foundation-models/social-media-campaigns

video-factory

A reusable pipeline for producing short, vertical (1080×1350, 4:5) explainer videos that pair narrated avatar clips with self-contained HTML/CSS kinetic-typography animations. Built for a daily publishing cadence to LinkedIn / YouTube.

The HTML is the source of truth — it is meant to be hand-edited. Everything else (per-section MP4s, the concatenated final.mp4) is regenerated from it.

Layout

video-factory/
├── engine/                     # shared, episode-agnostic tooling
│   ├── build.py                #   design system + kinetic keyframe helpers (library)
│   ├── build_master.py         #   assemble sections/ into one master player
│   ├── render.py               #   HTML section  ->  MP4 (animation + avatar + audio)
│   └── requirements.txt        #   pip deps for the renderer
└── episodes/
    └── 2026-07-03-prompt-injection/
        ├── episode.json        # titles + start/end card copy
        ├── build_episode.py    # (optional) regenerates sections/ from Python
        ├── sections/           # section-0..N.html  (hand-editable)
        ├── <slug>.mp4          # narration avatar clips (Git LFS)
        ├── prompt_injection_full.html   # master player (generated)
        └── final.mp4           # uploadable render (Git LFS, generated)

Requirements

• google-chrome (headless, for deterministic frame capture)
• ffmpeg / ffprobe
• yt-dlp (for reference deconstruction — make deconstruct-reference)
• Python 3 + uv (uv sync for deps; Whisper via uvx for transcription)
• git-lfs (for committing / pushing the .mp4 assets)

Daily workflow — a new episode

1. Create the episode folder and drop in the narration clips:

   cp -r episodes/_template episodes/YYYY-MM-DD-topic   # or copy an existing one
   # place each narration clip as <slug>.mp4 at the episode root
   

2. Edit the section HTML in sections/ (or regenerate via build_episode.py). Each section holds exactly one <video src="../<slug>.mp4">; the player advances when that clip's audio ends, so animation timing is duration-driven.
3. Build the master player (inlines every section, probes clip durations):

   python3 engine/build_master.py --project episodes/YYYY-MM-DD-topic
   

   Open the resulting HTML in Chrome to preview with sound.
4. Render the final video:

   python3 engine/render.py --project episodes/YYYY-MM-DD-topic            # all sections + concat
   python3 engine/render.py --project episodes/YYYY-MM-DD-topic --section 2 # one section
   python3 engine/render.py --project episodes/YYYY-MM-DD-topic --concat    # re-stitch only
   

   Output: episodes/YYYY-MM-DD-topic/final.mp4 (1080×1350, H.264 + AAC).
5. Commit and push (see below). Upload final.mp4 to LinkedIn / YouTube.

Reference video deconstruction

To capture pacing and visual style from a YouTube video you like (without passing the .mp4 to Claude):

make deconstruct-reference URL='https://www.youtube.com/watch?v=...'

Creates references/<video-id>/ with download, transcript, frames, and a scaffolded style-brief.md. An agent completes the brief using the frames, then uses it when drafting plan.json. See .claude/skills/video-reference-deconstruct/SKILL.md.

Manual HeyGen (no API)

If you generate avatar clips on the HeyGen website instead of make narrate, follow docs/MANUAL_HEYGEN_WORKFLOW.md — plan → manual clips → build_episode.py → make render.

Timing model

Section animations use percentage-based @keyframes whose total cycle equals the clip duration, so they stay in sync regardless of the clip's length. To retime, you only change the duration token — the percentages are duration-independent.

Version control & hosting (Hugging Face)

Media is tracked with Git LFS (see .gitattributes). Regenerable render caches (build/, .frames_*, headless-Chrome profiles) are git-ignored.

One-time, on your machine:

git lfs install

Publish to a public Hugging Face model/dataset repo (git-based, generous LFS):

# 1. create the repo (needs the huggingface_hub CLI + your token)
pip install -U huggingface_hub
hf auth login                    # paste a WRITE token from huggingface.co/settings/tokens
hf repo create video-factory --repo-type model   # or --repo-type dataset

# 2. point this repo at it and push
git remote add origin https://huggingface.co/<your-username>/video-factory
git add .gitattributes                # ensure LFS rules are in the first push
git lfs install
git add -A
git commit -m "feat: prompt-injection episode + engine"
git push -u origin main

If you prefer a browsable player, create a Static Space instead and push the same tree; the master HTML will play directly in the browser.