video-edit
"Turns raw 9:16 talking-head clips into a finished, premium Instagram reel end to end: picks the cleanest take of every line across your clips, cuts dead air, punches in from 4K, cuts you out of the background for words behind your head, applies one of three designed looks (bold kinetic words, cool girl, cool dude), builds animated UI inserts with parallel sub-agents, adds subtle sound design, and checks every cut frame by frame inside the Reels safe zone. Trigger with /video-edit or phrases like 'edit these clips', 'I dropped clips in my downloads', 'cut this into a reel', 'make this reel look fire'."
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/tenfoldmarc--video-edit-skill)HTML
<a href="/listings/tenfoldmarc--video-edit-skill"><img src="/api/listings/tenfoldmarc--video-edit-skill/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
No terminal needed.
Step 1: Open Claude Code
Open Claude Code: the Code tab in the Claude desktop app, or type `claude` in your terminal. (This skill renders video on your computer, so it needs Claude Code; a regular chat can't run it.)
Step 2: Paste this message
Copy-paste this into the chat and hit Enter:
Install this skill for me: https://github.com/tenfoldmarc/video-edit-skill
Claude downloads the skill into your skills folder and tells you when it's ready. Takes a few seconds.
Step 3: Run the skill
Type:
/video-edit
Hit Enter. On your first run, the skill asks your name, your Instagram handle and your favorite look, checks your tools and sets up everything it needs (about 2 minutes, once). Just follow the prompts.
<details> <summary>Prefer the terminal? Manual install</summary>
Open Terminal (Mac: `Command + Space`, type Terminal. Windows: Start menu, type PowerShell). Paste this one line and hit Enter (it needs [git](https://git-scm.com/downloads) installed):
git clone https://github.com/tenfoldmarc/video-edit-skill "$HOME/.claude/skills/video-edit"
Then type `claude` to open Claude Code and run `/video-edit`. </details>
---
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 169 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- ---
- name: video-edit
- description: "Turns raw 9:16 talking-head clips into a finished, premium Instagram reel end to end: picks the cleanest take of every line across your clips, cuts dead air, punches in from 4K, cuts you out of the background for words behind your head, applies one of three designed looks (bold kinetic words, cool girl, cool dude), builds animated UI inserts with parallel sub-agents, adds subtle sound design, and checks every cut frame by frame inside the Reels safe zone. Trigger with /video-edit or phrases like 'edit these clips', 'I dropped clips in my downloads', 'cut this into a reel', 'make this reel look fire'."
- ---
- # /video-edit
- Raw clips in, finished reel out. The user never opens an editor: they review a phone copy and give notes; every
- round is a ~2 minute re-render.
- Read before building: `references/looks.md` (the three looks, exact values), `references/layout.md` (safe zone,
- placement, behind-head formula), `references/gotchas.md` (every trap already hit once). Read
- `references/overlays.md` when the script calls for an animated insert.
- Works on macOS, Linux and Windows. setup.py detects the machine at the start of every run and every command below
- comes from that record, so never guess paths or shells: see "This machine" right after Step 0.
- ## Step 0: First-time setup (runs once)
- `SK` is this skill's folder (the one holding this SKILL.md), as a full path. Usually `~/.claude/skills/video-edit`;
- on Windows that is `C:/Users/<name>/.claude/skills/video-edit`.
- Look for `SK/config.json`.
- **If it exists:** load it (name, instagram handle, default look, folders), run the setup script exactly as in
- step 3 below (a few seconds: it only checks, and repairs the environment after an update), read "This machine",
- and go to Step 1 once it prints READY.
- **If it does not exist:**
- 1. Say: "Welcome to /video-edit! Let's get you set up. This takes about 2 minutes and only happens once."
- 2. Ask, in one message:
- - Your name (used on your Instagram profile insert)
- - Your Instagram handle (for your profile insert and to pull your reels for the orbit effect)
- - Default look: **bold** (kinetic words behind your head), **cool girl** (warm, yellow serif italic, your reels
- orbiting you) or **cool dude** (dark editorial, tracked caps + serif). Offer bold if they are unsure.
