Vertical Video Editing Skill

Turn a script + talking-head footage into a polished, creator-grade vertical (9:16) short — automatically. A Claude Code / Claude Agent skill that edits short-form video the way a senior editor would: engineered hooks, A‑roll/B‑roll cutting, eased camera moves, an editorial design system, layered SFX, and a render verification gate — rendered with HyperFrames.

npx vertical-video-editing-skill

Quick start · Features · How it works · Styles · Requirements · FAQ

What is this?

This is a packaged AI skill for editing short-form vertical video (TikTok, Reels, YouTube Shorts — 1080×1920, 9:16). You give your AI agent an approved script, your voiceover/talking-head media, and some gathered B‑roll/assets. The skill encodes the editing standard — the taste, the rules, and the workflow — so the agent produces a finished MP4 that looks like a real creator made it, not like generic AI slideshow output.

It does not reinvent rendering. It rides on top of HyperFrames (an HTML + GSAP video composition engine) and adds the layer that actually makes shorts good: the hook grammar, the cutting rhythm, the camera-motion physics, the design system, the sound design, and a hard verification gate.

Who it's for: creators, agencies, and developers building automated video pipelines who want creator-grade output instead of the uniform-1s-cuts, default-font, stock-template look that most AI video tools produce.

✨ Features

• 🎬 Engineered hooks (first 1–3s) — pattern interrupts and split-panels that grow into the full 9:16 frame. Never opens on a static talking head.
• ✂️ A‑roll / B‑roll cutting — your talking head is the spine; B‑roll covers and punctuates it. Non-metronomic cuts down to 0.2s, paced to the script's energy — never robotic 1‑second cuts.
• 🎥 Eased camera-motion system — zoom pushes and pans with an explicit anti-jitter contract (GPU-promoted layers, force3D, animate-the-wrapper) so motion is smooth in the render, not just the preview.
• 🗞️ Editorial "designed page" look — ten reusable layout patterns: mixed serif/sans headlines, paper & dark-grid grounds, stacked splits with seam captions, highlight-bar supers, name pills, headline cards, UI-card stacks, and more. Frames are designed pages, not video with text on top.
• 🧑‍🦱 Face-safe framing — automatically biases crops so faces stay centered (eyes in the upper third) instead of getting chopped by panel seams.
• 🔢 Themed motion graphics — number/stat counters, spec/JSON cards, callouts, poster splits, chart builds — all designed on your style theme, not a generic template.
• 🔊 Rights-free SFX + clean audio — synthesizes whooshes/clicks/risers/impacts with ffmpeg and mixes everything into one pre-mixed master track (no multi-<audio> echo/desync).
• 🎨 Pluggable visual styles — a FRAME.md design-token system. Ship your brand's colors, fonts, and components once; the editor themes every video from it. Add your own styles in seconds.
• ✅ Render verification gate — a bundled verify-render.mjs script asserts 9:16 dimensions, duration, h264/yuv420p, and audio presence with ffprobe. "Looks fine in preview" becomes a hard, scriptable pass/fail.
• 📋 Auto handoff report — every edit produces an edit-report.md: beat-by-beat decisions, assets used, SFX, and verification evidence.
• 📦 One-command install — npx vertical-video-editing-skill drops it into your Claude skills folder.

🚀 Quick start

1. Install the skill

# Global / user-level (installs to ~/.claude/skills/video-editing)
npx vertical-video-editing-skill

# …or into the current project only (./.claude/skills/video-editing)
npx vertical-video-editing-skill --project

# …or a custom skills directory
npx vertical-video-editing-skill --dir /path/to/skills

Re-run with --force to overwrite an existing install. Restart / reload Claude Code afterward so it picks up the skill.

