thought-to-text

A 20-second film about brain-computer interfaces, made by decompiling someone else's viral edit frame by frame and rebuilding it with code.

how do you speak, when you can't say a word? you dont. you just think it. through signal. intention. language. through ones own mind, straight to T E X T

The reference was Jordan Watkins' "i'd love to see AI try and recreate this": a kinetic poem built from pixel-art objects, thermal silhouettes, typewriter cards and a score that goes quiet in exactly the right places. This repo is the answer to that caption.

How it was made

reference.mp4
   |  tools/teardown.py      cuts, beats, palette, onsets, 0.15s frame strips
   v
reference/TEARDOWN.md       beat table, verbatim copy, motion vocabulary
reference/template.json     same structure, content replaced by slots
   |
   |-- film/   Remotion (React). 14 scenes, all in absolute seconds, seek-safe
   |-- score/  SuperCollider NRT score, every hit placed on a cut
   |           + ElevenLabs voiceover placed on the original VO's word timings
   v
thought-to-text.mp4

layer	what	tool
analysis	14 hard cuts, 108 BPM, every cut within 0.15 s of a music onset	tools/teardown.py, librosa, faster-whisper
picture	3D-extruded pixel sprites, thermal figures, grain, ink, pen loops, typewriter cards	Remotion 4, Google Sans
assets	14 sprites, 2 thermal plates, 1 paper scan	gpt-image (prompts in film/assets/prompts/)
score	lo-fi keys, 808, chip arps, spray hiss, typewriter clacks, silence for the coda	SuperCollider
voice	5 lines, "Will" voice	ElevenLabs

The rule that made it work: keep the timing, swap the nouns. The reference's pacing (slow open, 0.5 s strobing middle, 0.3 s per letter, a 3 s near-silent ending) was the expensive part. The new film reuses it exactly.

Run it

cd film && npm install
npm run studio           # scrub
npm run render           # -> film/out/thought-to-text.mp4

Rebuild the soundtrack (optional, film/public/audio/mix.wav is committed):

cp .env.example .env     # ELEVENLABS_API_KEY, only needed to regenerate VO
score/vo.sh              # new voiceover takes (optional)
score/render.sh          # SuperCollider -> music.wav, + VO -> film/public/audio/mix.wav

Layout

path	what
reference/	teardown of the original edit (text only, no footage)
tools/teardown.py	the decompiler (video-teardown)
film/src/theme.ts	the beat map. Every scene reads its times from here
film/src/scenes/	one file per act
film/STORYBOARD.md	beat-by-beat script, copy and VO timings
score/score.scd	the score, events placed by absolute time
.claude/skills/remake-an-edit/	agent guide for doing this to any edit

Credit

Reference edit by Jordan Watkins. None of the original footage, frames or audio is in this repo; reference/ is a written analysis.

MIT.