Linux · runs on a laptop CPU (no GPU needed)

Quick start · What's inside · How it works · Styles · Genres · Catalog · Changelog


✨ Just say what you want

> Edit this video: C:\clips\talk.mp4
> Make a 30 s Instagram reel from podcast.mp4 with the glass-ribbon style
> Cut a wedding highlight from the wedding/ folder on song.mp3, 4 minutes
> These are my shayari lines - put them as captions, keep the song at the end

video-director watches and listens first (speech, language, faces, music, light, length), picks the genre, style, format and platform, and runs the whole edit through to delivery and QC. You can also name any genre or style yourself.

🚀 What's new in v2.2

🎞️ Designed scene sequences Styles now change scene every phrase like pro reels: speaker scenes, full-screen type scenes, glass cards, typing search bars, blur transitions. Six recipes: ribbon, aurora, retro, explainer, eclipse, layercut, applied automatically by run.py.
🇮🇳 Hindi / Haryanvi first-class Creator profile (language, script, platform), no more Urdu-script captions, transcribe.py --text puts your own lyrics/shayari on the speech timing, grapheme-safe text animation, Devanagari thumbnails, a poetry-reel genre.
⚡ 2.3× faster renders Faster overlay capture with frame reuse, same pixels.
✂️ Cuts that sound professional Equal-power crossfades and room tone at every join, VAD silence detection, wind/rumble filter.
🧑‍🎨 Cleaner cut-outs Guided-filter matte edges, steadier masks, edge-contact detection so layouts never show frame edges; blurred or replaced backgrounds with a rim glow.
🖥️ Local review app python video-studio-core/app/server.py --root <folder>: watch, fix caption words, switch style, re-render. Bound to 127.0.0.1 with a token.
✅ Honest QC Catches wrong caption script, missing glyphs, text over the face and black frames at the end.

📦 What's inside

count decides examples
🎬 Director 1 what the video should be, then runs it video-director
✂️ Genre skills 42 structure, pacing, cut rules, sound, deliverables retention-reel, podcast-episode, wedding-india, music-video
🎨 Style skins 46 captions, fonts, graphics, scene recipes glass-ribbon, glow-cutout, poster-pop, halo-cut
🧰 Utility skills 5 finishing, platform packs, accessibility, NLE hand-off, review rounds platform-pack, editor-handoff
⚙️ Core engine 1 transcription, cuts, timeline, reframe, colour, motion, sound, delivery, QC video-studio-core

Run any genre with any skin: wedding-highlight + doc-real, or retention-reel + glass-ribbon.

⚡ Quick start

Option A: Claude Code plugin

claude plugin marketplace add Viprasol-Tech/claude-video-editing-skills
claude plugin install video-studio@claude-video-editing-skills

Option B: installer (copies every skill to ~/.claude/skills)

git clone https://github.com/Viprasol-Tech/claude-video-editing-skills
cd claude-video-editing-skills
powershell -ExecutionPolicy Bypass -File install.ps1   # Windows
bash install.sh                                        # macOS / Linux

Then install the tools once

pip install faster-whisper playwright numpy pillow opencv-contrib-python mediapipe fastapi uvicorn fonttools
python -m playwright install chromium
winget install ffmpeg   # Windows  (macOS: brew install ffmpeg · Linux: sudo apt install ffmpeg)
python ~/.claude/skills/video-studio-core/scripts/check_env.py

check_env.py prints the exact fix for anything missing. Restart Claude Code, then say "edit this video: path/to/clip.mp4".

🧠 How it works

flowchart LR
  A[🎥 Footage] --> B[video-director<br/>reads it, decides]
  B --> C[prep · transcribe<br/>autocut · matte]
  C --> D[make_plan + scene recipe]
  D --> E[render<br/>HTML motion · FFmpeg]
  E --> F[finish<br/>grade · mix · loudness]
  F --> G[deliver + QC<br/>Reels · Shorts · YouTube · SRT]
  D -.-> R[review app<br/>fix captions, re-render]
  R -.-> D
layer what it decides examples
Director (the boss) watches and listens to the footage, picks genre, skin, format, platform and clean-up video-director
Genre skill (editorial logic) structure, pacing, cut rules, transitions, sound, loudness, deliverables podcast-episode, wedding-india, trailer-teaser
Style skin (the look) captions, fonts, colours, graphics, scene recipe glass-ribbon, essay-mode, news-desk
Core (engine + editor brain) transcription, cuts, timeline, reframe, colour, motion, sound, delivery, QC video-studio-core

The editor brain scores cut points with Murch's Rule of Six, sets pacing targets per genre, checks hook and retention beats, keeps overlays inside safe zones and limits caption reading speed, then reviews its own plan (editor.py review --fix). These are heuristics: they flag problems, and Claude and you decide.

