zaddy99999/video-editing

video-editing (router)

You edit video by running small, single-purpose scripts and reading their JSON output, not by watching the video. Look at frames (proof sheets, full-res stills) only at decision points.

Which skill

The user wants...	Skill
a long recording split into clips, cut points suggested, dead lead-in removed	auto-cut
hook / title / "POV:" text on the video	hook-overlay
subtitles, captions, a transcript, cuts that respect words	captions
loudness fixed, silence removed, clean joins, crossfades	audio-polish
a seamless loop, a GIF, a meme GIF, a sticker	loops-gifs-stickers
to know if a render is OK, a contact sheet, an approval step	render-qc
the video on YouTube	youtube-publish

Default pipeline (raw screen recording to published Short)

auto-cut  ->  hook-overlay  ->  captions  ->  audio-polish  ->  render-qc  ->  youtube-publish
suggest,      plate PNG +      transcribe,    trim silence,     verify,         dry-run,
approve,      overlay          ASS, burn      loudnorm          proof sheet,    upload private,
split, trim                    (last visual)  (last audio)      user approves   ledger

Not every job needs every step: a music clip skips captions; a GIF skips publishing. Ask when the goal is unclear; otherwise state the plan and go.

Hard rules (all skills)

1. Never touch the original. Outputs go to edit/ beside the source (scripts default to this).
2. Plan, then cut. Propose cut points / hooks / settings and get the user's OK before destructive or slow steps.
3. Never cut inside a word. For speech, snap cuts to word gaps (captions/snap-to-words).
4. Every audio cut gets a ~30 ms fade. Clicks come from cutting on non-zero samples.
5. Clips open on content. No black frames, loading screens or the previous screen's tail (auto-cut/trim-lead, checked by render-qc).
6. Text stays in the safe zone. Hook plate in the top area, max 2 lines; captions above the bottom platform UI; nothing overlapping.
7. Order matters. Cuts, then overlays, then captions (last visual step), then loudness (last audio step), then QC.
8. QC every final render with render-qc/verify-render, then show the user the proof sheet or the real MP4. Do not judge from downscaled previews.
9. Approval is human. Never pass --approved or --allow-public unless the user said so in this conversation.
10. Secrets stay out of git and out of logs. OAuth clients and tokens live in env vars or gitignored files; never print them.
11. Prefer re-running a script with different flags over hand-writing ffmpeg. Every script answers --help and --help --json.

Setup check

ffmpeg -version && ffprobe -version     # required (any recent build; drawtext NOT required)
node --version                          # >= 18
npm install                             # once, at the repo root (installs sharp)
python3 -m pip install faster-whisper   # only for captions/transcribe.py

If ffmpeg is not on PATH, set FFMPEG_PATH and FFPROBE_PATH.

Style packs

brands/example.json holds canvas size, font, colors, hook-plate geometry and caption style. Copy it per channel/brand and pass --brand brands/<name>.json. Use system fonts or open-license fonts (--font-file); never commit commercial fonts.