A folder of clips becomes a trailer, a longer episode, or a film of a whole trip. The tools do the boring part (shots, quality, beats, speech); the agent picks the story; the person judges. Every rule below was paid for by a cut someone called boring or broken. If the project has its own editing skill, that one holds the commands; this one holds the rules.
Workflow
- Analyse once, store beside the sources. Shots, frame quality, beats, speech with word
timings. Write the results as sidecar files next to each clip (
clip.mp4→clip.analysis.json), so a month later a long film needs no re-analysis. Take shooting time from container metadata (ffprobe -show_entries format_tags=creation_time), never from file names: camera clocks drift. Tools: any shot detector + whisper-family STT;npm i life2film-engine(WASM) covers beats, shots and frame scores on-device. - The day as one text file. All speech in shooting order, numbered by clip and sentence. Read it like a scriptwriter and propose the story before touching a timeline.
- Write the cut as a script, not a timeline. A small JSON: which sentences of which clip are said, which shot covers which line, which silent walks keep their own sound, what the cold open is. Render from it. A complaint ("X is mentioned, never shown") becomes a one-line edit.
- Voice as one continuous track, pictures on top. Cut the audio from the same pieces as the picture, quantised to frames; check that film length equals voice length (drift > 1 frame = bug).
- Look before you render, every time (see Eyes below), then render, then check with
ffprobethat duration, codec and audio length match the plan. - Masters per platform from one archival master (
references/delivery.md). Publish private, watch on a phone, then public.
Editorial rules
- The cutaway shows what is being said, from the same place and time (within minutes of the line), never from a scene the story has not reached or has left.
- Never cover a shot where lips move. It reads as broken sync. Cutaways go into silent windows.
- A mentioned object must be seen, ideally just before the line about it.
- Cold open, then hook. 8–10 s of the best frames, then one line with a twist. The hook is hand-picked from a contact sheet: automatic ranking picks sharp and colourful (an empty path through trees), which is what people scroll past. Fire, water, neon, night, motion toward the camera, a face with an expression, the film's object.
- Scenes, not best phrases. A conversation goes in whole: lead-in, the talk, the reaction. A "best line" without its lead-in confuses anyone who did not see what was cut.
- Find promise → payoff pairs in the transcripts ("tomorrow we'll buy…" → the purchase) and keep both halves; they are the thread of a multi-day film.
- Trim filler words at the ends of a line; mid-line only when the join lands in silence.
- Take something from every clip. The person filmed it, so it mattered. After the cut, list the clips that did not make it, each with a reason.
- Live music in frame is one continuous span with picture cuts on top. Cut per shot and the beat restarts at every cut.
- Length follows the material. Cut stammers and repeats, not thoughts. Don't stretch to a target.
Eyes — numbers miss what a glance catches
- Read the whole film as text in order before rendering: does each line make sense to someone who did not see the cut material?
- A sheet of every cutaway with its line under it. Measured once: 12 of 54 automatic cutaways were off topic (a frame upside down, a car over a line about a garage).
- A storyboard grid, one frame per shot. Catches "same face ten times" that every metric passed.
- A critic subagent, two passes minimum. Pass one finds repeats and sagging scenes, pass two the lines cut mid-word at joins.
Sound
Score written to the cut, a hole in it (not ducking) wherever something plays live, live music as
one continuous span, other people's music stripped with Demucs before upload, chapters levelled to
−16 LUFS. Rules, numbers and commands: references/sound.md.
Delivery
One archival master, a file per door: HEVC for upload (36 min = 1.6 GB vs 3.2 GB x264), YouTube in
1440p even from 1080p, h264 for messengers, native vertical untouched or a face-following 9:16
window with burned subtitles for feeds. ffmpeg recipes, the outro card without re-encoding, cover and
archive rules: references/delivery.md.
Gotchas
- Analysis written into a symlinked subset folder ends up beside the links, not the originals — resolve symlinks before writing sidecars.
- Never leave a rendered film inside the source folder: the next run treats it as one more clip.
- STT turns unknown names into celebrities. Keep a per-trip glossary and fix transcripts once.
- A long trip film is cut from the finished day films (colour, sync, subtitles already fixed), one file per chapter cached by its inputs, so a critic note re-renders one chapter, not an hour.
- Put the real shooting date in every description; a week in, nobody can tell the days apart.
- Archive sources to cloud storage after the films are out, never mid-edit, with the sidecars.
- Verify on at least one real captured file, not only synthetic ones. A pipeline green on
lavfitest sources can still break on camera files: a 48 kHz silent card joined to a 96 kHz recording passed every synthetic test and drifted the audio clock 1.39 s at the seam. Keep one short real clip per camera/phone in the verification set, and check the delivered file itself.
Don't
- Don't trust an automatic hook, cover frame or cutaway list without looking at it.
- Don't re-encode video to change audio, trim, mute or add an outro: copy the stream (
references/delivery.md, "Copy first"). - Don't crop horizontal footage blindly to 9:16.
- Don't chop live music per shot, or duck the score under it.
- Don't render before reading the cut as text.