9k-montage

9k Montage

Overview

You are the editor. A video goes in, and a finished 9k montage comes out: tightened cut, jump-cut punch-ins, word-by-word captions, 9k motion graphics on the right words, SFX, end card. His voice stays as recorded: no clean-up, at the source level. The pipeline is scripted. Your job is the judgment: what to cut, what the words really are, which graphic goes on which word, and an honest critique of your own frames.

One gate only. You ask the user nothing before the plan. Read the footage, decide, write docs/plan.md, and wait for "OK". After the OK, render, self-review, fix, and deliver without asking again.

Styling comes from 9k-motion (references/brand.md there: colors, type, springs). This skill does not run 9k-motion's questionnaire: a montage has its brief inside the footage.

Workflow

1. Scaffold: if <name>-9k/ already sits next to the video, this is a re-edit: work in that folder (its README has the rebuild commands) and skip to the step the request touches. Otherwise bash <skill>/scripts/new-montage.sh "<video>" [project-dir] (<skill> is this skill's base directory) creates the project (default ~/films/montage-<folder-slug>/). Then cd into it. It links sounds and images found next to the video into assets/. Never write into the source folder until delivery.
2. Ingest: node scripts/ingest.mjs (add --lang ar for Arabic). Read work/transcript.md and work/ingest.json, and look at work/contact.png. Follow references/craft.md §1: face box, burned-in text, inserts already in the source, framing.
3. Cut: write edit.json → cut (drops for retakes and flubs on raw takes; just tightening on rough cuts) and tail (3.2 for an end card, decided now because cut.mjs applies it), then run node scripts/cut.mjs. From here on, every time is OUTPUT time from work/transcript.cut.md.
4. Transcript: run node lib/captions.mjs, fix every misheard word in captions.fixes, and set keywords. Repeat until every chunk reads right (craft §3).
5. Design the edit: fill in camera, graphics (with the endcard at the start of the tail) and audio. Schema: references/edit-spec.md. Choices: craft §4–6. Work sentence by sentence: one graphic per idea, held for the sentence and built on his words. Then run bash scripts/review.sh for overlay-over-footage stills every 2 s, read the coverage line it prints, look at the stills, and fix every bare stretch and anything obvious.
6. Plan → GATE: fill in docs/plan.md with source facts, cut summary, transcript fixes, and the beat table (time · words · caption accent · graphic · camera · sound). Show it, together with out/contact.png. Wait for OK. Apply any notes, then continue without asking again.
7. Render: node scripts/render-overlay.mjs, then node scripts/compose.mjs.
8. Critique loop: run bash scripts/review.sh out/final-1080x1920.mp4 <climax-t>. Open contact, phone and strip, score them with the rubric in craft §7, fix the 3 worst problems, and re-render. Do at least 2 rounds. Log each round in docs/review_log.md.
9. Deliver: copy the final next to the source as <name>-9k.mp4, then run bash scripts/archive.sh: it saves the editable project next to the source as <name>-9k/ (edit.json, transcript, plan, review log, scripts, real copies of the assets, the final). Always, on every delivery and every later re-render, so he can change the film months later. Never overwrite the source. Report both paths, duration (before → after), LUFS, what you fixed in critique, and one line on what you'd improve next.

Quick reference

Need	Where
Every edit.json key, all 18 graphic types	references/edit-spec.md
Cut rules, whisper fixes, graphic choice and placement, sound, review rubric	references/craft.md
9k colors, type, springs, Nino, wordmark	<skill>/../9k-motion/references/brand.md
Live preview of overlay over footage	open overlay.html in Chrome (scrub the video)
Quick preview of one range	node scripts/compose.mjs --from 15 --to 25 --half
Labelled frames from any video	node lib/sheet.mjs <video> out.png --every 1

Timing: on Apple Silicon, ingest takes ~20 s and cut ~60 s per 100 s of 4K. Overlay takes ~6 s, and compose ~50 s.

Red flags: stop

Thought	Reality
"Let me run the 9k-motion questionnaire / ask about music, captions, CTA first"	No questions before the plan. Defaults and the evidence in the footage decide, and the plan gate is where he corrects you
"A 4 s gap in the words means a pause I can cut"	Whisper skips speech. Cuts come from measured silence (cut.mjs does this). Read the ⚠ lines
"The transcript is close enough"	Every misheard tech name ends up on screen in orange. Fix every chunk
"I'll put captions where they always go"	Burned-in text and source photos are already on screen there. Hide those ranges, or cover them
"Graphics at the top look clean"	His face is at the top. Use the band below the chin
"More graphics = more motion"	One graphic plus the caption at a time, one band, one title and one panel per film
"A 2 s pop on the keyword is enough"	The graphic holds for the sentence and builds on his words. Bare face for 5 s reads as unfinished: read the coverage line
"Same box, same corner, it's consistent"	It's flat. Alternate sides and change layout: column, lower third, title, split panel, band
"I'll build captions by hand in a canvas scene"	captions.mjs + overlay.js already do RTL/LTR runs, proclitic splits and keyword pills
"The mp4 is delivered, the project in ~/films is enough"	He looks next to the video, not in ~/films. No <name>-9k/ folder there means the delivery isn't finished
"I rendered it, so it's done"	It isn't done until you've looked at the review sheets for 2 rounds
"Sample the logo's colours / use Inter"	9k tokens only. Logos are monochrome by default (iconColor: 'brand' only if he asks)

Common mistakes

Mistake	Fix
Planning graphics in source time, then cutting	Cut first. Plan from transcript.cut.md
A graphic lands before or after its word	Use the word's s from work/words.cut.json
Latin label overflowing a panel in an RTL film	Components fit labels automatically. Shorten the label if it still overflows
The burned-in hook line peeks out under the new hook	hook.cover = the exact rect, padded by 1–2 %
Synth music under his voice	No music unless a track is supplied, and then ducked
Cleaning or normalizing his voice	The voice is always the original: no denoise, high-pass or compression, at the source level. voice.clean and a numeric lufs only when he asks
16:9 graphics centred on his chest	Give them a side column, side: "start" / "end", alternating (craft §4)
A card open and empty, waiting for its line	Open it ≤ 1 s before at, or give the wait to another graphic
Editing edit.json and re-running only compose	Check the re-run table in edit-spec.md (captions/graphics need render-overlay.mjs)