Ported from the owner's skills library (port 1.0.0, upstream version 1.0; the source hash and what it needs are in manifest.json next to this file). Paths written {skill_dir}/… are this skill's own folder, {skills_root}/<name>/… a sibling icloneOS skill, {agent_home}/… the agent's home. Runs on macos only. Needs ffmpeg on PATH, ffprobe on PATH, python3 on PATH, $F5TTS_VENV, $VIDEOEDITING_PIPELINE_DIR. Not ported (the owner's own records — notes, logs, examples, project data — never ship): references/CHANGELOG.md, references/LEARNED.md. In this copy a folder of the owner's machine is written <your project folder> (in code, the PROJECT_DIR environment variable); folders of the owner's machine are environment variables the host sets: VIDEOEDITING_PIPELINE_DIR (the reference video pipeline: its voice engine and its assets); a path into the owner's own folders is written <your project folder>/… (in shell code, ${PROJECT_DIR:?}); the owner's own folder and project names are written sample project (in the case and joiner they had); the brand-vignette assets are written vignette_<width>x<height>_<fps>.mov.

Video Editing — the owner's standard

The golden rule: the voice in the video is HIS voice. Nothing is re-read, nothing is dubbed by a voice actor, nothing is "improved" by a generic model. A hesitation is cut from the real recording; a wrong word is regenerated in his timbre and only that stretch is replaced. If a model cannot say the word the way he would say it, his stays — his voice with a small flaw is worth more than a synthetic stretch with the same flaw.

The second rule: nothing is claimed without measuring. Every correction passes numeric gates, every delivery passes verify3.py. "Looks fine" is not a result. If a gate fails, state the number that failed.

Language. Everything on screen and in code in English (subtitles, robot boxes, file names). Replies to the owner in his language, short. Final report short: what changed, the numbers, what is left.

---

0. Before touching anything: the state of the file

Three questions, every time, in this order:

1. Does it already have the brand vignettes at the start and at the end? → python3 tools/detect_vignettes.py <video.mp4> and references/vignettes.md. If it does: NEVER add them again. Work only on the body; the vignettes are copied from the file itself, intact. If it does not: they can be added naturally (11 s at the head + 11 s at the tail, asset of the right resolution) — or exactly as the owner says. In doubt between "has" and "has not", assume it has and ask in one line.
2. Is the picture already approved? (subtitles and robot panel burned in, the owner has already said "it's approved") → then it is audio-only work: references/picture-pipeline.md § audio-only mode. The picture is kept and is re-timed by the voice time map; text is never rendered again.
3. What is the source? Raw recording (VFR!) or an already delivered file? A VFR source goes through fps= before cutting.

---

1. The pipeline map

Root of the reference pipeline: $VIDEOEDITING_PIPELINE_DIR (voice engine in sample_project/, venv $F5TTS_VENV).

new voice                                       picture
──────────────────────────────────────────      ──────────────────────────────────────────
find_fixes.py      speech↔subtitle differences
prep_edits.py      differences → real tasks
evidence.py        2 whisper models per task
report_tasks.py    prep/<k>_brief.md
   ↓ EDITOR agent (references/agents.md)
edits_<k>.json  →  check_edits.py
   ↓
run_fix.sh <k> out_<k>          engine: cuts/regenerates, gates T1/T2/T3
   ↓                            → out_<k>/<k>_voice_fixed.wav + <k>_timemap.json
run_critic.sh <k> out_<k>       measurements per correction (+ --episode)
   ↓ CRITIC agent               verdict per id
apply_verdicts.py edits_<k>.json out_<k>/critic/verdicts.json <round>
   ↓ repeat until 0 redos
                                           apply_to_video.sh <k> out_<k>
                                           ├─ apply_voicefix.py   "virtual raw" (picture that follows the voice)
                                           ├─ retime_cues.py      subtitles to the new timing
                                           ├─ activity.py         screen activity
                                           ├─ edit3.py plan/cut   speed-ups + cut
                                           ├─ run_r3.sh boxes     robot + keys + mix
                                           └─ run_r3.sh final     subtitles + MP4 + deliver
                                                   ↓
                                           validate_text.py  (BEFORE the render)
                                           verify3.py        (AFTER: delivery gates)

Audio-only mode (picture already approved): skips the whole right-hand side and uses sample_project/v2/fix_audio_only.py <finished.mp4> <out_dir> <key> <out.mp4> — keeps the picture, re-times it by the map, rebuilds only the body's sound and copies the two vignettes from the finished file.

