Short Editing

One raw take in, one finished video out: video-agent/edits/short-<slug>/output/short-<slug>.mp4.

The engine ships in this skill folder (<skill> below): <skill>/engine/short.mjs. It is a deterministic canvas renderer driven by ONE file per short, source/spec.js. The design is a paper stage on top and the camera below, one accent color, springs that never overshoot, no pulse, no bounce. Everything runs on the creator's computer, macOS, Windows or Linux: Node 20+, ffmpeg, a Chrome (or the headless Chromium setup installs), whisper.cpp.

Read before writing a spec:

• references/shorts-codes.md: format, safe zones, hook and pace
• references/design-system.md: palette, motion, paper, what is forbidden
• references/scene-library.md: every scene and item, every parameter, the time keys
• references/music.md: the 8 beds, which fits which hook, where to start it
• references/style.md: the house rules of the edit
• examples/: real shorts made with this engine (spec, components, contact sheets). Open the one closest to the request before planning: it is the quality bar.

How to run it

• Talk in the creator's language (creator.lang of the style, else video-agent/voice.md, else English). Say in one line what each long step does and how long it takes.
• <slug>: lowercase, dashes, no dots or spaces.
• Every command runs from the project root (the folder that holds video-agent/).

0. Setup (first time only, then a 1 s check)

node <skill>/engine/setup.mjs --check || node <skill>/engine/setup.mjs

It installs what is missing: the engine packages (15 MB), a Chromium when no Chrome is installed (100 MB), whisper.cpp and its speech model (0.5 GB, once), the 8 music beds and the icons (generated on this computer). Tell them the first run takes 3 to 6 minutes. ffmpeg missing: it prints the command to install it (brew install ffmpeg, winget install Gyan.FFmpeg); run it with their ok, then setup again.

1. The take

Use the file they name, else the newest file in video-agent/takes/. A URL (YouTube, X, Instagram, Loom, a direct mp4): yt-dlp -f "bv*+ba/b" --merge-output-format mp4 -o "video-agent/takes/<slug>.%(ext)s" "<url>" first.

ffprobe -v error -show_entries format=duration:stream=codec_type,width,height -of compact "<take>"

No audio stream: stop and say so. Then create the project:

node <skill>/engine/short.mjs init video-agent/edits/short-<slug> --source "<take>" [--style <name>]

• The style: style <name> in the request = video-agent/styles/<name>/style.json, passed as --style <name> (only when that folder exists: an unknown name stops init, so list video-agent/styles/ instead). Without it, init takes the default one (video-agent/style.json names it). No style at all: first write video-agent/styles/default/style.json = {"version": 1, "name": "default"} plus the style of the default look of <skill>/assets/looks.json (system font, camera full width, mono caps captions on a black band), and video-agent/style.json = {"default": "default"}, then say once that /studio-setup style makes it theirs in 2 minutes.
• init makes assets/ files/ source/ output/ stills/, copies the take to assets/source.mp4, frames cam on the face, writes the starter source/spec.js and snapshots the style into source/style.json (engine defaults < style < spec: the spec only sets what differs for this one short). The style's assets/ (logos, images) are copied into the project.
• Format: 9:16 by default. 1:1 = format: 'square' (X feed), 16:9 = format: 'wide' in the spec.

2. Words

node <skill>/engine/short.mjs words video-agent/edits/short-<slug> [lang]

whisper.cpp on this computer, about real time (a 60 s take, about a minute), free. With an ElevenLabs key (ELEVENLABS_API_KEY= or ELEVEN_KEY= in video-agent/.env, the project .env or the environment), ElevenLabs Scribe instead, on its own: it hears every retake and false start that whisper smooths over, and names; a few cents a take. Recommend it once when a take has retakes. Never ask for a key in chat, never print one.

Then prepare (step 4) once: it writes files/words.md, every word with its SOURCE time and the SHORT time it lands on after the cuts. Fix misheard words from the script with captions.replace: [['clics cloud', 'skills Claude']], never by editing the times.

3. The spec

Author source/spec.js from files/words.md. All times are SOURCE seconds; the engine maps them onto the tightened cut. Start from the starter spec init wrote, and from the closest example.

