Compute
Prepare podcast and video episodes
Produce a transcript, chapters, clip candidates and draft copy from one episode set. Each output has its own readiness, and none of them is an edited file or a published post.
What you get
Four outputs per episode, and they are not equally finished. Read the readiness column before you plan around them.
| Output | What it is | Readiness |
|---|---|---|
| Transcript | Timestamped, speaker-labeled text | Usable as delivered |
| Chapters | Titled time ranges covering the episode | Usable after a title pass |
| Clip candidates | Start and end times, with a reason and a cited quote | Candidates only; an editor picks |
| Draft copy | Title options, episode description, show notes | Drafts; edit before use |
Nothing here is an edited audio or video file, and nothing is published.
What you need
Three episodes and a manifest. Three shows you a short one, a long one and an interview with crosstalk.
episode_id,path,duration_seconds,episode_title,speaker_count
EP-0112,masters/ep112-solo.wav,1740,Pricing without panic,1
EP-0113,masters/ep113-interview.wav,4020,Field notes from a rebuild,2
EP-0114,masters/ep114-panel.mp4,5310,Three views on hiring,4
Declare episode_id as your item_id_field. Use your finished master, not a rough cut, because chapter and clip times are only useful against the file you will actually release.
Run it
- Inspect locally:
lithi media inspect ./episode-fixture. Confirm durations match your masters exactly. - Read the workflow:
lithi workflow explain <workflow-id>. Its parameters set clip length bounds and how many candidates you want per episode. - Quote it:
lithi quote --workflow <workflow-id> --input ./episode-fixture/manifest.csv. The quote states a maximum charge and an expiry, and starts nothing. - Approve the quote as a separate act. Read the maximum charge first: what a quote does and does not do.
- Submit with
lithi batch submit, then track withlithi status <job-id>. - Download the bundle once the job is terminal.
Read the result
Open result-manifest.json, then read data/part-* output by output rather than episode by episode.
Check the transcript covers the full duration, with no gap at the end. Check chapters tile the episode without overlapping, and that the first chapter starts at zero. For clip candidates, open three at the cited times and confirm the quote in lithi_evidence is really said there. A candidate whose quote does not appear at its timestamp is a defect.
Read the draft copy last, and read it as a draft. It will repeat your own words back, which is the point: what each bundle file holds.
Crosstalk in the panel episode is where confidence should drop. If it does not, the speaker labels are worth checking by hand.
Where it stops
No editing happens. There is no cut file, no burned-in captions, no thumbnail and no render.
No publishing happens either. Nothing reaches a hosting platform, a feed or a social account, and no post is scheduled. The clip candidates are timestamps to hand an editor, and the draft copy is text to rewrite in your own voice.
The chapter titles and copy are generated language, so they carry no promise of accuracy. Every claim they make traces back to a cited moment in your episode, and a person decides what ships.