Command-line pipeline (what Claude runs for you)
python CORE/scripts/run.py clip.mp4 --style SKIN --genre retention-reel --smart-cuts --retakes --work edit-x --stop-after plan
python CORE/scripts/editor.py review edit-x/plan.json --genre retention-reel --fix
python CORE/scripts/render.py edit-x/plan.json --frames 1,4,8,12     # contact sheet: look at it
python CORE/scripts/render.py edit-x/plan.json
python CORE/scripts/finish.py edit-x/plan.json --platform reels
python CORE/scripts/deliver.py edit-x/plan.json --to reels,tiktok,shorts
python CORE/scripts/qc.py edit-x/final.mp4 --plan edit-x/plan.json

CORE is the video-studio-core folder, SKIN a style folder such as ~/.claude/skills/glass-ribbon. run.py --help lists every option: --clips, --timeline, --auto montage|story|highlights, --reframe 9:16, --broll DIR, --music, --platform, --deliver youtube,reels. Hindi lyrics/shayari: python CORE/scripts/transcribe.py --work edit-x --text lines.txt. Creator profile: python CORE/scripts/profile.py show.

Genre skills (42)

Each one has a matching preset: python CORE/scripts/editor.py genre NAME shows its pacing, cut rules, captions, loudness and deliverables.

Short-form

skill what it makes
retention-reel 15-60 s talking-head reel: hook by 1.5 s, open loop, payoff, loop or CTA
clip-miner-pro the 5-15 strongest stand-alone moments from a long recording, reframed 9:16
reaction-duet reaction, duet or stitch: source clip plus facecam, source audio ducked
micro-drama binge-able 60-120 s vertical episodes with a recap open and a cliffhanger
gaming-vod stream highlights from audio and motion peaks, facecam over gameplay

YouTube and long-form

skill what it makes
longform-retention retention-structured long-form: minute-one promise, re-engagements, quick end
video-essay script-first, voice-over-led essay with archival, quote cards, maps and charts
vlog a day of A-roll and B-roll as a chronological story
tutorial-screen software how-to: promise, numbered steps, recap; waiting cut out
product-review verdict tease, specs, macro B-roll, pros and cons, score, disclosure
faceless-explainer offline voice-over plus one visual per sentence, no presenter
news-explainer inverted-pyramid explainer with attribution on screen

Podcast and talks

skill what it makes
podcast-episode multicam episode switched by who is talking, light filler removal, -16 LUFS
conference-talk keynote or panel with title card, slide inserts, trimmed dead air

Film and documentary

skill what it makes
interview-doc paper edit, radio edit, then B-roll over every interview cut
vox-pop street interviews: question on screen, answers grouped by intent
scene-editor scripted coverage cut by line strings, best performance per line
short-film-finisher per-shot colour, mix to delivery loudness, masters after picture lock
trailer-teaser three-act trailer with copy cards, risers into hits, button after the title

Ads, brand and corporate

skill what it makes
ad-ladder 60/30 s story master, rebuilt 15 s and 6 s bumper, every ratio
ugc-ad-matrix 3-5 hook variants x one body x CTAs, cut native to TikTok and Reels
brand-film 2-3 min mission film plus cutdowns, brand-guide lower thirds
testimonial-film real customer case study: problem, solution, numbers, recommendation
saas-launch-demo 60-180 s launch or demo from a screen recording, cursor zoom, callouts

Weddings, events and music

skill what it makes
wedding-highlight 3-5 min cinematic highlight on the music, vows as the spine, teaser
wedding-india multi-day Indian wedding sorted by event, highlight, per-event reels, long film
event-recap 60-120 s beat-cut recap with speaker soundbites and a 9:16 version
concert-multicam 2-4 angles switched to the music over the board mix
music-video performance takes synced to the master, intercut narrative, cut on downbeats
lyric-video every lyric line timed to the vocal with a karaoke fill
memorial-tribute photo-led chronological life arc, slow and family-approved

Lifestyle and how-to

skill what it makes
recipe-video 30-60 s recipe reel: money shot first, ingredients, sped-up process
fitness-video workout reel or follow-along with rep counters and interval timers
fashion-lookbook beat-cut lookbook with outfit-change transitions and true-colour grade
travel-cinematic music-led travel film with match cuts, whips and speed ramps
real-estate-tour walk-through in buyer's-visit order with price, beds and baths cards
course-module 3-10 min lesson: one objective, steps, quiz cards, recap, accessible captions

Specialist

skill what it makes
kids-content bright, simple, repetitive, sing-along captions at slow reading speed
asmr-video long takes, untouched trigger audio, no music, slow dissolves
regional-devotional bhajan, aarti, kirtan and temple films with Indic-script lyrics
localise-dub translated subtitles and a time-fitted offline voice-over dub
sports-highlight plays from crowd and commentary peaks, freeze-frame spotlight, player card

