Orchestrates one narrated vertical short from idea

sk-pipeline: one video, idea to hand-off

Announce: "Using sk-pipeline for <id>."

This skill is not the rules

The rules live in SOUL.md and docs/soul/. This skill only fixes the order and points at the skill and § for each step. Where it disagrees with a §, the § wins and this file is the one to fix. Do not copy rules into skills: two copies drift and nobody knows which is current.

Read SOUL.md first (index, three first rules, pipeline order, checklist). It holds this channel's editable defaults; changing them is sk-improve's job. Then open only the § for the step you are on; do not load the whole of docs/soul/. Exact commands per step are in docs/making-a-video.md.

The steps

#	Step	Skill	Rules	Evidence on disk before moving on
0	If SOUL.md's change log has no entries, the rules were never adapted: offer first-run setup	sk-improve	SOUL.md	change log has entries, or user declined
1	Idea: one sentence the viewer can repeat after watching	sk-brief	§1	sentence written in the brief
2	Facts: every claim sourced and recorded	sk-facts	§9	bun run facts clean
3	Flow + visual choice, image licences checked	sk-brief	§2, §5, docs/flows/, docs/visuals/	bun run new <id> --flow <X> done
4	Brief	sk-brief	§2, §5	briefs/<id>.md with fact ids
5	Narration	sk-narration (polish: sk-style)	§1, §3, §3b, §3c	src/videos/<id>/narration.json
6	TTS	sk-narration	§4	voice.json + public/audio/<id>/voice.mp3, listened to
7	Lint	sk-narration	docs/soul/thresholds.json	bun run lint <id> 0 errors, warnings read
7b	Storyboard: one row per camera shot, approved by the user before any scene code	sk-scene	§12, docs/storyboard.md	bun run storyboard <id> clean, shots approved
8	Scenes (+ optional SFX, docs/sfx.md)	sk-scene	§5, §6, §7, §8, §12, §13	bun run check clean
9	Still review	sk-scene	§7	stills in out/<id>/ opened and looked at
10	Render	sk-render	§11	out/<id>/<id>.mp4, frames checked
11	Cover	sk-handoff	§10	out/<id>/<id>-cover.png checked at 200 px
12	Normalize	sk-render	§8	out/<id>/<id>-loud.mp4, loudness measured
13	Encode	sk-render	§11	out/<id>/<id>-upload.mp4, ffprobe checked
14	Hand-off	sk-handoff	§10, §11	caption, pinned comment, credits in out/<id>/

A recorded voice instead of TTS (§5j)

When the user brings a recording (a filmed take, several clips, or a voice-only audio file), the voice is what was said, not written. Start with bun run from <file|dir>...: it does steps 3, 5a and 5b below in one go and prints what to ask the speaker. Steps 5 and 6 become ingest and cut; the rest keeps its order:

#	Step	Command	Evidence
3	Flow + visual: talking-head shows the take; any other visual uses only its audio	bun run new <id> --flow <X> --visual talking-head or --visual <type> --voice recorded	narration.json voice.engine recorded
5a	Ingest: normalise, Whisper transcript, proposed cuts	bun run ingest <id> <file> --lang <code> [--prompt "names"]	transcript.json, review read, every ⟪words missing?⟫ heard
5b	Cut: edit cuts.json (reasons), fix misheard words with spelling, never retype a line	bun run cut <id>	footage.json, voice.mp3 listened to end to end
2	Facts, AFTER the transcript: every claim said is verified, cut, or "claim": false	sk-facts	lint has no "cites no fact" error

Ask the user for the exact words wherever Whisper is unsure (? in the review, numbers, names, ⟪words missing?⟫): an agent cannot hear the take, and Whisper writes what it expects. Then regroup lines into the flow's beats, declare face for talking-head, and continue at step 7.

The step list the preview shows lives in docs/pipeline.json. A channel that adds, removes or reorders a step changes it there (with sk-improve), and the preview follows.

After the user has posted and has numbers: sk-improve.

Three orderings that are rules, not habits

1. Facts before words (step 2 before step 5). No sentence is written before its claim has a verified record. Writing first and sourcing later is how a good-sounding sentence survives with no source behind it.
2. Visual language before words (step 3 before step 5). Decide how each claim gets on screen, and check the licence of every image, before scripting. A script built around an image you may not use has to be rewritten.
3. Script is final before TTS, TTS before scenes. Scene length comes from the real voice track, never set by hand. Editing text after TTS means running TTS again, or captions drift silently.

The cross-cutting principle

A step is done when there is evidence on disk, not when the command exits 0. This pipeline fails silently: an mp4 that exits 0 but is truncated, a still that renders but has text under the platform UI, a voice track that loops a phrase. Open the file, look at the frame, listen to the audio.

Hard limits

• Never auto-publish. The pipeline ends at files in out/<id>/. The user uploads by hand. Do not write or run upload scripts.
• Never hand-edit voice.json. It is generated by bun run tts <id> (or bun run cut <id> for a recorded voice).
• A video never imports another video. Shared code goes to src/shared/. bun run check enforces this.

Checklist

Use the pre-hand-off checklist in SOUL.md, not a copy of it anywhere else.

Resuming a video

Look at what exists: briefs/<id>.md, src/videos/<id>/narration.json, voice.json, out/<id>/*. The first missing piece in the table above is the next step. If narration.json is newer than voice.json, TTS is stale.