I stopped editing videos because it took too long. Here's the Claude setup that brought me back.

How I gave Claude 782 tagged clips, my Final Cut habits, a sound bank it can read and 227 of my own rules, and went from 2 to 4 hours per reel to under an hour for a concept and all its versions.

A prompt that asks about your product first, then adapts every step. Works in ChatGPT too.

Hey!

My name is Lambert. I run growth at Kai, and I’ve loved editing videos since I was 16.

Today I edit our reels with Claude. This is the setup behind it: how Claude learned our 782 clips, the way I edit in Final Cut, the beats of our songs and the 227 notes I’ve given it. A reel used to take me 2 to 4 hours. Now a concept and all its versions take under an hour.

Claude building one of David's reels, replayed in 25 seconds.

The summers I spent editing

At 16 I watched everything Rory Kramer and ThatOneBlondKid put out. Rory’s video for Avicii’s The Nights is his own life cut to the song: cliff dives, car surfing, water-skiing. ThatOneBlondKid’s That Kid On Tour videos with The Chainsmokers (and part 2) were on repeat, and later the Up & Down music video he directed. Fast cuts, the music landing on the exact frame, text that moves with the people in the shot.

I wanted to make that. So I spent my summers trying.

Eight seconds of one of my summer edits.

Then I stopped. Getting the result I had in my head took way too much time, and I never had enough of it.

Kai brought it back, and it still took too long

This year we started filming everything we do at Kai: hackathons, demos, the kitchen, the pool. By May we had a real bank of footage, so I tried again.

The intro video for David, our CEO, took three months from script to shoot to edit. The edit alone took a month and a half.

David's intro, on Instagram. A month and a half in Final Cut.

Even a simple reel, like our POV of building a startup with the boys in Switzerland, took me 2 to 4 hours. And that one has almost no editing in it: find the shots, pick the music, put them on the timeline, add the text.

The Switzerland POV, on Instagram. Almost no editing, still 2 to 4 hours.

In a startup, 3 hours for a reel you’re not even sure will work isn’t doable. So I parked it again.

End of September: “improve it”

At the end of September I gave Claude an old edit of mine and literally typed: improve it.

I watched what came back and I was like, “Wait… it can do THAT?”

I’M EDITING AGAIN!

Well, almost. Next I gave it a reel I’d saved on Instagram and said: let’s cook something like this with our clips. The edit was great and the clips were wrong. It had no idea what was in our footage, so it grabbed whatever looked close.

Claude could edit. It didn’t know my footage, my taste or my sounds. So I gave it those, one at a time.

Step 1: a bank of real footage, tagged by Claude

We have 782 clips catalogued today, from Zurich in February, Maz in May, Caparica in July and our weeks in Lisbon this fall. They sit in a shared Google Drive.

Claude goes through every clip and writes down what someone would need to find it without watching it: who’s in frame, what they’re doing, where, the light, the energy, the shot size, where there’s room for text, and whether it’s worth keeping. It does that per segment, every time the framing or the action changes.

Here’s a real entry. It’s me, ranting at my AI agent from a rope net in Caparica:

{
  "file": "Lambert-talking-to-claude-2.MOV",
  "start_s": 6.0,
  "end_s": 84.0,
  "who": ["lambert"],
  "action": "working-laptop",
  "place": "house",
  "light": "day",
  "energy": "hype",
  "shot": { "size": "medium", "angle": "high", "camera": "handheld" },
  "text_space": ["top"],
  "keep": "yes",
  "moment": "Lambert argues out loud with his AI agent from the rope net, hands chopping the air"
}

So when I ask for “me working, high energy, room for a title at the top”, Claude knows which 78 seconds to pull.

A Finder-style window titled Kai, Video, Footage: a grid of 42 clip thumbnails from Zurich, Maz, Caparica and Lisbon on the left, and on the right the entry for Lambert-talking-to-claude-2.MOV, 114.5 seconds, with a timeline split into three segments (the middle one, 6.0 to 84.0 seconds, selected) and its tags: lambert, working-laptop, house, day, hype, medium, high, handheld, text space top, keep yes. — Part of the bank, and the entry above as the catalogue holds it: three segments, the middle one is 78 seconds of me ranting.

When it picks a shot I don’t want, I say no and the shot goes on a reject list it checks before every edit. 14 shots are on it today.

Why I don’t generate clips

You can generate clips with AI now. For content about you or your company, I think that’s a mistake.

People follow humans. That’s why building in public works for a company: people watch real people cooking, arguing and shipping, and they decide they like you long before they need your product.

So every frame in that bank is one of us.

Demo night in the Zurich flat: six of the team on a sofa and armchairs with laptops, all laughing at something on screen. — Demo night in Zurich, February.

Step 2: give Claude the edits that already worked

My taste was already sitting in Final Cut: years of projects, the titles I always reach for, my filter, my sound effects.

So I let Claude read my projects. It doesn’t click around in Final Cut. It reads the project files, writes down what I do, and renders the edit itself with code.

