A kit for Claude Code

Edit your podcast by talking to it.

The AI Video Editor turns Claude Code into a video editor. Drop in your camera files and your audio, say "edit the podcast", and it syncs, cuts, frames, captions and renders. Then it cuts the reels.

Updated November 2026 · 33 cards · 4 caption styles

Ask for an invite

The kit lives in a private GitHub repo. Send me your GitHub username and I'll add you.

1I invite you2You accept the email3Sign in to GitHub4Paste three commands See the install commands

What it does

The jobs an editor does on every episode

It's built for the most common job: two or three cameras, a conversation, and short clips cut from it afterwards. Each step is a script Claude runs for you. You never type a command.

Multicam

Switches to whoever is talking

Picks the camera from the mics, adds close-ups when the footage is 4K, and holds wide shots for back-and-forth.

Reels

Cuts short clips from the long one

Finds the strongest moments and keeps each clip a complete story with a start, a point and an ending.

Captions

Burns in captions in your style

Uses your font, colour and size. The words come from the saved transcript, so they match what was said.

9:16

Reframes wide video for phones

Follows faces to turn 16:9 into 9:16 or 4:5, leaving room above the head.

Sync

Lines up every camera and mic

Matches cameras to the audio by sound, so you don't need a clap or a slate.

Cut

Takes out the dead air

Cuts from the transcript. It leaves a breath after a sentence ends and never cuts in the middle of a word.

Transcript

Writes down who said what

Transcribes on your own computer, free, with the speakers named. It's saved, so it never runs twice.

Sound

Masters the audio

Evens out the voices and sets loudness to a standard target: one for the full episode, another for reels.

Checks

Tests the render before you see it

Checks for leftover dead air, sound out of sync, blank frames and wrong length. A clean finish is not enough proof on its own.

How it learned

How it learned to edit like me

It started on 13 July 2026. Claude edits, I watch and correct, and each correction becomes a rule or an automatic check. Anything I used to catch by eye, like dead air, sync or loudness, is now measured before I see the video.

1,021 saved changes
89 working days
215 episode transcripts saved
22 multicam masters with a named transcript
128 of my corrections turned into permanent rules
90 locked edit rules ship in the kit
It saved me over 300 hours in ten weeks. And every note I give it once becomes permanent, so the next edit needs fewer notes.
Daniel Ro
CapCut

It read my CapCut projects to learn how I cut, trim and place B-roll.

Measured from the editor's history, 28 Oct 2026. Started 13 July 2026. Hours saved: 31 episodes and 257 reels from 19 Aug to 29 Oct, counted at 4 hours per finished hour and 1 hour per reel by hand, minus the 23.5 hours I spent directing the edits.

How it works

From install to your first edit

  1. 1

    Install Claude Code

    Anthropic's coding assistant. It runs in a terminal or inside VS Code.

    claude
    Claude Code is ready
  2. 2

    Add the kit

    Paste the three install commands. Claude Code downloads the kit from GitHub.

    /plugin marketplace add
    /plugin install
    kit installed ✓
  3. 3

    Run setup

    The third command runs it. It asks a few questions, installs what's missing and does a test render.

    /ai-video-editor:setup
    Show name? My Show
    test render passed ✓
  4. 4

    Point it at footage

    "Edit the podcast in this folder." It shows you the plan before it renders.

    edit the podcast
    11 min of dead air out
    render it? yes
  5. 5

    Review and approve

    Watch it and tell it what to change in plain words. It re-renders only what you named.

    B camera looks orange
    fixed and re-rendered ✓

Setup asks you

Your show becomes a settings file

Your answers are saved as one file, shows/your-show.json. Every edit reads from it, so the look stays the same across episodes. For a second show or client, run setup again to get a second file. You never copy the kit.

1Show or brand name
2Mac or Windows
3Brand colours
4Headline and body fonts
5Caption style
6How long a reel should be
7Camera switching: calm, normal or lively
8Where finished videos go
9Your API keys

Cards

33 on-screen cards, in your brand

Cards are the graphics on the video: a key point, a big number, a list beside the speaker, a map, a chart. They're built with Remotion, which makes video graphics from code, so they stay sharp and every card on the show matches.

Setup puts in your colours and fonts. For each episode, you write the words and Claude places them where they're said. Cards work the same on Mac and Windows.

Podcast cards

9

Full-screen and side-by-side cards for the main points of an episode.

Key pointA frosted label under the speaker with a badge like THE POINT.
Big numberOne figure, large, with the video pulled to the side.
Side statementThe speaker full height on one side, a short headline on the other.
ListTwo to four points beside the video, as they're said.
Highlight listBolder keywords beside the video, for the result or the takeaway.
CompareThe usual way against the better way, in two columns.
StatementA sentence full screen with a small heading. The most used card.
Section card
SectionOpens a chapter of the episode, full screen.
IntroOpens the episode with a quote and what it's about.

