Every tool in What to build reads and writes plain files in one project folder. Nothing lives only in a tool's memory: if a step is lost, the files say where the film is. This page is the contract; the site's own data (data/*.json) follows it too.
The project folder
latency/
book.json the shot book: the one file every tool reads
song/
take.wav the take a person picked, untouched
stems/ vocals.wav, drums.wav, bass.wav, other.wav (Demucs)
words.json the timed words (from the lead stem)
bars.json the beat and bar map
lines.json the line table
plates/ empty worlds from Midjourney, one file per plate
frames/ first frames with him in them, one per shot
windows/ the audio window of every sung shot
takes/ every video take that came back, never deleted
cut/ edit.json, the renders, the review sheets
ledger.jsonl every paid call, one line each
log.md what happened, in the order it happenedTimed words · song/words.json
Read off the lead-vocal stem, not the mix: the mix smears onsets.
[
{ "w": "Your", "t0": 12.48, "t1": 12.71, "line": 1 },
{ "w": "message", "t0": 12.71, "t1": 13.20, "line": 1 }
]Times are seconds from the first sample of take.wav, to the hundredth.
The bar map · song/bars.json
{ "bpm": 96, "beats_per_bar": 4, "first_downbeat": 0.42,
"bars": [ { "n": 1, "t": 0.42 }, { "n": 2, "t": 2.92 } ],
"sections": [ { "name": "Intro", "from_bar": 1, "to_bar": 4 } ] }A cut is allowed only within 0.03 s of a bar line in this file.
The line table · song/lines.json
One row per sung line, built from the lead stem's level (a line starts where the level rises above the floor and ends where it falls back), then named from the timed words.
[ { "n": 1, "text": "Your message came at midnight", "t0": 12.48, "t1": 14.90,
"section": "Verse 1", "bars": "5–6", "held": false } ]The shot book · book.json
The board, the windows and the jobs in one file. Each shot:
{ "id": "S06", "bars": "17–19", "section": "Chorus 1", "look": "night-coat",
"kind": "sung", // sung | silent | no-face | code
"lines": [7, 8], // rows of lines.json it sings
"plate": "plates/train-window-2.png",
"frame": "frames/S06.png",
"window": { "file": "windows/S06.wav", "seconds": 6, "lead_in": 0.7 },
"prompt": "…", // written from the canon, checked by lint
"takes": ["S06-a", "S06-b"],
"pick": "S06-b" }Rules the lint checks: the identity text appears once and word for word; a sung window is a whole number of seconds and opens on about 0.7 s of silence; every segment of the timeline has one clause of movement; no mood words.
The take registry · takes/takes.json
[ { "id": "S06-b", "shot": "S06", "model": "seedance-2.5", "where": "higgsfield-api",
"inputs": { "image1": "frames/S06.png", "image2": "kit/hal-face-sheet.png", "audio1": "windows/S06.wav" },
"seconds": 6, "usd_list": 2.40, "file": "takes/S06-b.mp4",
"verdict": "keep", "usable": [0.7, 5.6], "note": "lips late on 'way', fixed by sync" } ]Verdicts: keep, maybe, drop. A dropped take stays on disk: the failure is data.
The ledger · ledger.jsonl
One line per paid call, written before the call returns. The spend cap reads the sum.
{"t":"2026-11-20T21:04:11Z","what":"seedance","shot":"S06","usd":2.40,"job":"…"}The edit list · cut/edit.json
[ { "shot": "S06", "take": "S06-b", "in": 0.70, "out": 5.60,
"at": 41.92, "path": "cut/paths/S06-b.json" } ]at is the time in the song where the clip starts; path is the sync alignment of the take to its window.
Review notes · cut/review.md
One heading per version (## v1, ## v2), one bullet per note, each with the shot id and what to do. The film's making-of quotes these notes; the site renders them in the record.
The image record · data/images.json
What the site shows for every generated image: id (<batch>.<name>.<variant>), film, batch, title, model, aspect, refs, match (face similarity to his sheet, 0–1, or null), look, shot, note, and the prompt exactly as it ran.