Here’s what it found in one project, the third version of David’s intro:

116 titles, almost all in Arial Bold, most of them at size 60.
109 of those titles with the same blur-in effect.
One full-length layer I’d named “Long”: a faint Newsprint halftone with a vignette on top. That’s my look, and it now goes over every reel.
A half-second shutter click I place 15 to 22 times per video. In that project, 16 of the 22 land within 2 frames of a caption word appearing.

I’d never noticed that last one. I just did it.

The replay's Final Cut-style window: the first clips of David's intro landing on the timeline over the Denzel Washington speech, a garden shot in the viewer. — Five frames from the replay: the shots land, the titles go in, then the Long layer goes over everything.

The timeline nearly full, a pool shot in the viewer. — Five frames from the replay: the shots land, the titles go in, then the Long layer goes over everything.

One square shot of the team tying on matching aprons, split down the middle: the left half raw, the right half with a faint dot texture and darker corners. — Left, the raw shot. Right, the same shot under the Long look, rendered by Claude's code: Newsprint at 6%, then the vignette.

Step 3: a sound bank Claude can read

Music is half the edit, so every track lives in a bank with its measurements. Each one starts on its first sound, and its beats are measured on the audio, not by ear. In Runaway (yes, the Kanye one), the piano hits land every 1.381 seconds and the drop is a jump of more than 50 dB below 120 Hz. Claude cuts on those numbers.

Runaway's first 46 seconds: a waveform with a vertical line on each of the 28 piano hits, 16 on E then D sharp, C sharp, A, G sharp and E, a red line at the drop at 37.92 seconds, and below it the level under 120 Hz sitting around minus 65 dB before jumping 53 dB at the drop. — The first 46 seconds of Runaway, measured: one cut per piano hit, then the low end jumps 53 dB on the drop.

Speeches work the same way. When a reel runs on a famous speech, every word has a timestamp, and every new sentence gets a new shot.

Step 4: a skill that asks the right question at the right time

All of this lives in one Claude Code skill. When I start a video, it asks only what it can’t guess: who it’s for, which format, what carries it (a track, a famous speech or one of us talking) and, when I give it a reference, whether I want an exact copy or only its rhythm.

A Claude Code terminal: the request /reel same as this one, for us, with an Instagram link, then Claude saying the reference gives it the format and the sound, and asking two things: who it's for, and an exact copy or its spine and rhythm with our own patterns. — How /reel opens when I paste a reference and nothing else.

Then I describe the idea. I open Wispr Flow and talk for about 10 minutes: the concept, the moments, the feeling I want at the end.

Every note becomes a rule

This is what takes the most time, and it’s why the rest works.

Every time I correct a video, the correction becomes a dated rule in a file Claude reads before every edit. There are 227 of them, written between 2 and 30 October.

One example. I asked for a reel in the format of someone else’s video, and v1 copied his 35 cut times, his fisheye vignette and his type. My note: “I’m scared you copied him exactly, when I thought you’d just take the learnings.”

The rule it wrote: “A reference is a pattern to learn from, not a video to copy, unless Lambert confirms an exact copy.” Now it asks me once, every time.

A window titled skill-rules.md with three pairs: a note from Lambert, and below it the dated rule it became. Cropped clips became 4:3 when most of the footage is horizontal; the copied reference became a reference is a pattern to learn from; fun night shots became a day-versus-night cut's night is only us at work. — Three of my notes, and the rules they became.

That’s where your own expertise comes in. Claude can edit. It can’t know that you hate a slow pace, or that your captions breathe after a pause, until you tell it and it writes it down. Every one of the 227 rules started with me watching a version and saying what was off.

What it still gets wrong

It still picks the wrong shot sometimes. I say no, and the list grows.

A window titled rejected-shots.tsv with three crossed-out frames: a dim curved monitor in the Zurich flat, a balcony over a sea of fog, a tray of croissants filmed sideways, each with Lambert's note, the date and the video. — Three of the 14 shots on the reject list, with what I said about them.

It’s not great at emotion. It can tell a laugh from a frown, but it can’t always tell which laugh makes you feel something. So I watch every version before anything goes out.

And some ideas just die. Claude’s first rebuild of David’s intro reordered my lines and slowed everything down. The rhythm was gone. A reel called Make It Harder got dropped after 5 versions.

Where I am now

A simple reel used to take me 2 to 4 hours. Now a concept with several versions takes 30 minutes to an hour, and 10 of those minutes are me talking into Wispr Flow.

What happens after the edit (testing several versions of each idea on Instagram and reading where people stop watching) is the next article. Jim and I talked about it in How We Grow #018.

If you’ve got a folder of footage, start where I started and get Claude to tag it. The “Run this playbook with Claude” button at the top of this page sets up the same system around your footage, your edits and the way you review :)

Then tell me how it went on LinkedIn. I read every single message!

Video

Video Editing

Claude Code

Build in Public