- - Where your raw clips usually land (default `~/Downloads`) and where projects should live (default `~/reel-edits`)
- 3. Run the setup script with whichever Python the machine has. Try these in order until one prints STATUS lines:
- `python3 "SK/scripts/setup.py"`, then `python "SK/scripts/setup.py"`, then `py -3 "SK/scripts/setup.py"`.
- (On Windows `python3` is often a Microsoft Store stub that prints "Python was not found": just try the next one.)
- None of them runs: Python is missing. Give the fix for their system and stop (Windows:
- `winget install --id Python.Python.3.12 -e`, Mac: `brew install python`, Linux: `sudo apt install python3 python3-venv`).
- Then read the STATUS lines:
- - ffmpeg / node / python venv MISSING: the line ends with the exact fix for THIS machine (winget on Windows,
- brew on Mac, apt on Linux). Give it to the user and stop until the script prints READY.
- - Windows: the user pastes the winget lines into their own PowerShell window (winget asks them to agree to its
- terms the first time), then quits Claude Code completely and reopens it, or the new programs are not found.
- - The script downloads the 6 sound effects itself. If it still reports them MISSING (offline), that is optional:
- the reel just has no SFX until they are added.
- 4. Optional, Apify (pulls reels and profile stats automatically): try a simple Apify MCP call. If it is not
- connected, say it is optional and how to add it:
- 1. Create a free account at https://apify.com and copy the API token from Settings > API & Integrations.
- 2. In a terminal, paste this one line with the token in place of YOUR_TOKEN:
- `claude mcp add apify -e APIFY_TOKEN=YOUR_TOKEN -- npx -y @apify/actors-mcp-server`
- (heads up: Claude Code stores the token in plain text in `~/.claude.json`; keep that file private).
- 3. Reopen Claude and run /video-edit again.
- Without Apify, the skill asks them to drop their reel files into the project instead.
- 5. Save `SK/config.json` (no secrets, ever):
- ```json
- {"name": "...", "instagram": "handle-without-@", "defaultLook": "bold", "clipsDir": "~/Downloads",
- "projectsDir": "~/reel-edits", "setupComplete": true, "setupDate": "YYYY-MM-DD"}
- ```
- 6. Say: "Setup complete! Running /video-edit now..." and continue.
- ## This machine (read on every run)
- Read `SK/.platform.json`, written by setup.py. It records the OS (`os`: macos, linux or windows), the CPU and the
- exact commands for this computer:
- - `PY` = its `python` value: the skill's own Python (`.venv/bin/python` on Mac and Linux,
- `.venv/Scripts/python.exe` on Windows). Run EVERY Python script of this skill with it, never a bare `python3`.
- - `HF` = `PY "SK/scripts/hf.py"`: starts the pinned HyperFrames (`npx --yes hyperframes@0.8.34 ...`) with the
- right Node on any OS. No `nvm use`, no `npx.cmd` trouble.
- - `cutout`: `CoreML` (Apple Silicon) or `CPU` (Windows, Linux, Intel Macs: no GPU help, so the cutout step is the
- slow one; say so, run it in the background and keep building).
- setup.py rewrites this file on every run (Step 0), so it is always current. File missing, or `PY` no longer
- exists: setup did not run or did not reach READY; run it as in Step 0.3 and follow its fix.
- One command style that works in zsh, bash, Git Bash and PowerShell:
- - Full paths in double quotes with forward slashes (fine on Windows too): `"PY" "SK/scripts/ingest.py" "<project>"`.
- - In PowerShell (Windows without Git for Windows, or whenever the PowerShell tool is the one you have) put `&` in
- front: `& "PY" "SK/scripts/ingest.py" "<project>"`.
- - One command per call. No `&&`, no `VAR=1 command`, no `~` inside arguments, no `source`, no `zsh`.
