| name | videogen |
| description | Orchestrate the full call pipeline end to end โ from a transcript to a published call page (cleaned transcript, chapters, AI summary) and a rendered video. Inspects the call markdown to see what's already done, proposes the remaining steps, and composes the per-stage skills, confirming before the render. Use for the whole flow; for a single part, run that step's skill directly. |
Videogen Skill (orchestrator)
Composes the call pipeline. It does not itself clean transcripts or render โ it
identifies the call, inspects what's already done, proposes a pipeline, and runs
the per-stage skills with confirmation at each gate. The real work lives in:
diarize โ (weird case, on-demand) fix speaker labels when one Zoom
participant is actually two+ people. Most calls skip this.call-markdown โ clean transcript + chapters + AI summary โ calls/<id>.md
(the "update the markdown" half).render-call โ render the MP4 + YouTube handoff block (the "make the
video" half).
You can always run those directly when you only want one part. Use videogen
when you want the whole flow driven and gated for you.
Arguments
$ARGUMENTS may be a bare call number, file paths, or both. Resolve the call id
first (from args, filenames, or by asking). <id> = the calls/<id>.md basename,
e.g. 20260612_054; <NN> = the call number.
Step 0 โ Identify the call (confirm)
- Find the call's markdown in
calls/ (match the number: frontmatter). If none
exists, ask the user to create the agenda first (or offer /draft-agenda).
- Published-call guard. If the markdown already contains BOTH a
# AI Summary
AND a ```webvtt transcript block, the call is already published. Warn that
re-running stages would overwrite published content; only continue on an
explicit stage the user asks for, never a blind full re-run.
- Read the most recent completed call (one with a
# AI Summary + webvtt
transcript) โ the style template for the summary/structure. Replicate it.
- Confirm the call identity (number, title, date) with the user.
Step 1 โ Inspect state and propose the pipeline
Read calls/<id>.md and detect which stages are done, by content marker:
| Stage | Skill | Done when the markdown hasโฆ |
|---|
| Transcript | call-markdown (A) | a ```webvtt block under ## Full Transcript |
| Chapters | call-markdown (B) | NOTE chapters inside that block |
| Summary | call-markdown (C) | a # AI Summary section |
| Video | render-call | a rendered <YYMMDD>-etccc-<NN>.mp4 exists (gitignored) |
Also check inputs: is there a transcript at transcripts/<id>.vtt (or a
transcripts/<id>.diarized.vtt)? If neither exists, prompt the user to drop the
original transcript at transcripts/<id>.vtt (see call-markdown for why
that location).
Does it need diarization? Only if the user says so, or transcripts/<id>.vtt
shows one Zoom label covering two obviously different people (a shared
mic/account). If so, recommend running the diarize skill first (it writes
transcripts/<id>.diarized.vtt, which call-markdown then prefers). Do not
auto-run it; most calls don't need it.
Then present the proposed pipeline and confirm before doing anything, e.g.:
"Transcript โ, chapters โ, summary โ, video โ. Plan: finish the summary via
call-markdown, then โ with your OK โ render via render-call. Sound good?"
If the user only wants the markdown (no video), stop after call-markdown.
Step 2 โ Run the stages (gated)
In order, skipping what's already done:
- (if needed)
diarize โ separate, on-demand; produces the diarized
transcript. Hand its output to the next stage.
call-markdown โ runs the missing text stage(s) (clean / chapters /
summary / write), each with its own confirm gate. This produces the complete
calls/<id>.md.- Confirm before rendering. This is the key gate: only proceed to video once
the user approves. Many runs legitimately stop here.
render-call โ draft โ final MP4 + the paste-ready YouTube handoff.
Invoke each sub-skill rather than re-implementing it; this file is only the
conductor.
Important notes
- Confirm at every gate; never run the whole pipeline unattended, and always
pause before the render.
- Single source of truth is the markdown โ never write the legacy
public/videogen/<NN>/{job,meta,transcript} files.
- Never commit or push media.
calls/<id>.md, transcripts/<id>*.vtt,
speakers/ (yaml + avatars), and the committable <NN>-audio.mp3 are
committable โ but only commit when the user asks.
