Can Claude edit videos? How I built a reel studio that learns from my feedback
I edit my reels with Claude Code, FFmpeg and HyperFrames, all on my Mac. A reel went from an hour of editing to 5-10 minutes of my time. The clever part is that every note I give becomes a rule, so each reel starts better than the last.
Short answer: Yes. Claude can edit videos when you give it the right tools. I record on my phone in Instagram Edits, AirDrop the raw video to my Mac, and Claude Code runs a small studio there: FFmpeg cleans up the sound and cuts out the dead air, whisper writes the captions, and HyperFrames adds my text and renders the reel. Nothing leaves my laptop and a render costs nothing. A reel used to take me an hour to edit. Now it takes about 30 minutes, and only 5 to 10 of them are me. The real trick is not the editing. Every time I give feedback, Claude writes it down as a rule in the studio's instructions, so the next reel starts from everything I have already said. Since 6 October that rule book has grown from 180 lines to 293.
Why did I build my own video studio?
Because editing took an hour per reel, and that hour is exactly what kills a posting plan. I tested it in my series Will AI Fix It?, where I pick one real problem in the scaleup and give AI a week to fix it. Week 2 was video editing.
The verdict was KEEP. For the first time ever, all of Roccai's content for the rest of October is planned, edited and scheduled: two and a half weeks of content made in two days, while I did my normal CTO tasks in Roccai.
I didn't start from zero. I set it up with the free YouTube editor kit from WeAreNoCode. It is built for landscape YouTube videos, and I only make reels, so I had Claude rebuild it around one format: vertical 1080x1920 for Instagram and LinkedIn, with a square version as a fallback for the LinkedIn feed. Fewer formats, fewer decisions.
How does the studio work?
It is a folder on my desktop with a handful of scripts. Each one does one job:
| Step | What happens |
|---|---|
| Record | I record on my phone in Instagram's Edits app, with its teleprompter on when I need it. No filters, captions or music in the app. Then I AirDrop the raw video to my Mac and save it in the studio folder |
| Ingest | The sound is normalised to −16 LUFS, the video is set to 30 frames per second, whisper transcribes it and every pause is mapped |
| Tighten | Pauses longer than 0.6 seconds are shortened to 0.25. Shorter breaths stay, because removing every breath is what makes edited talking-head video sound wrong |
| Reframe | Only if the recording is landscape: it becomes vertical |
| Make the reel | A HyperFrames composition is built with my captions, the series title, the hook line, the point labels and the end frame |
| Render | One command, and the finished reel lands in a renders folder |
For messy recordings with lots of retakes there is a second set of tools. They keep the last take of every sentence, cut restarts, check that no cut lands in the middle of a word, and keep going until there is no dead air left. When I fluff a line, I pause for two seconds and say the whole sentence again. The studio knows the last take wins.
Music comes from Epidemic Sound only, and every track is written into a licence ledger before it can be used. Six months from now, nobody will remember which track was in which reel. The ledger will.
What can Claude do to a video?
In the reel that goes with this post, I just say what I want, and it happens on screen. Everything I ask for in it, the studio can do today:
| I say | What happens | What does it |
|---|---|---|
| "Give me captions" | One word at a time, under my chin, following my face when I move | whisper, on my Mac |
| "Zoom in" | A punch-in, anchored at the top so my head never gets cropped | HyperFrames keyframes |
| "Remove my background" | The wall behind me disappears | A local AI model in HyperFrames. Nothing is uploaded |
| "Put text behind me" | Big text sitting between me and the background | The background on its own layer, then the text, then me on top |
| "Music" | A low bed from Epidemic Sound | The Epidemic connector and my licence ledger |
| "Record our app" | Our product list scrolling, the guide played and swiped through, the results card | Claude in its own browser |
In real life I don't say it to the camera. I record first, then tell Claude what I want, and it builds the edit. But the commands are that short.
The one that saves me the most time is the last one.
Can Claude record screen footage of my app?
Yes, and this is the part I would not give back. Claude Code in the Claude desktop app has its own built-in browser. I sign in to Roccai myself, and then Claude opens the app, scrolls through the products and the question cards, and plays the guide as a visitor would, swiping cards until it lands on the results.
The browser takes screenshots, not video, so Claude builds the motion from stills: it scrolls a page in steps, stitches the screenshots into one long strip and pans across it, and it plays the guide on a phone-sized screen so the footage is already vertical. It stops before anything broken, like a missing image or a draft product. And it never fills in a contact form in the guide.
All I did was sign in. I did not press record once.
Can I edit from my phone?
I have tried. You can edit directly in Claude from your phone, and it works. But the best and fastest experience I get is in Claude Code on my computer, where the whole studio lives: the scripts, my brand, the music ledger and every rule from earlier reels. So my phone is for recording and my Mac is for editing.
How does it get better every time?
The idea comes from the WeAreNoCode kit: after every round of notes, you tell Claude "Save this to the skill." I made that a rule, so Claude does it without me asking.
Claude Code has two places to remember things. One is the studio's instruction file, SKILL.md, which describes how a reel is made. The other is a small memory folder with notes about me: what I like, what I rejected and why. When I give feedback, it fixes the reel, and then it writes the feedback down in one of those two places.
