fb-ad-video-studio

High-converting Facebook / Instagram / TikTok video ads, built as code in Claude Code — no editor, no After Effects, no $3k/mo motion designer.

Your best static ad still loses to video. You know it. You're still not shipping video.

Not because you can't. Because the pipeline is hell.

A single 60-second founder ad is a timeline editor, a VO booth, a silence-cutting pass, caption sync, sound design, render queue, and four rounds of "move that text up 40px." Miss the sync by 200ms and it feels cheap. Outsource it and it's $1,500 and a week per variant — so you ship one, it fatigues in eleven days, and you're back to static because video "takes too long."

Meanwhile the accounts crushing it aren't making better videos. They're making more of the same proven shape — same hook arc, same pacing, same caption rhythm — with a fresh script and a new face.

That shape is the asset. This skill is that shape, as code.

fb-ad-video-studio turns Claude Code into your motion department. It's an opinionated layer on top of HyperFrames (HeyGen's HTML-video framework): you bring a script or a recording, Claude authors the composition, and it renders to a finished MP4 — using a structure distilled from 23 production iterations of a real founder ad, not a blank timeline.

Because every ad is a composition file, you get what a timeline editor structurally can't:

• The hook arc is baked in. Cold-open motion hook → speaker PIP arc (hidden → PIP → full → PIP → full → lockup) → whip-pan scene cuts. The rhythm that tested best across 23 versions is the default, not something you rediscover at 1am.
• Sync is automatic, and tight. VO is transcribed to word-level timestamps; captions, SFX hits, and text reveals anchor to actual spoken words — never script estimates. 0.18s drift cuts, 0.25s reveals. It feels expensive because the timing is frame-accurate.
• The audio pipeline is solved. ElevenLabs Chris VO with dialed-in settings, silence-cutting that trims dead air without clipping words (77s→67s typical), and a Mixkit SFX rail with tight ad-length trims. Every "free audio" dead end is already mapped so you skip it.
• It version-controls and templatizes. git diff a headline change. And the killer move — share a winning video ad and Claude reverse-engineers its arc, pacing, and caption rhythm into a reusable template. The winner stops being a competitor's video and becomes your repeatable format.
• It scales sideways. One composition → a dozen on-brand variants for testing. New script, new VO, new offer — same proven shape, rendered, not rebuilt.

Two families ship ready:

It's MIT-licensed and free. The render runs locally. The next video ad you would have paid $1,500 and waited a week for is one npx hyperframes render away — and the one after that is a script swap.

Install it, bring a script, and ask Claude for the ad. ⬇️

Install

As a Claude Code skill (recommended)

git clone https://github.com/ai-agents-for-agencies-coaches/fb-ad-video-studio.git \
  ~/.claude/skills/fb-ad-video-studio

cd ~/.claude/skills/fb-ad-video-studio
npm install                    # pulls the HyperFrames CLI
npx hyperframes doctor         # checks Node >=22, FFmpeg, Docker

Claude auto-discovers the skill via SKILL.md. Then just ask: "make me a 20-second motion-graphics ad for…" or "build a founder ad from this recording" or "make a video ad like this one: ".

This skill is the ad recipe on top of HyperFrames — it doesn't reimplement the framework. If the base HyperFrames skills aren't present: npx skills add heygen-com/hyperframes.

Requirements

• Node.js ≥ 22, FFmpeg, and Docker (render runs in Docker on Linux — apparmor blocks puppeteer otherwise)
• ELEVENLABS_API_KEY (only if generating VO via scripts/tts.py)
• pip install faster-whisper (for caption/SFX sync timestamps)

Production pipeline

The order that survived 23 iterations. Scripts live in scripts/.

script.md
  → scripts/tts.py            VO (ElevenLabs Chris)         [or record a presenter]
  → scripts/cut-silences.py   trim dead air                 [talking-head only]
  → scripts/whisper-words.py  word timestamps (the spine)
  → scripts/reencode-footage.sh  fix sparse keyframes        [talking-head only]
  → edit the template         anchor captions/SFX to words
  → scripts/fetch-sfx.sh      Mixkit SFX rail, tight-trimmed
  → npx hyperframes lint && inspect
  → scripts/render.sh         MP4 (Docker)

Templates

Real HyperFrames compositions in templates/ — generic brand tokens (--accent, --ink, …) and [BRACKET] placeholder copy. Scaffolds, not constraints.

• motion-graphics-spot/ — kinetic-typography spot. Beat arc: hook · problem · solution · proof · CTA. VO + SFX, no footage.
• talking-head-founder-ad/ — the proven speaker PIP arc, whisper-synced captions, SFX rail, whip-pan cuts, closing lockup.

Reverse-template workflow

The highest-leverage use: turn a winning video ad into a reusable composition. See references/reverse-template.md. You extract structure and pacing (not ownable) — never a competitor's footage, VO, music, or copy.

References

• references/patterns.md — Lottie/GSAP/caption/pacing/render gotchas from 23 builds
• references/audio-sources.md — what works for SFX, music, VO, silence-cutting
• references/reverse-template.md — extracting a template from a winning ad

License

MIT — see LICENSE.