- - `ffmpeg` and `ffprobe` are on PATH on every OS (setup checks it), so one-off ffmpeg commands are the same everywhere.
- ## Hard rules
- 1. Everything inside the Reels safe zone: top 220, bottom 450 (nothing below y 1470), sides 35, right 100 from y 1155.
- 2. Never cover the face with text or cards. Behind-head words show at least ~70% of each letter.
- 3. Captions say what was actually said: fix Whisper's mishearings of names and keywords ("Claude" often comes out
- as "cloud"/"claws", "skill" as "scale"). No em dashes in on-screen text.
- 4. Sound effects subtle (template `SFX_GAIN = 0.75` of the base volumes), never louder than the voice.
- 5. Credit other creators on screen (@handle) whenever their reels appear. Real numbers on the user's own profile;
- anything else (commenters, chat history) is a generic placeholder, never a real person, never fabricated proof.
- 6. Prove cuts from the final render, not from snapshots.
- 7. Never delete the user's files. Originals are read, never modified.
- ## Workflow
- ### 1. Intake
- - Clips: the files the user names, else the newest video files in `clipsDir`. Create `<projectsDir>/<short-slug>/`.
- (`~` in config.json is the user's home folder, `C:/Users/<name>` on Windows. The skill's scripts expand it.)
- - Reference links (IG reels, Pinterest pins) for a style: download them and look at frames before designing anything
- (see overlays.md for how to fetch them).
- ### 2. Ingest
- ```bash
- PY SK/scripts/ingest.py <project> <clips or folder> [--newest 5]
- ```
- Look at `work/sheets.jpg` (setups and angles) and read `transcript.txt` (per-clip lines, words with times, speech map).
- ### 3. Pick takes -> `<project>/edl.json`
- - Rebuild the script from the takes: people repeat a line until it lands. Use the LAST complete clean take unless
- the user says otherwise ("the better section starts at 0:47" means one continuous take from there).
- - In/out points from the speech map: start ~0.05-0.07s before the first word, end at the silence start after the
- last word. Kill dead air between sections; confirm short pauses with silencedetect before trimming (Whisper can
- invent a leading word at a cut).
- - Two segments in a row from the same angle: punch one in (zoom 1.3-1.9 from the 4K original, cx/cy on the chest).
- - Unsure words (keyword, CTA trigger): re-transcribe that snippet with medium.en and a prompt containing the names.
- - Format: `[{"id": "hook", "clip": "c1", "in": 12.66, "out": 15.72, "zoom": 1.0, "cx": 0.5, "cy": 0.5, "line": "..."}]`
- - Post the EDL to the user as a short table (line, clip, look per section) and keep going; do not wait.
- ### 4. Assemble, transcribe, cut out
- ```bash
- PY SK/scripts/assemble.py <project> # aroll.mp4 clean 30fps CFR + segments.json with exact frames
- PY SK/scripts/transcribe_cut.py <project> "<names and keywords in the script>"
- PY SK/scripts/cutout.py <project> # run in background; asserts cutout frames == aroll frames
- PY SK/scripts/headpos.py <project> # after the cutout: head/body position per segment
- ```
- One cutout at a time (about 1.5 GB of RAM). It prints a progress line every 30 seconds.
- ### 5. Inserts in parallel (as soon as the EDL is locked)
- For each animated insert the script calls for (an app on screen, the profile follow on the CTA, a file scroll),
- spawn a sub-agent in the background with the brief template in `references/overlays.md`: fixed size, fixed
- duration, beats locked to the word times, transparent webm out. Keep building the main edit while they work.
- ### 6. Build
- Copy `SK/assets/template/` into the project (build.py, hyperframes.json, package.json, assets/fonts, assets/sfx),
- then adapt build.py. It is the finished build of a real reel: keep the machinery, rewrite the content (its
- docstring lists which is which).
- - Which look where: `defaultLook` for the whole reel; a line that names a look switches its section.