Utility skills (5)

skill what it does
finishing-pro correct, match, look, texture, mix, loudness, legalise, encode, QC on any locked edit
platform-pack one master into every platform version, plus captions, chapters and thumbnails
accessibility-pack SDH subtitles, clean transcript, audio-description script and described version
editor-handoff EDL, FCPXML and OTIO timelines, ProRes 4444 graphics layer, mix WAV for an NLE
versions-review numbered versions never overwritten, A/B contact sheets, notes rounds

Style skins (46)

Each skin is a folder with theme.json and theme.css. Pass it as --style.

category skins
Talking-head reel (28) glow-cutout† poster-pop†* glass-ribbon† neon-marquee* grid-stage†* halo-cut depth-type* split-proof neon-orbit* paper-cut chaos-stack bold-bounce minimal-mono frame-studio vhs-tape y2k-chrome* news-desk doc-real cinematic-phone sticker-pop terminal-tech comic-panel swiss-poster* soft-gradient countdown-list before-after dual-subs hook-lab
YouTube explainer (6) editorial-cards essay-mode whiteboard-host chapter-pro highlighter-study tutorial-steps
Explainer (1) data-story
Podcast / repurposing (3) clip-miner pod-frame quote-card-reel
Product / app promo (8) device-fly kinetic-launch ugc-ad-kit logo-sting testimonial-card event-promo product-spotlight loop-ambient

* = text sits behind the speaker, † = background blurred or replaced around the speaker (both need the person matte). The full list with one-line looks is in docs/catalog.md.

What the engine does

area what you get script
Cutting and the brain pauses, fillers (English and basic Hindi), retakes, cuts at the quietest point, kept breaths, J-cuts at sentence joins; selects and paper edit; story outline, hook candidates, reorder with a constructed-quote warning; Rule-of-Six scoring, pacing, retention, overlay and caption review with safe fixes; 69 genre presets autocut.py selects.py story.py editor.py
Timeline multi-clip assembly with 58 FFmpeg xfade transitions plus dip-black, dip-white and whip; J/L cuts; constant speed and speed ramps; optical-flow slow motion; freeze, reverse, Ken Burns stills; beat-cut montage; frame-exact timing timeline.py
Analysis shot, black, freeze and flash detection; face tracking and active speaker; beats, downbeats and waveform peaks scenes.py track.py beats.py
Reframe face-follow reframe to any aspect, smooth, fast or locked camera; two-person stacked layout; multicam switching by loudest mic track.py multicam.py
B-roll places B-roll from a folder on matching words, covering jump cuts first broll.py
Colour auto correction (white balance, levels, contrast, saturation), approximate log to Rec.709 for 5 cameras, official LUT input, shot matching, 51 looks with skin protection, film texture, scopes and legal levels. Colour is applied to the footage, not to captions and graphics color.py luts.py
Motion 45 element types, 34 in/out animation names, 30+ easings, springs, keyframes, 10 text effects, 6 caption presets with safe zones, reading-speed limits and face dodge, 11 transition overlays, beat pulse, face anchoring runtime/engine.js
Sound dialogue chain (high-pass, de-hum, de-click, de-noise, EQ, de-ess, gate, compressor), room-tone fill, 49 synthesized sound effects, speech-aware music ducking, music cut to length on bars, offline voice-over, stems audio_chain.py music_fit.py tts.py stems.py
Delivery and QC 23 platform presets (22 encodes plus a youtube-hdr note) with two-pass loudness, reframe and size budgets; SRT, VTT and TTML subtitles to broadcast rules; chapters; thumbnails; post pack; EDL, FCPXML and OTIO; ProRes 4444 overlays; automated QC report deliver.py captions.py export_edl.py qc.py

Loudness targets per platform: -14 LUFS for social and YouTube, -16 for podcasts, -23 for EBU broadcast, -24 for ATSC and -27 for Netflix (approximate, not dialogue-gated).
Every move, with its exact plan field or flag and an honest Ready / Upgrade / Heavy tag, is in docs/catalog.md.

Requirements

  • Claude Code (or Cowork on desktop) with skills enabled
  • Python 3.10 or newer (tested on 3.14)
  • FFmpeg 7, 8 or 9, full build, with ffprobe on PATH
  • pip install faster-whisper playwright numpy pillow opencv-contrib-python mediapipe fastapi uvicorn fonttools, then python -m playwright install chromium
  • About 500 MB of disk for the first Whisper model (small). The person-matte model (~16 MB) downloads on first use.
  • Internet only for the first model downloads and Google Fonts. Offline, fallback fonts are used.

