← Back to the vault

From @rajeevkugan

How I make reels with Claude.

The reel you just watched was built from one script file. Claude times the words, places the sound, writes the 3D and renders it. No After Effects, no keyframes. Here's the method.

The method

Five layers, one file

01

Rhythm

Whisper gives every word a time. A small script snaps each one to where the sound really starts, and the captions land 4 frames early, so they read on time instead of late.

02

Sound

Claude picks effects from my own library and writes a cue list: at, id, peak. Each sound starts early by its own peak time, so the hit lands exactly on the frame.

03

3D

Scenes like the laptop are three.js, rendered frame by frame in headless Chrome to a clip. Claude writes the object, the light and the camera move as code.

04

One script

Every on-screen word comes from a file. Change a word there and the whole reel re-renders with it. The video is Remotion: React components drawn per frame.

05

Voice

The voiceover is my cloned voice. I approve the script first, then generate each hook as its own take and the body once.

06

Direction, not keyframes

My job is the brief and the notes. Claude renders stills of the risky frames, checks them, then renders the full video and the mix.

Setup

Build it in Claude Code

  1. Install the basics. Node 20 or newer, ffmpeg, Google Chrome and Claude Code.
  2. Get a voiceover. Record yourself, or use a voice clone you have the rights to. Save it as a WAV.
  3. Open Claude Code in an empty folder and paste the first prompt with your script.
  4. Review stills, give notes, then add sound and 3D with the next prompts.
1. Script to timed reel
Build a vertical reel (1080x1920, 30 fps) in Remotion from my script and my voiceover file vo.wav.

- Put every on-screen word in one script file (JSON) and have the composition read it. Nothing typed into components.
- Time the words with whisper.cpp (word-level timestamps), then snap each word to its real onset in the waveform. Use the script for the text and whisper only for the timing.
- Show captions word by word. Each word appears fully visible 4 frames before it's spoken: no fade or blur in, at most a 3-frame settle from 4% oversize.
- Keep text at least 110 px from the left, 150 px from the right and 230 px from the top, and nothing but captions below y 1450.
- Solid text colours only: no background-clip text and no mix-blend-mode.
- Load fonts from local files before rendering.

Before the full render, render stills of the first word of every section and check them yourself. Then render and tell me where the file is.
2. Sound on the frame
Add sound effects from my folder ./sfx.

- Index every file: length and the time of its loudest point (the peak).
- Write a cue list file: at (seconds), id, volume. Start each sound at (at - peak) so its peak lands exactly on the frame.
- One different whoosh per section change, a soft hit when something locks into place, and nothing on every word.
- One master SFX gain (start at 0.4) so the voice always stays on top.
- Mix: two-pass linear loudnorm to -14 LUFS, then a limiter. Check the final file for loudness and dropouts.
3. A 3D shot as code
Add a three.js laptop reveal: the lid opens, the camera dives through the screen until my screenshot (16:10) fills the frame.

- Render it in headless Chrome with a renderFrame(f) function that draws frame f the same way every time, then encode the frames with ffmpeg into a silent 30 fps MP4.
- Show me stills at the start, middle and end before the full render.
- Play the clip in the reel where the script says "laptop".
4. Change it with one word
In the script file, change the title to "MADE BY CLAUDE" and re-render. Then show me the frame where it first appears.
Go deeper. The sound side is in my Claude SFX setup and the animation library in my Claude animation setup. Remotion is free for individuals and small teams, but larger companies need a licence, so check their terms. Only clone a voice you own or have permission to use.

Want it done for you?

If you want scroll-stopping video for your brand, that's what we do at FROZN Media. DM me @rajeevkugan.

DM on Instagram ↗
Rajeev Kugan@froznmedia · @djicykle · @devbites · #EveryPixelCounts
Follow ↗

Tools and commands can change. Check each tool's own docs if a step doesn't match.