- - Orbit and profile grid use the user's own reels (overlays.md: Apify post scraper with their handle, or files they
- drop into `assets/reels/`).
- - Iterate from the project folder: `HF lint` (0 errors), then `PY build.py --safe`, then
- `HF snapshot --at <10-20 moments> --no-end --describe false`. Read the contact sheets, fix layout, readability
- and safe zone. (`HF` picks the right Node by itself, so there is no `nvm use`.)
- A build.py copied from an older version of this skill ignores `--safe`: if the red guide is missing from the
- snapshots, build that project the old way, `SAFE=1 "PY" build.py` (such projects only exist on Mac and Linux).
- ### 7. Render and QA
- ```bash
- PY build.py # without --safe
- HF render -o renders/<slug>-vN.mp4 --quality high # ~1.5-2 min for a 25s reel on Apple Silicon
- PY SK/scripts/check_cuts.py <project> renders/<slug>-vN.mp4 --phone
- ```
- Look at `work/cut_check.jpg`: every after-cut frame shows the new shot cleanly. Check audio peaks (`volumedetect`,
- max around -2 to -5 dB) and the silence at section joins.
- ### 8. Deliver
- Send or point to the phone copy (720p) and the full-quality render. Report in plain terms what changed and anything
- that needs the user's call. Keep every version in `renders/`; back up `assets/` + `segments.json` to `work/vN/` before
- re-assembling.
- ## Defaults that came out of real feedback (do not re-litigate)
- - Change the grade and captions to create a look; do not replace the background with an AI-generated room.
- - No film grain. Dark is fine, but keep the subject lit and skin from going red.
- - Lead-in captions centred in front of the speaker read better than lines pinned to the left edge.
- - Words behind the head are the signature move, but only when readable.
- ---
- Built by [@tenfoldmarc](https://instagram.com/tenfoldmarc). Follow for daily AI automation builds, real systems, not theory.
README.md · README · 131 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- # Video Edit
- ### A Claude Code Skill by [@tenfoldmarc](https://www.instagram.com/tenfoldmarc)
- Film your talking-head clips on your phone, drop them in a folder, type `/video-edit`, and Claude turns them into a finished Instagram reel. It picks your best take of every line, cuts the dead air, cuts you out of the background so words can sit behind your head, and applies a designed look with captions, motion graphics and sound. You never open an editor: you watch the result, say what you want changed, and it re-renders in about two minutes.
- ---
- ## What It Does
- 1. Finds your clips, makes contact sheets and transcribes every take with word-level timing.
- 2. Rebuilds your script from your takes (you can repeat lines as many times as you want) and keeps the cleanest one of each, with zero dead air between sections.
- 3. Cuts the reel straight from your original files, with sharp punch-ins from 4K for variety.
- 4. Cuts you out of the background, so big words can sit behind your head.
- 5. Applies one of three looks:
- - **Bold:** kinetic words around you, the big ones behind your head.
- - **Cool girl:** a warm grade with a yellow serif italic, and your own reels orbiting around you.
- - **Cool dude:** a dark editorial grade with tiny spaced capitals and big serif words.
- 6. Builds animated inserts in parallel with sub-agents: an app on screen behind you, your Instagram profile getting followed on your call to action, a file scrolling.
- 7. Keeps everything inside the Instagram safe zone, checks every cut frame by frame, and hands you a phone-ready copy.
- ---
- ## Requirements
- - A Mac, a Windows 10/11 PC or a Linux computer. Apple Silicon Macs are fastest for the background cutout; on Windows and Linux that one step runs on the processor, so it takes longer.
- - [Claude Code](https://docs.anthropic.com/en/docs/claude-code) installed and working
- - [ffmpeg](https://ffmpeg.org)
- - [Node.js 22 or newer](https://nodejs.org)
- - Python 3.10 or newer ([python.org](https://www.python.org/downloads/))
- - Optional: an [Apify](https://apify.com) account (free tier works) to pull your reels and profile stats automatically
- Don't worry about connecting these manually. On first run the skill checks your computer, figures out whether it's a Mac, a PC or Linux, and gives you the exact line to paste for anything that's missing.
