Pipeline

How it works

Every clip travels the same path, in the same order. The stages are separate on purpose: each one can be re-run on its own, and the sixth is a stop where a person decides whether the clip goes any further.

01 · Ingest

A long-form source video is added to the local library — a podcast episode, an interview, a talk. The file stays on the operator's machine; the library only records where it came from, how long it is, and what it is allowed to be used for. Nothing is fetched from a platform at this stage.

02 · Transcribe

Speech is transcribed locally, with timestamps, so the selection stage has something to read rather than something to watch. The transcript never leaves the machine, and it is kept alongside the source so a re-run doesn't repeat the work.

03 · Select

Candidate moments are proposed from the transcript. A candidate is only useful if it stands on its own, so each one carries three things: a start, an end, and a reason it works without the surrounding hour. Candidates that begin mid-sentence or end on a dangling clause are discarded here rather than fixed later.

04 · Render

The selected passage is cut to vertical format. This is where the clip actually gets made:

  • the speaker is framed for a 9:16 canvas,
  • captions are burned in, timed to the transcript,
  • audio is levelled so the clip doesn't arrive quieter or louder than the feed around it,
  • and the finished file is written to the local render directory.

05 · Check

Automatic checks run against the rendered file before a person ever sees it. They catch the faults that are cheap to detect and expensive to miss: clipped audio, a sentence cut off at either end, framing that crops the speaker, a caption that runs past the safe area, a duration outside the range the target platform rewards. A clip that fails is flagged with the specific fault, not a generic error.

06 · Approve Manual

This is the gate. Every clip is opened, watched and judged by a person. The reviewer either approves it — at which point a publish job can be created — or sends it back with a written reason, which stays attached to the clip so the same mistake is visible next time. There is no path from stage 05 to stage 07 that skips this step, and no setting that turns it off.

07 · Publish

An approved clip is uploaded to one chosen account at the scheduled time, through that platform's official API. One job means one clip on one account: jobs carry an idempotency key, so an upload interrupted halfway resumes instead of producing a second post. What each platform's upload actually looks like is described on the integrations page.

What happens when something fails

A failed stage stops that clip and nothing else. The source stays in the library, the transcript is kept, and the run can be resumed from the stage that failed rather than from the beginning. A failed publish job is retried a bounded number of times and then left in a failed state for a person to look at — it is never silently re-queued onto a different account.