A few of my real notes and what they became:
- "The music is too high." "The whooshes sound weird." On one reel the music was ducked under my voice and pumped back up in every pause. I kept hearing a whoosh on every cut. Now the rule is: no whooshes, no ducking, a steady music bed about 16 dB under my voice.
- "It sounds like paper tearing." That was a synthesised pop. Now there is at most one pop, on the one moment that matters.
One reel, "clicks not reasons", took eight versions. Every one of those versions left a rule behind, and that is why a reel now takes 5 to 10 minutes of my time instead of an hour.
That is the whole point. I am not getting a better tool. I am training an editor, and it does not forget.
What is in the rule book?
The instruction file has my brand, the text layout I approved, how captions sit under my chin and follow my face when I move, how long a label stays on screen, when a sound is allowed, and a checklist before anything goes out. Watch it once muted. Watch it once at phone size.
There is one rule about the rules: when a new note contradicts an old rule, the old rule is replaced. Never both. Otherwise the file slowly fills up with arguments with itself.
How do I set it up myself?
I made my setup into a free kit, so you get the reel version straight away: download the Reel Editor Kit.
It is built on what I learned from the WeAreNoCode YouTube editor kit, rewritten from scratch for vertical reels. If you make landscape YouTube videos, use theirs instead.
You need a Mac and the Claude desktop app with Claude Code. Plan on about 30 minutes.
1. Install the plugins
Two, and you install both inside Claude:
- HyperFrames, the plugin that turns your edit into a video. In Claude Code, type
/plugin marketplace add heygen-com/hyperframes, then/plugin install hyperframes@hyperframes. - Epidemic Sound (optional), for licensed music. Connect it under Connectors in your Claude settings. It needs an Epidemic Sound subscription.
Everything else, Claude installs for you in step 3. The browser that records your app is already built into the Claude desktop app.
2. Unzip the kit
Put the folder somewhere you will keep it, for example on your desktop. In the Code tab of the Claude desktop app, choose that folder as your project.
3. Paste this prompt
Set up the reel editor in this folder.
1. Run bin/setup.sh and install anything it says is missing. Tell me in plain words what you installed.
2. Install skill/SKILL.md as a skill called reel-edit in this project's .claude/skills folder.
3. Ask me about my look, one question at a time: my colours, my font, how I want captions, and the names
and words whisper might get wrong (my name, my company, my products). Write the answers into
STYLE-GUIDE.md and update templates/reel/compositions/_brand.css to match.
4. Then ask me for a short test clip of me talking, and edit it as a reel.4. Edit your first reel
Record on your phone with nothing on it: no captions, filters or music. Move the raw video into the kit's projects/inbox folder, then tell Claude: "There's a new clip in projects/inbox. Edit it as a reel."
Watch it on your phone and say what you don't like. The kit saves every note as a rule by itself, and that is the part that makes it better.
Is it better, or just faster?
Faster, yes. Better? Too early to say. So far every rule comes from me: what I hear and see when I watch a draft on my phone. That only tells you what I like, not what makes you keep watching.
That is week 3 of Will AI Fix It?. The loop:
- Pull the numbers for each reel from Instagram 48 hours after posting: average watch time, views, skip rate, saves, shares and comments.
- Screenshot the retention graph, because the second-by-second drop-off is not in the API as far as we have found.
- Match the second where people drop off to the timed script for that reel.
- Turn what we find into a rule in the studio, about the edit or about the script.
- Edit the next reel with the new rules, and measure again.
If average watch time goes up on the reels made with the new rules, it is better, not just faster. I will share the verdict.
FAQ
Can Claude edit videos?
Yes, with tools. Claude Code can run FFmpeg, whisper and HyperFrames on your computer, so it can cut, caption, add text and render. It does not edit in a timeline you drag around. It edits through scripts and plain HTML.
Do I need to code to edit videos with Claude?
No, but you need to be comfortable with Claude running commands in a terminal on your computer. I describe what I want and Claude runs the scripts.
Do I need editing skills to make professional videos with Claude?
No. I never open an editing program or drag clips around a timeline. Claude does the cutting, the captions, the text and the render. What you do need is to know what you like and say it out loud: "the music is too high", "that sounds weird". Your taste is the skill, and every note you give becomes a rule, so you only have to say it once.
Which plugins do I need to edit videos with Claude Code?
The HyperFrames plugin, and the Epidemic Sound connector if you want music. Claude installs the rest itself when you set up the Reel Editor Kit.
Is the Reel Editor Kit free?
Yes. It is MIT licensed, so you can use it and change it. It builds on what I learned from the free WeAreNoCode YouTube editor kit, and it credits them.
Does my footage get uploaded anywhere?
Not in my setup. Transcription, cutting and rendering all run locally on my Mac, so no footage leaves the machine and there is no cost per render.
How does Claude remember my feedback between videos?
Through files. The studio's instructions live in SKILL.md, and Claude Code keeps a memory folder of notes about my preferences. When I give feedback, Claude updates one of them, and the next session reads them before it starts.
What is HyperFrames?
A framework that turns an HTML composition into a video. My captions, titles and labels are plain HTML and CSS, which means Claude can edit them as easily as a web page.
Can Claude add captions to a video?
Yes. Whisper transcribes the audio, and Claude places the captions. Mine show one word at a time, white on a dark box, just under my chin. Read them before you publish: whisper mishears names.