2. Make sure the engines are present

npx hyperframes --version   # the rendering engine
ffmpeg -version             # for SFX, audio mixing, and verification

3. Hand the agent your inputs and ask for a short

Point it at a folder like this:

your-project/
├── script.md          # approved script with beats/timestamps + on-screen supers
├── voiceover/         # your VO / talking-head clips (the A-roll)
├── assets/
│   ├── manifest.json  # gathered B-roll + a per-beat asset map
│   └── ...            # the media
└── edit/              # the agent writes the finished video here

Then prompt: "Edit this into a 9:16 short using the vibe-life style." The agent runs the non-negotiable init → preview → lint → render → verify workflow and hands back a verified MP4 plus an edit-report.md.

🛠 How it works

The skill enforces one workflow (it never hand-authors HTML from scratch — that breaks live-reload and structure):

1. npx hyperframes init edit — scaffold the project.
2. Build the edit beat-by-beat in index.html (+ sub-compositions) applying the editing grammar.
3. Theme from a style — turn styles/<style>/FRAME.md into a concrete design spec (palette tokens, local @font-face, motion personality).
4. Author SFX + one master audio track with ffmpeg.
5. npx hyperframes preview — screenshot key beats, confirm pacing/layout/motion.
6. npx hyperframes lint — fix every error.
7. npx hyperframes render — final MP4.
8. Verify — node scripts/verify-render.mjs <file>.mp4 + ffmpeg contact-sheet frame checks.

The full standard lives in skills/video-editing/SKILL.md, including inlined references for the editorial look, the camera-move anti-jitter contract, SFX synthesis, style theming, and the report template.

🎨 Styles

A style is a design-token pack (FRAME.md) that decides how every video looks — colors, type, spacing, components, and motion — without touching the editing logic. Different style → different look, same craft.

This repo ships:

Add your own style

cd ~/.claude/skills/video-editing/styles   # (or this repo's skills/video-editing/styles)
cp -r _template my-brand
# edit my-brand/FRAME.md — colors, fonts, components, motion personality

Then tell the agent to "theme from styles/my-brand/FRAME.md." See styles/README.md for the full guide on what makes a good FRAME.md (grounds vs. rationed accents, reading vs. display type ramps, local fonts, component tokens, motion personality).

📦 Requirements

📁 Repo structure

vertical-video-editing-skill/
├── README.md                       ← you are here
├── package.json                    ← npx installer + metadata
├── bin/install.mjs                 ← copies the skill into your Claude skills dir
└── skills/
    └── video-editing/
        ├── SKILL.md                ← the editing standard (the skill itself)
        ├── scripts/
        │   └── verify-render.mjs   ← the render verification gate
        └── styles/
            ├── README.md           ← how to add your own style
            ├── _template/FRAME.md  ← copy this to start a style
            └── vibe-life/          ← worked-example style + HTML showcase

❓ FAQ

Does this generate the video by itself? No — it's the editing brain. It drives HyperFrames (rendering) and ffmpeg (audio), and it needs you to provide a script and media. It supplies the taste, the rules, and the verification.

Is it only for 9:16? The pacing and hook grammar are written for vertical short-form first (the default and the strongest use case). The style packs also define 16:9 and 1:1 behavior, so the same look can be retargeted.

Can I use it without Claude Code? The skill is a structured SKILL.md standard. It's designed for Claude Code / Claude Agents, but the reference material (editorial patterns, camera-move physics, SFX recipes, the verification script) is useful to any editor or pipeline.

How do I update it? Re-run npx vertical-video-editing-skill --force.

How do I uninstall? Delete the installed folder, e.g. rm -rf ~/.claude/skills/video-editing.

🙏 Credits

Built on HyperFrames. Adopts conventions from hyperframes-helper (GSAP power-eases, glow captions, the lint gotcha list) and extends them with a 9:16 short-form grammar, an eased camera-motion system, a style-theming pipeline, rights-free SFX synthesis, and a render verification gate. See the provenance section in SKILL.md for the full adopted-vs-added breakdown.

📄 License

MIT © adriiita