As promised in Content Lab
The Claude skill I use to cut my reels
One raw take goes in. A finished vertical reel comes out - hook burned on top, captions from the first second, every fumble and silence cut away. It all runs on your own machine.
Free. No email needed, no sign-up, no strings.
I record in one take and never stop when I fumble
I am not good at editing video and I find it deeply boring. So I stopped doing it. I turn the phone on, say my thing, and if I mess up a sentence I just say it again. Then I hand the file to Claude and this skill does the rest:
- Cuts the dead air and the restarts - the cut follows the actual silence in the audio, not the transcript, so nothing quiet gets lost.
- Burns in the captions from the first second, broken on phrase boundaries so each line reads like a sentence, never ending on "the" or "and".
- Puts your hook on top in your brand colour, inside the reels safe zone, and drops it the moment you stop saying that sentence.
- Exports a 1080x1920 MP4 ready to upload. No music - it competes with your voice and adds nothing.
It runs locally: ffmpeg for the cut, Whisper for the transcript. Nothing is uploaded anywhere.
Four steps, about twenty minutes the first time
Upload the ZIP into Claude
You do not unzip it and you do not touch a terminal. In the Claude desktop app - or on claude.ai, it is the same account - open Settings, then Customize, then Skills. Click Add, choose Upload a skill, and drop in the ZIP you just downloaded.
It lands in your account, so it is there in the desktop app, in the browser and in Claude Code - you do not register it anywhere. If you would rather keep it as a folder on disk, unzip it into ~/.claude/skills/ (C:\Users\<you>\.claude\skills\ on Windows) and restart Claude Code instead.
Say yes to what it installs
It shows you the list first and installs nothing until you agree: ffmpeg (~80 MB, does the cutting), faster-whisper (~200 MB, does the transcript), pillow (~5 MB, draws the hook), and the Whisper model (~1.5 GB, downloaded once with your first reel).
You say yes and Claude installs it - there is nothing for you to type. All of it runs on your machine. The first reel is the slow one; after that the model is cached.
Answer six questions, once
One question at a time, each with a default - "whatever you think" is a valid answer.
- Where your raw videos land. A folder on this computer, or Google Drive / Dropbox / OneDrive / Box through a Claude connector. It then checks it can actually reach it - and if the connector is off, it says so instead of guessing.
- Where finished reels go. Same place as the first answer, by default.
- Where the working files may live. A few hundred MB of temporary files that have to sit on a local disk - and get deleted once you approve the reel.
- Your language. English by default.
- Which font. No font comes in the download - it shows you the ones already installed on your machine and you pick one, or hand it a path to your own TTF. Out of that family it takes the heaviest weight for the hook and a bold one for the captions.
- Your brand colour for the hook box.
The answers are saved in your home folder (~/.talking-head-reel/profile.json), not inside the skill, so re-uploading or re-installing never wipes your setup. It never asks again. To change something later you just say so - "use Poppins instead" - and it rewrites that one answer.
Ask for the reel - then approve it
Drop the take in your input folder. Then one message to Claude, with your hook - the sentence you want burned over the opening:
Cut a reel out of the newest video.
Hook: "3 things I stopped doing before every launch"
Then, in order: transcribe → align the silence → it shows you the planned cuts → render → it checks the frames (it cannot watch the video, so it reads single frames to prove the hook and captions landed inside the safe zone) → you look at it and say yes → saved to your folder, working files deleted.
Nothing is filed as finished and nothing is deleted before you approve it. The delete also refuses to run unless the saved copy checks out byte for byte.
What it is not
- It is not a full video editor. One speaker, one take, vertical. That is the whole job it does.
- It adds no music and no b-roll. On purpose.
- The first run downloads the Whisper model (~1.5 GB) and is slow. Every run after that is fast.
- On a CPU the transcript takes roughly as long as the video. A GPU is faster when the drivers cooperate; if they do not, it falls back to CPU on its own.
- It is my working setup, not a product. It comes as it is, with the notes and the warnings I wrote for myself.
Take it
The pack: SKILL.md, a README and eight Python scripts. No font, no binaries - it uses a font you already have.
What people asked in the comments
Do I need Claude Code to use this?
No. Claude Code makes it a one-sentence job, but the pack is plain Python scripts. You can run them by hand following the steps in SKILL.md.
Does anything get uploaded?
No. The transcript and the render both run on your machine. The video never leaves your computer.
What happens to the temporary files?
While it works, the skill keeps a few hundred MB in a folder you chose - a copy of the video, a wav of the audio, the render. Once you have looked at the reel and approved it, it saves the finished MP4 and the caption text to your output folder, verifies that copy byte for byte, and then deletes the whole working folder. If the copy does not verify, it refuses to delete anything. Want the raw material kept? Say so, or set "cleanup": false in your profile.
Does it work in languages other than English?
Yes. Whisper handles the transcript in any language it supports - set the language code in config.json. The caption line-breaking knows the function words of English, Czech, German and Slovak; anything else falls back to the English rules, which still works, it just breaks lines a little less gracefully.
How long does it take?
The first run also downloads the Whisper model, about 1.5 GB. After that a 4-minute video takes a few minutes to transcribe on a CPU and a couple more to render.
Can I change the fonts and colours?
Yes, and you already chose the font during setup - no font ships with the skill. To swap it, just say so: "use Poppins instead", or point it at your brand TTF. It can also list every font installed on your machine and let you pick. The hook colour is one hex value, changed the same way.
Can I use it commercially and change it?
Yes, freely and with no attribution required. The pack is text files only - no font, no binaries - so there is nothing in it with somebody else's licence attached.
I share the next one the same way
I build these for my own work - reels, funnels, invoices, whatever is eating my week - and I write down what worked and what broke. If you want that in your inbox, leave your email. No schedule, no filler, unsubscribe in one click.