SuperAPI AI Directory

youtube-edit-kit

Design & creation

View repository

Updated 10-16 05:04

About this skill

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

Task Where
Tool installs + verify commands references/SETUP.md
Full pipeline: EDL, fillers, render, burn references/EDIT.md
Caption correctness: mapping math, corrections, styles references/CAPTIONS.md
Shorts: selection, layout, hooks, first-frame rule references/SHORTS.md
Runnable CLIs (transcribe → EDL → fillers → cut → terms → SRT → shorts) scripts/

Red flags — STOP, you are rationalizing

Thought Reality
"I'll re-transcribe the edit for captions" ~1s drift per cut, compounds. Map through the piece table.
"Burn captions first, overlay after" Overlays hide captions. Silent failure. Captions LAST.
"One-pass filtergraph is simpler" Double-encodes when anything is added. Per-segment + concat.
"The cut sounds fine without fades" Pops at boundaries on headphones. 30ms afades, every edge.
"User is offline, I'll just render" Strategy confirm is blocking. Wait.
"Exit code 0, ship it" Look at the frames. Every self-test is visual.
"The transcript looks clean, skip the term review" ASR errors are systematic, not obvious. Run scan_terms.py and READ.

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).