- | | Mac | Windows (PowerShell) | Linux |
- |---|---|---|---|
- | ffmpeg | `brew install ffmpeg` | `winget install --id Gyan.FFmpeg -e` | `sudo apt install ffmpeg` |
- | Node.js | `brew install node` | `winget install --id OpenJS.NodeJS.LTS -e` | [nodejs.org](https://nodejs.org) |
- | Python | `brew install python` | `winget install --id Python.Python.3.12 -e` | `sudo apt install python3 python3-venv` |
- **Windows support is new.** The skill was built and tested on a Mac, and Windows has not had that level of real-world testing yet. If something breaks on your PC, please [open an issue](https://github.com/tenfoldmarc/video-edit-skill/issues) with the error text and the `STATUS` lines the setup printed.
- **On Windows:** each line is one paste into PowerShell. After installing, close Claude Code completely and open it again, or it won't see the new programs. Git for Windows is optional (the skill works with or without it). On a Windows ARM laptop (Snapdragon), add `--architecture x64` to the Python line.
- ---
- ## Install
- No terminal needed.
- ### Step 1: Open Claude Code
- Open Claude Code: the Code tab in the Claude desktop app, or type `claude` in your terminal. (This skill renders video on your computer, so it needs Claude Code; a regular chat can't run it.)
- ### Step 2: Paste this message
- Copy-paste this into the chat and hit Enter:
- ```
- Install this skill for me: https://github.com/tenfoldmarc/video-edit-skill
- ```
- Claude downloads the skill into your skills folder and tells you when it's ready. Takes a few seconds.
- ### Step 3: Run the skill
- Type:
- ```
- /video-edit
- ```
- Hit Enter. On your first run, the skill asks your name, your Instagram handle and your favorite look, checks your tools and sets up everything it needs (about 2 minutes, once). Just follow the prompts.
- <details>
- <summary>Prefer the terminal? Manual install</summary>
- Open Terminal (Mac: `Command + Space`, type Terminal. Windows: Start menu, type PowerShell). Paste this one line and hit Enter (it needs [git](https://git-scm.com/downloads) installed):
- ```bash
- git clone https://github.com/tenfoldmarc/video-edit-skill "$HOME/.claude/skills/video-edit"
- ```
- Then type `claude` to open Claude Code and run `/video-edit`.
- </details>
- ---
- ## Usage
- **Edit today's clips:**
- ```
- /video-edit I just dropped 5 clips in my downloads
- ```
- Claude finds them, shows you which take it picked for each line, and sends you the finished reel.
- **Name the look per section:**
- ```
- /video-edit use the cool girl look when I say "the cool girl aesthetic", cool dude after that, bold everywhere else
- ```
- **Give notes on a version:**
- ```
- the captions feel too small, the grain is too much, and cut the pause before "or you can get"
- ```
- It fixes exactly that and re-renders in about two minutes.
- Tips for filming: shoot vertical in 4K if your phone allows (punch-ins stay sharp), keep the camera still for each setup, and repeat a line until it feels right. The skill picks the best take.
- ---
- ## Updating
- Paste this into Claude:
- ```
- Update the video-edit skill from https://github.com/tenfoldmarc/video-edit-skill
- ```
- Or in Terminal / PowerShell: `git -C "$HOME/.claude/skills/video-edit" pull`
- ---
- ## Credits
- Built on [HyperFrames](https://github.com/heygen-com/hyperframes) (HTML-to-video rendering and background removal) and [faster-whisper](https://github.com/SYSTRAN/faster-whisper). Fonts from Google Fonts (SIL Open Font License). Sound effects from Pixabay, downloaded on setup.
- ---
- ## Built By
- [@tenfoldmarc](https://www.instagram.com/tenfoldmarc). Follow for daily AI automation walkthroughs, real systems, not theory.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.