Documentary: transcripts, logging, and the paper edit

An interview documentary is a writing problem with pictures attached. Forty hours of people talking has to become ninety minutes of argument, and the assembly that works is almost never discovered on a timeline; it is discovered in the transcripts. The paper edit is old advice and still the right advice. What changed is the paper: it can answer questions now. Here is the pipeline that gets an interview-heavy project from a shelf of drives to a radio cut, and then the case for the part everyone skips, the archive that outlives the film.

Transcribe everything on your own machine

Transcription used to be the tax on documentary: per-minute cloud rates that turn forty hours into real money, or interns and weeks. On an Apple silicon Mac it is neither. ClipLogger transcribes on-device as clips open, free, with audio never leaving the machine, and a larger Whisper model is one download away when accents or crosstalk demand the accuracy. Transcripts feed full-text search directly: "the first time she mentions the flood" is a query that returns clips, not an afternoon with a legal pad and a growing suspicion it was on the other drive. So transcribe all of it, including the material you suspect is nothing, because selective transcription is a bet on your own memory, and the whole point of this system is to stop making that bet. The economics matter here: when transcription is free and local, the marginal hour costs electricity, so transcribing stops being a budget line and becomes a default.

Themes and speakers become fields

A transcript records what was said. It cannot say what the material means to the film, and that gap is what logging is for. Keep the structure small: a multi-select theme field with a closed vocabulary (the flood, the lawsuit, before, after), the speaker linked as a subject, and a rating for how alive the delivery is, because a flat read of a crucial sentence is a fact worth recording. Speakers as subjects earn their keep outside the interviews: ClipLogger groups the same face across the footage, so naming a group once means the archive knows every appearance, sit-down or vérité, and "her, but not talking" becomes a filterable idea. The AI proposes against your theme vocabulary, and it is decent at it, but nothing lands until you confirm, and the vocabulary itself deserves a human author. The theme list is the film's thesis in noun form. The thesis is not delegable.

The paper edit becomes a query

Assembly by query looks like this: theme is the lawsuit, has a transcript, speakers narrowed to the two principals, flagged takes only. Read the survivors as a string-out, in order, and mark exact spans as segments while you listen. Promote the keepers to virtual clips: each pull gets its own name and its own row, nothing copied and nothing transcoded, a first-class clip that is a decision about a range. The radio cut assembles from pulls, leaves as FCPXML with an EDL beside it, and the real cutting starts in the NLE with the argument already standing. What used to be index cards taped to a wall is now a saved query, and unlike the wall, it re-runs itself next month when the pickup interviews land. The wall never told you it was stale. The query cannot help being current.

Read before you cut

The discipline underneath all of this predates every tool on this page and still works with printed pages and a highlighter. Read every transcript once, away from the footage, and mark selects for what is said, not how it looks; the camera will re-argue its case later, loudly. Build the paper edit, the argument in quote order, before opening a timeline, because timelines seduce you into cutting for rhythm before the story exists, and rhythm is very good at hiding a missing spine. Then test the paper the cheap way: read it aloud, or hand it to a producer, who will absorb thirty pages in an evening but will never, ever scrub your timeline. If the argument does not survive being read flat off a page, footage will not save it. The films that assemble fast are the ones somebody actually read first.

The outtakes are the sequel's archive

Ninety minutes ships. Thirty-eight hours goes on a shelf, and the shelf is where documentaries lose their memory: the drive still spins in five years, but nobody alive remembers what is on it, and the director who did has moved on. Logged this way, the shelf remembers itself. Every transcript, theme, speaker link, marker, and note rides in a plain JSON sidecar beside its file, and the content key re-attaches metadata after renames and reorganizations, so even a drive tidied by a well-meaning intern still knows itself. When the sequel gets funded, the new AE mounts the drive, points a Mac at the folder, and inherits the entire understanding without reprocessing an hour. And if there is no app on that machine at all, the sidecars are readable JSON a ten-line script can mine. The film ends. The archive should not, because the archive is where the next film is already hiding.

Write the film from the transcripts, cut it in the NLE, and leave behind an archive that can still answer questions when the sequel calls.