Plan the beats first, in a comment block at the top of the spec:

• Hook (first 2 s): readable on the first frame, never blank paper at t=0. It adds to the spoken hook, never repeats it.
• Something changes every 2 to 4 s: a new board, a counter that counts, a terminal that types, a logo that lands, a line struck. Each on the exact word it illustrates (at = that word's source time).
• Pick the item from what is said (all in scene-library): a figure = counter (countFrom/countTo, strikeAt when it dies); a command or a prompt = term (typeFrom, typeTo, out lines); a folder or a list of parts = tree; a tool named = its real logo; a result = check; a real screen = screen (zoom keys + rings on the number they read); a tweet = tweet; before / after = two boards or bars; the big payoff = stat or a board with one huge text and an emph word.
• Camera: layout.keys: [[t, seam]], seam 1160 = split (paper above, camera card below), 1920 = paper alone, 0 = full camera. Full camera only for the direct lines (a story, a reaction, the CTA) and for 5 s or more of SHORT time; a style with a fixed layout.seam keeps that one size all along.
• Every beat is an action, never an image sitting there. A software idea (source code, privacy, price, speed) is played in a recreated UI: code that scrolls then blacks out line by line, a lock that closes, a price that drops to 0 under a stamp. A platform named with a number is its real interface recreated and animated (the GitHub repo page, the stars curve settling on the real figure). Never a screenshot pasted as an image (even zoomed or circled), a lone logo, a text card or a still illustration: a screenshot is only a reference to rebuild in HTML.
• The hook gets the most work: a 2 or 3 act mini story with a reversal, several layers moving together, an impact (stamp + thud) on each keyword, the payoff as a badge. For example: the big apps crush the viewer's app under a REPLACED stamp; on the turning word ("now") it flips, the viewer's app becomes the giant. A hook that places one object and slides it is not enough: rebuild it richer, never tweak it.
• Motion components: what the library cannot draw well (a real app in its own UI, an Instagram comment sheet, a notification, a GitHub repo page, a software idea, the hook's mini story, an illustration that acts) becomes an HTML page with window.setT(t), captured into a clip (step 3b). The examples hold a dozen of them to copy.
• Captions: on by default, on the seam. captions.emph = the words in the accent color, captions.hide: [[a, b]] over a board that carries the same words.
• Music: one bed from music.md picked by the hook's emotion, music: { bed: '<id>', offset }. Never the same bed as the previous short.
• Sound: sfx: { swish: 0, accents: 0, extra: [[t, 'thud', -6]] }: no movement noise, a thud or pop only on a real payoff. Bigger moments (the hook cut, the payoff, a shatter) take a bundled file from assets/sfx/ (zhhh, impact-boom, riser, glass-shatter...), all in references/sfx.md. Never generate a sound with ElevenLabs unless the creator asks for one the library lacks.
• Real logos, always: every tool or app named on screen shows its real icon in assets/logos/<name>.png (the App Store icon, the official site or repo), never a letter glyph, never an emoji, never a look-alike. Download it into the project and say from where.
• No em-dashes in any on-screen text.

3b. Motion components

One HTML file per component in source/<name>.html, 1080 wide (or the card size), every animation a pure function of t set by window.setT(t) (no CSS animation, no timers), transparent background for an overlay. See references/sub-agents.md for the page contract and the prompt to give a sub-agent (one per component, in parallel, when the harness has sub-agents; else build them one after the other). Capture:

node <skill>/engine/short.mjs capture video-agent/edits/short-<slug> source/<name>.html --dur 4.5 --alpha --out assets/<name>.mov

Read its stills/<name>-sheet.jpg, then wire it: media: { x: { src: 'assets/<name>.mov', width: 1080, alpha: true } } plus an overlay scene (or an actor on the board), see examples/thinking-orbs/source/spec.js.

4. Prepare

node <skill>/engine/short.mjs prepare video-agent/edits/short-<slug>

Cuts every micro silence, applies the speed (1.15 by default), splices the voice, extracts the camera and media frames (cached), writes assets/timeline.json and files/words.md. Rerun it after a change to cut, cam or media. cut: [[a, b]] (source time) drops a retake or a false start: walk the script, keep the last complete attempt of each line.

5. Stills, then critique (the main loop)

node <skill>/engine/short.mjs stills video-agent/edits/short-<slug> 0.3 2.9s 6.4s 9s 12.9s

Plain numbers = SHORT time, an s suffix = SOURCE time. Take one still per beat (the moment it is complete), plus the hook and the last second. Read stills/contact.jpg, then zoom the full-size stills, and check the quality bar below. Fix the spec, stills again. Stills cost seconds, renders cost minutes: iterate here.