---

2. The non-negotiable laws

Layer	Law	Where
Voice	cut from the real recording or regeneration in his timbre; only the replaced stretch	references/voice-correction.md
Voice	gates T1 (words heard), T2 (no leftovers ≥ 0.25 s), T3 (pitch/level vs the REPLACED stretch)	same
Vignettes	if they are already in the file, they are not added again; they are copied intact	references/vignettes.md
Delivery	11 + body + 11 (±0.06 s), 50 fps, H.264 High yuv420p bt709, AAC 48 kHz stereo	references/delivery-gates.md
Delivery	program −18..−15 LUFS, vignettes −14..−11, peak ≤ −0.9 dBTP	same
Delivery	sync by correlation within a 1× segment, never by an isolated peak	same
Subtitles	≤ 20 chars/s, 1.0–6.5 s, ≥ 0.08 s gap, ≤ 2 lines, ≤ 42 chars/line	references/subtitles-and-boxes.md
Robot	≥ 12 s between boxes, end ≥ 3 s before the cut ends, t ≥ 6 s, 7–14 s, text "Name: …" ≤ 118 ASCII chars	same
Mix	voice −16.5 LUFS · bed −34 LUFS from 0:14 with ducking · keys −30 dBFS · alimiter=0.891	references/audio-mix.md
Process	one writer per file: never two integrations of the same episode at the same time	references/traps.md

---

3. How the work is done (pace)

• Parallel up to the machine's limit, but never two writes to the same file. The voice engine is sequential per episode; different episodes run at the same time. Chrome renders: 4 in parallel.
• Everything resumable. MLX segfaults now and then; every long step saves per item and resumes (run_fix.sh, run_critic.sh, evidence.py already do this).
• Two agents per episode: an EDITOR that decides and an adversarial CRITIC that tries to fail it. The critic only approves with numbers. references/agents.md has both contracts.
• Always confirm the file's date before calling a result good (ls -la): the engine stops without writing the wav when two corrections overlap, and the old file stays there, misleading.
• Requests to the owner: short. Work: fast and measured.

---

4. Self-improvement (mandatory)

This skill must get sharper with every job. At the end of every job, without the owner asking:

1. references/CHANGELOG.md — one entry: date, file/series, what was done, the final numbers (durations, LUFS, how many corrections, how many refused), how long it took.
2. references/LEARNED.md — each new lesson in three lines: Rule / Why (measured) / How to apply. A correction from the owner goes in immediately, with the prefix owner: and his words.
3. If a lesson shows up twice, it is promoted: it moves to the reference file of the respective layer (and the metadata.version of this SKILL.md goes up 0.1). A promoted lesson stops being a "lesson" and becomes a law in the §2 table.
4. If a new gate proves useful, it is added to the gates script, not just to the text.

The test: an identical job, done later, cannot repeat a mistake that is already written here.

---

5. References

• references/voice-correction.md — the engine, the gates, the overrides, what to do when one fails
• references/vignettes.md — the vignette law, how to detect, how to add them if missing
• references/delivery-gates.md — formats, loudness, sync by correlation, proofs
• references/subtitles-and-boxes.md — Apple TV-style subtitles and robot boxes, and the re-alignment after the voice
• references/audio-mix.md — voice, music bed, keys, ducking, limiter
• references/picture-pipeline.md — virtual raw, plan/cut, audio-only mode
• references/agents.md — EDITOR and CRITIC contracts
• references/traps.md — everything already paid for (read before debugging)
• references/LEARNED.md · references/CHANGELOG.md — the skill's living memory
• tools/detect_vignettes.py · tools/probe_delivery.sh — the two start-up checks