Watch on Instagram The idea
Claude is good at deciding what a video says and how it’s laid out. It’s unreliable at making pixels and audio by itself. So we split the job:
- Claude writes the script, the on-screen text, the scene timing and the code for each scene.
- Free tools turn that code into frames, voice, music and files. The same input gives the same output every time, so you can fix a mistake and re-render.
Part 1 is setup: getting the five tools installed.
The 5 tools you need
Only HyperFrames installs as a Claude skill. The other four are tools Claude runs from your project folder.
| Tool | What it does | Get it |
|---|---|---|
| HyperFrames (skill) | Turns HTML scenes into MP4 video | npx skills add heygen-com/hyperframes (needs Node.js) |
| OmniVoice | Voiceover, including cloning your own voice | Install steps in its README (needs Python) |
| NumPy | Makes the music bed and sound effects | pip install numpy |
| FFmpeg | Trims, speeds up and mixes audio, handles video files | winget install Gyan.FFmpeg |
| Playwright | Screenshots carousel slides from HTML | pip install playwright then playwright install chromium |
Setup in 4 steps
Install Node.js and Python
HyperFrames needs Node.js and the other tools need Python. On Windows, open PowerShell and run:
winget install OpenJS.NodeJS.LTSClose PowerShell and open a new one before the next step.winget install Python.Python.3.12Make a project folder and open Claude Code
mkdir "D:\Videos\video-edits"; cd "D:\Videos\video-edits"; claudePaste the setup prompt
This is the prompt from the reel. Claude installs all five:
Hey Claude, install this skill: https://github.com/heygen-com/hyperframes Also install OmniVoice (https://github.com/k2-fsa/OmniVoice), NumPy, FFmpeg and Playwright CLI.
If something is already installed, Claude skips it and tells you.
Check it worked
Ask Claude to make a tiny test video:
Using /hyperframes, make a 5-second test video with my name in big text, then render it.
If you get an MP4, you’re ready for Part 2.
What each tool does in the pipeline
- Script. Claude writes the hook, the beats and the on-screen text.
- Voice. OmniVoice reads the script aloud, one scene per file.
- Clean-up. FFmpeg trims the silence and sets the speed.
- Music and sound. NumPy builds the music bed and effects, mixed quietly under the voice.
- Scenes. HyperFrames renders each HTML scene to MP4.
- Carousels. Playwright turns HTML slides into 1080x1350 PNGs.
Use your own voice, or one whose owner agreed in writing. OmniVoice needs a 3-10 second clip of the voice.
What we learned the hard way
| Problem | What to do |
|---|---|
| Reels with no face and a synthetic voice lost viewers in about 2 seconds | Keep a real person on camera. Use the graphics as cutaways between shots. |
| Decorative motion looked slick but said nothing | Every moving thing should show a number, a change or a cause and effect. |
| Checking every frame by eye is slow | Let Claude run npx hyperframes check, then look at one small contact sheet. |
| Made-up numbers destroy trust | Label every example as an example. Use real figures only from real records. |
Common errors
Windows error 1455 when loading the voice model
Your PC has run out of memory. Close other apps and try again. OmniVoice runs on a 4 GB graphics card in half precision."npx is not recognized"
Node.js isn’t installed or PowerShell hasn’t picked it up. Reinstall it (step 1), then open a new PowerShell window.Playwright says the browser is missing
Runplaywright install chromium once.