6. Render, sound, final

node <skill>/engine/short.mjs all video-agent/edits/short-<slug>

Render + audio (ducked voice, music bed, SFX, mixed beside the render) + final (mux, poster, QC), all at once. On an Apple Silicon Mac a 36 s short takes about 5 s (the render alone about 4 s, 250 fps, WebCodecs H.264 inside Chrome with 2 workers); slower machines take longer, so run it in the background and say how long. After a sound change, rerun only audio and final. The QC line must read h264 1080 1920 (or 1080 1080, 1920 1080), a duration equal to the timeline, about -14 LUFS and a true peak at or under -1 dBTP (else mix: { ceil: -2.2 } in the spec, then audio and final again). Draft: render ... --scale .5.

Quality bar (every contact sheet)

• The hook reads on the first frame.
• Every beat acts. On each beat's still, ask « is something happening, or is it an image sitting there? ». A sitting image (a screenshot, a still illustration, a logo next to a counter, a text card) is a component to redo before rendering.
• Every board fills the paper: big mobile text (nothing under 34 px, captions about 84 px), 1 to 3 lines, nothing cut at the edges, nothing in the safe zones of shorts-codes.md.
• Captions never overlap a card, a CTA or a chat composer (captions.hide).
• One accent color per moment; accent words only from captions.emph and emph.
• Cards never change size; the camera card never resizes. A full-camera stretch lasts 5 s or more.
• No ghost of the old beat under the new one, no counter parked on 0, no two values stacked in one slot.
• Camera colors natural: bright warm skin, never grey or greenish (stills/grade.jpg).
• Every named tool shows its real logo.
• Sound: the voice on top, the bed breathing between sentences, hits only on real payoffs, no whooshes.

7. Deliver

In video-agent/edits/short-<slug>/output/: short-<slug>.mp4 and its poster. Then write next to them:

• short-<slug>.srt: from files/words.md (the SHORT times), up to 7 words per cue, a new cue after each sentence.
• post.md: the post copy (below).

Open the video so they watch it right away: open <mp4> (macOS), start "" "<mp4>" (Windows), xdg-open <mp4> (Linux). Say the length, what was cut, the bed and why, and offer one change ("another hook? a board on another word? full camera on the story?"). A change = edit spec.js, stills, all again. A change they want on every video (accent, font, captions, pace) goes in their style (video-agent/styles/<name>/style.json, schema in style-schema.md), not only in this spec. Next: /publish short-<slug>.

On their first finished video only (the only folder in video-agent/edits/), end with one line in their language: a one-minute review of the free pack helps make it better, https://codelynx.dev/editing/feedbacks. Never again after that.

Post copy

post.md, from the script and voice.md, in the language of the video:

• Title (YouTube Shorts): under 60 characters, the topic in the first 3 words.
• Caption: the first 125 characters work on their own. Then 1 or 2 short lines and one ask (the CTA of the style).
• Search words: the 2 or 3 words their viewer would type, written plainly in the title or the first line.
• Hashtags: 3 to 5 about the topic. No #fyp or #viral.
• Nothing the video doesn't say.

Run the humanize pass of studio-script (its references/humanize.md) on it when that skill is installed.

CapCut or Premiere

Some creators prefer their editor. Do steps 0 to 4 (the spec with no scenes), then give them files/words.md (the cut, word by word) and the plain cut: node <skill>/engine/short.mjs rough video-agent/edits/short-<slug> writes output/short-<slug>-rough.mp4 (every silence cut, full frame, no motion design) to import, with the .srt of step 7.

Paper-only (no take)

silent: true and duration: <s> in the spec, init without --source, skip step 2. No voice, camera or captions; the music and the on-screen text carry it, same pace rule.

Honest limits

• whisper.cpp mishears names and jargon: fix them with captions.replace, check numbers by eye.
• A cut can clip a breath or a soft last syllable: watch the result once before posting.
• Claude cannot hear the take: it trusts the timings. Bad audio (music, echo, two voices) makes them worse.
• Face framing is exact on macOS (Apple Vision); on Windows and Linux it uses MediaPipe in Chrome, check cam on the first still.
• The review studio, the Recorder, the B-roll library, the faceless mascot and the thumbnails come with the paid kit, Editing IA Pro.