Optional add-ons

None of these are bundled. The engine works without them and uses them when they are installed.

add-on for install licence
Kokoro best offline voice-over pip install kokoro soundfile Apache-2.0
Piper alternative voice-over separate program, called on the command line GPL-3.0 (run as a separate program only, never imported)
rembg alternative person matte, sometimes cleaner hair pip install rembg onnxruntime MIT
Demucs split vocals from music pip install demucs (slow on CPU) MIT
librosa alternative beat tracker pip install librosa ISC
PySceneDetect alternative shot detection pip install scenedetect BSD-3
OpenTimelineIO OTIO through the official library pip install opentimelineio Apache-2.0
Windows SAPI / macOS say / espeak-ng built-in fallback voices already on the system system

Heavy models are listed by check_env.py for reference only. Some are personal-use, non-commercial models: they are optional, never bundled, and never downloaded by this pack.

model for licence note
ProPainter object removal non-commercial (S-Lab), needs a large GPU
XTTS-v2 / F5-TTS voice cloning non-commercial weights; personal use only, and only with the speaker's consent
MusicGen music generation CC-BY-NC weights; personal use only
Real-ESRGAN / RIFE upscale / frame interpolation BSD-3 / MIT

Check each upstream licence yourself before you install it.

Make your own style

  1. Copy any skin folder, for example halo-cut, and rename it. The folder name must equal name: in its SKILL.md.
  2. Edit theme.json. It holds the format, fonts, CSS variables, speaker layout, the auto rules (captions, keywords, headlines, numbers, lists, hook, end card, cut sounds) and finish.
  3. Edit theme.css to style the elements (.cap-preset-*, .card, .lowerthird, .chip and so on).
  4. For the look, set finish.grade to any of the 52 look names (for example portra or teal-orange-pro). It is applied to the footage, not to your graphics.
  5. Test it with run.py clip.mp4 --style YOUR-SKIN --stop-after plan, then render.py plan.json --frames 1,4,8,12.

References: video-studio-core/references/theme-schema.md, plan-schema.md and motion.md.

Make your own genre

  1. Add a preset to video-studio-core/scripts/genres.json under genres. Copy the closest one and change the aspect, duration, autocut rules, ASL range, captions, music, SFX density, allowed transitions, hook rules, structure beats, loudness and safe zone.
  2. Check it: python CORE/scripts/editor.py genre YOUR-GENRE.
  3. Make a skill folder with a SKILL.md (front matter name: = folder name, plus a description: that says when to use it). Write how a pro cuts this kind of video: structure, pacing, cut rules, sound, colour, deliverables, common mistakes, and the exact run.py ... --genre YOUR-GENRE commands. Use any existing genre skill as the template.
  4. autocut.py, make_plan.py, editor.py review and story.py all read --genre.

References: video-studio-core/references/craft.md, timeline.md, color.md, audio.md and delivery.md.

Limits (honest)

  • Limited real-footage testing. Tested on generated clips and on real low-resolution phone footage (Hindi/Haryanvi). Face tracking, the person matte and skin protection have not been tested on a wide range of faces, skin tones, lighting or profiles. Check every contact sheet.
  • No GPU features by default. Object removal, upscaling, AI frame interpolation, depth effects and generative video are not included. Optical-flow slow motion uses FFmpeg minterpolate, which is slow and can warp complex motion.
  • No lip-sync and no voice cloning. Dubs are voice-over style, with synthetic offline voices.
  • Log to Rec.709 LUTs are approximate. They are generated from published curve and gamut formulas, not from manufacturer data. Use the official camera LUT (color.lut) when you have one.
  • Skin protection is hue-range based inside the LUT, not a face mask.
  • Face detection misses faces smaller than about 5% of the frame width and profile faces. With no face, the reframe centre-crops.
  • Multicam switching does not sync cameras by waveform; every file must start at the same moment.
  • Netflix loudness is measured on the full programme, not dialogue-gated. SCC captions are not produced (there is no free CEA-608 encoder). HDR deliveries are not produced.
  • Automated QC is a set of heuristics: average-luma flash detection, stream-timing A/V sync and box-based safe zones. It does not replace a watch-down.
  • Translation, chapter titles, post copy and audio-description scripts are drafted by Claude from the transcript and the picture. Check them.
  • Use only music, footage, fonts and reviews you have the rights to. testimonial-card and testimonial-film are for real customers and real words only.

Licence

MIT. See LICENSE. Optional add-ons and models keep their own licences.
See CHANGELOG.md for what changed.


If this saves you editing time, please ⭐ the repo - it helps other creators find it.

Made by Viprasol · Issues and pull requests welcome

Reviews (0)

No results found