Motion cards

13

Short animated moments, three to eight seconds, for things you show or name.

Proof highlightA real web page fills the screen and a marker sweeps the line that proves the point.
Flow mapBoxes joined by lines that draw themselves, until the result lights up.
Floating pageA website screenshot flies in tilted, then settles.
Toolkit gridA grid of boxes, each one lights up as you name it.
Task chipsLabels pop in beside the speaker as they list things.
Chapter tagA small tag in the corner naming the part of the video you're in.
Takes, crossed outEvery attempt as a line, crossed out until only the one that worked is left.
Word, then meaningA term in the accent colour, then what it means.
Post on screenA real social post rebuilt as a clean card, the key line swept.
ChatA chat window types the question and the answer streams in.
Card stack counterScreenshots flip by like cards while a number climbs.
Framed shotThe camera shot shrinks into a rounded card with a title.
Typed searchA search bar grows across the screen and types what's being said.

Bold cards

11

Big, high-contrast text for the lines you want to land hard.

Two-tier headlineThe set-up line in white, then the punch line big and bright across the bottom.
Keyword beside the speakerThe shot slides the speaker left and big text fills the right half.
Label pillsOne big label each side of the speaker, the second a beat later.
Question, then answerThe question in a dark bar, the answer lands under it.
Labels across the bottomWords rise in one at a time along the bottom.
Icon listIcon and word rows stack down one half of the frame.
Icon calloutAn icon pops, the word lands, a line drops to the next word.
Icon checklistFull screen, one step at a time, each with an icon.
Timeline dotsA dot lands for each step, then they pull together into the result.
Bar chartBars grow in and a trend line draws over them.
Silhouette headlineA silhouette on one side, the set-up line, then the punch words circled.
Real cards from finished videos. Yours use your colours and fonts.

Captions

Captions that match what was said

Every reel gets captions burned in. The words come from the saved transcript, so they're right the first time, and the timing follows the voice word by word.

Setup picks your font, colour and size. The word that matters takes your accent colour. There's no box behind the text, only a soft shadow, so it reads on any shot.

The four defaults

PopItalic, two or three words at a time, the key word in your accent colour. The default.
SpotlightThe whole line shows and the key word lights up as it's said.
LineOne or two words at a time, large, for slow and dramatic moments.
CleanUpright and simple. The key word takes your brand colour, orange here.

All four defaults play the same sentence, so you can compare them. Setup asks which one is your default.

Or make them yours

These are the defaults. Your captions can look like anything: your font, colours, size, position, even chat bubbles. Tell Claude what you want and it saves it to your show.

Teal key word
Red key word
Chat bubbles
Gold key word

Colour

Colour grading, by asking

Tell Claude what you want: lift the shadows, more contrast, add an S-curve, warmer skin. It pulls frames from each part of the video, shows you before and after side by side, and renders the one you pick. White balance, matching two cameras and straightening a tilted shot happen on their own.

Pick a look

Finishing

For frame-by-frame finishing, DaVinci Resolve is faster than going back and forth. Export the edit and every cut comes with it.

A woman on a gym bike, graded: Clean
A woman on a gym bike, graded: Original
Original Clean

Your editor

Rough cut here. Finish in your editor.

The kit makes the rough cut: the cuts, the camera switches, the captions and the cards. Then it exports it as a project, so you can open it in your own editor to fine-tune it and colour grade it there.

Already on the timeline

  • Every cut
  • Camera switches
  • Captions
  • Cards
  • Audio
Ready

CapCut. Opens as a CapCut desktop project.

Ready

DaVinci Resolve. Opens as a Resolve timeline (FCPXML), tested in Resolve Studio 20.

Premiere Pro. Coming soon: Premiere reads a different file type, not tested yet.

Proof

Every video on my YouTube channel is edited with this

Same kit, same steps as above. These are the latest ones.

See the channel

What you need

Mac or Windows

Setup installs the free parts for you. This is everything that ends up on your computer, and what it costs.

Mac

Apple chips

Recommended
Claude Code
Claude Pro, Max or an API account
Cost: Your plan
ffmpeg, Python and Node
Installed by setup
Cost: Free
Disk space
About 4 GB for the tools, plus room for your footage
Cost: Your drive
Transcription
Runs on your Mac, faster on Apple chips
Cost: Free
Fast encoding
Built into Apple chips
Cost: Free
Fast external SSD
Recommended, USB-C or Thunderbolt. Keep about 3× your footage free while editing.
Cost: Your drive

Windows

Faster with an NVIDIA card

Claude Code
Claude Pro, Max or an API account
Cost: Your plan
ffmpeg, Python and Node
Installed by setup
Cost: Free
Disk space
About 4 GB for the tools (measured on a Mac), plus room for your footage
Cost: Your drive
Transcription
Runs on your PC, faster with an NVIDIA card
Cost: Free
Fast encoding
With an NVIDIA card, otherwise slower
Cost: Free
Fast external SSD
Recommended, USB-C or Thunderbolt. Keep about 3× your footage free while editing.
Cost: Your drive

Optional: ElevenLabs or HeyGen, only for an AI voice or avatar, at their pricing.

