The film editor
Thirty people record a message. A finished film and a printed book come out the other end.
“You still owe me a pizza from 2009.”
29 people · 67 clips · cut by theme, not by person · 6 min 59 s
Sample with invented data. The real one looks the same.
What it does
- When an organiser picks the occasion, the questions and the people, the website invites each person, reminds them and closes on the deadline.
- When the clips arrive, it listens to every one, in English and Hindi, and marks where each sentence really ends, so no cut lands mid-word.
- When it plans the edit, it cuts by theme instead of person by person: running jokes, callbacks, a bloopers tail and a final wall of "happy birthday".
- From the same edit it makes a short trailer, captions in a second language and a printed photo book where a QR code on each page plays that person's clip.
How it works
1 · Hears
Every clip is cleaned up and transcribed on the laptop's graphics card, word by word, with the language set per clip.
2 · Thinks
One Claude Opus call reads the whole set and returns the cut as a list. Code checks the list. A no-AI baseline cut exists, and the AI has to beat it.
3 · Does
ffmpeg renders the film from the list: cuts snapped to word edges, captions, name boxes, a music bed. A QA pass looks at a frame of every item.
What broke
Sentences still cut off
"Is Opus not doing a good job? How can we confirm what's wrong?" Measuring showed the renderer was cutting, not the model. Cuts now snap to sentence ends.
Hindi heard as three other languages
Automatic language detection labelled Hindi clips as Urdu, Maori and Javanese. Language is now set from the sign-up form.
Music nobody could hear
The music bed was approved by listening to the track alone. In the final mix it sat 13 dB under the voices. Now every check runs on the finished file.
By the numbers
- 29 of 30
- people recorded for the first film
- 10 min
- first order, start to delivered, unattended
- $49
- for a film; $109 with the book
The call that mattered
The model never touches pixels. It decides the cut; code and ffmpeg do the rest.
Built with
WhisperX on an RTX 3060 · Claude Opus (one decision per film) · ffmpeg · Cloudflare Workers + D1 + R2 · Python book layout