YouTube Edit Kit

One source video in → a clean YouTube cut out: silence + disfluency cuts, burned corrected captions, SRT + chapters, and optional vertical shorts. Free and local — ffmpeg + faster-whisper, no API key.

Every gate below is MANDATORY, in order. Skipping a gate or a self-test is a failure, not a shortcut. Violating the letter of these rules is violating their spirit. Each rule exists because an agent already broke it in production.

Gates (run in order)

digraph gates {
    "G0 Setup check" -> "G1 Transcribe + read" -> "G2 Strategy confirm" -> "G3 Cut + render" -> "G4 Captions" -> "G5 Shorts (optional)" -> "G6 Self-review";
    "G0 Setup check" -> "STOP: give user references/SETUP.md installs, wait" [label="tool missing"];
    "G2 Strategy confirm" -> "WAIT for user OK" [label="no approval"];
}

G0 — Setup. Verify EVERY tool in references/SETUP.md with its verify command. Missing → hand the user that file's install instructions and wait. Never substitute a "similar" tool.

G1 — Transcribe. references/EDIT.md §1: word-level verbatim, cached per source, then READ the entire packed transcript. Cut decisions come from reading.

G2 — Strategy confirm. Present cut approach + estimated length (+ shorts count and hook list if applicable) and WAIT for the user's OK. Rendering before approval is a violation, not initiative.

G3 — Cut. references/EDIT.md §3–5: EDL (silence + disfluency) → per-segment render → lossless concat.

G4 — Captions. references/CAPTIONS.md — read it before writing any cue. Term review first (scan_terms.py + reading; YOU are the AI reviewer for proper nouns/jargon/onomatopoeia — ask the user what you can't verify), then corrections dictionary, output-timeline mapping, burn LAST.

G5 — Shorts. references/SHORTS.md — moment selection, hook titles, first-frame rule, blocking self-tests.

G6 — Self-review. references/EDIT.md §7: look at frames, check duration, spot-check sync. If you wouldn't ship it, don't present it.

Hard rules (production correctness — non-negotiable)

1. Never cut inside a word — snap every edge to a word boundary, then pad it.
2. Never re-transcribe an edited file for captions — map original timestamps through the piece table (CAPTIONS.md §2).
3. Captions burn LAST, after every overlay, with -c:a copy.
4. Per-segment extract → lossless -c copy concat; identical codec params on every clip; 30ms afades at every boundary.
5. Cache transcripts per source; never re-transcribe an unchanged file.
6. No render before the user approves the strategy; no publish without an explicit per-destination "go".
7. Verify by LOOKING at extracted frames, never by exit code alone.
8. Captions never burn without a REVIEWED corrections.json — scan_terms.py + full transcript read; empty [] only after the review actually happened.

Quick reference

Red flags — STOP, you are rationalizing

Upgrade path (optional richer skills, if installed)

• longform-to-content — webinars/lectures at scale: relayout + PIP, cold open, CTR thumbnails, scheduled publishing.
• embedded-captions — designed caption identities with subject occlusion.
• talking-head-recut — kinetic graphic overlay cards (lower-thirds, callouts).