What setup installs

What goes on your computer, and why

Setup installs these first, before it asks you anything. All of them are free.

1

Homebrew

Mac, if you don't have it

Size: 0.3 GB

The standard installer for developer tools on a Mac.

Setup uses it to install ffmpeg, Node and Python. It asks for your Mac password, so you install it once yourself, from brew.sh. On Windows, setup uses winget, which comes with Windows.

2

ffmpeg

Size: 0.4 GB

The video engine inside most editing apps.

It cuts, joins and exports your video, masters the sound, and burns in Clean captions.

3

Python and its libraries

Size: 0.6 GB

The language the editor's scripts are written in.

Face tracking for framing, sound analysis for camera switching, and the code that runs Whisper.

4

Node and Remotion

Size: 1.5 GB

Remotion turns code into video. Node runs it.

It draws the cards and the Pop, Spotlight and Line captions, with its own copy of Chrome to render them.

5

Whisper

Size: 1.4 GB

A free speech-to-text model that runs on your computer.

It writes the transcript with the time of every word. Cuts, reels and captions all come from it, and nothing uploads.

6

Face, hand and noise models

Size: 11 MB

Three small model files.

Finding faces for framing and close-ups, hands for gestures, and cleaning up background noise.

Total About 4 GB, plus room for your footage

Limits

What it doesn't do yet

This is a working framework, not a finished app. Know these before you start.

Mac is the tested setup

Everything was built and proven on Apple Silicon Macs. Windows should work, but test it on one episode before relying on it.

Speed depends on your computer

On an Apple Silicon Mac, transcribing a 90 minute episode takes about 50 minutes, and it only runs once. Older or slower computers take longer.

You bring your own keys

The kit has no accounts or keys in it. Each person uses their own and pays for their own use.

Posting is an add-on

The kit makes the files. Posting to YouTube and Instagram is available as an add-on.

Built around conversations

Podcasts, interviews and talking-head videos work best. Fast-cut music videos and heavy effects aren't what it's for.

Get the kit

Install it in Claude Code

These go into Claude Code in the Terminal (PowerShell on Windows), not into the Claude desktop app and not straight into the Terminal. First install Claude Code, open the Terminal, type claude and press Enter. Then paste these one at a time. The kit installs as a plugin, Claude Code's word for an add-on, so you get updates when new versions come out.

1 /plugin marketplace add danielro-ai/ai-video-editor-public
2 /plugin install ai-video-editor@ai-video-editor-public
3 /ai-video-editor:setup

The repo is private for now. Sign in to GitHub on your computer first, or the first command can't find it.

Request access

The repo is invite-only. Ask once and you get the invite on GitHub.

Questions

Before you ask

A repo is the folder of code on GitHub. A skill is a set of instructions that tells Claude how to do one job. A plugin packs the code and the skills together so Claude Code can install them in one step. The kit is all three: a repo, installed as a plugin, full of skills.

No. Your video and audio stay on your computer, and transcription runs locally too. The only things that leave your computer are your chat with Claude and any calls to services you set up yourself, like an AI voice.

Yes. Each one gets its own settings file. Say which show you're editing and it uses that one's look.

Yes. That's what it's for. The code is plain Python and TypeScript, and the rules that make the edits look right are written down in the repo.

In a file named .env on your computer. The kit never commits it, so it never ends up on GitHub.

Privacy and safety

What stays on your computer

The questions developers ask first, answered from the code.

1

Privacy

  • Your video never uploads. Transcribing, cutting, cards, captions and rendering all run on your computer.
  • Claude Code reads the transcripts and the stills it checks, not your footage. That text goes to Anthropic like anything else you do in Claude Code.
  • Nothing reports back to me. No tracking, no usage stats.
2

Keys

  • Your keys live in a .env file that only your user account can read, and git ignores it.
  • An Anthropic API key is only needed for the AI reel pass. It's paid per reel and only runs when you ask for it.
  • None of my keys are in the repo. Every update is scanned for keys before it ships.
3

Your files

  • Your camera originals are never changed. Every edit writes a new file.
  • The only files it deletes are its own temporary ones.
  • Setup asks before it installs anything on your computer.
4

Backups

  • Finished videos go to the export folder you pick at setup.
  • Transcripts are saved, so a long episode is only transcribed once.
  • Your camera files are yours to back up. The kit doesn't copy them anywhere.

AI Video Editor · Built by Daniel Ro from a production editor used on real podcasts every week.

Not sure where to start? Ask me.