Skip to main content
These projects demonstrate what’s possible when you combine Video Agent with AI coding agents, browser extensions, CI/CD pipelines, and more.

​
README-to-Video

Auto-generate video walkthroughs from GitHub README changes. A GitHub Action watches for README changes, uses Claude to write a scene-by-scene production prompt, and sends it to Video Agent. The rendered video is automatically embedded back in the README.
  • HeyGen features: Video Agent API
  • Stack: TypeScript, GitHub Actions, Claude
  • Key insight: The quality gap between good and mediocre videos comes down to the prompt. This project uses an LLM to write production-quality briefs with specific visual directions — not just narration scripts.
  • Cost: ~$0.05–0.15 per video
Instead of passing the README text directly to Video Agent, the system uses a two-stage prompt:
  1. Meta-prompt — Claude receives the README content + instructions on how to write a great Video Agent prompt (scene structure, visual style, B-roll descriptions, pacing)
  2. Video prompt — Claude outputs a detailed production brief that Video Agent can execute
This “prompt-that-writes-a-prompt” pattern is reusable for any content-to-video workflow.

​
Viral Video Pipeline

Research trending topics, then batch-generate short-form videos. Researches trending self-improvement topics via web search, then generates 6 TikTok/Reels/Shorts-ready videos in one run. Fully automated — no camera, no mic, no editing.
  • HeyGen features: Video Agent API (batch generation, portrait mode)
  • Stack: Claude Code + HeyGen Skills
  • Key insight: Rate limit handling is critical for batch generation. This pipeline fires videos sequentially with 5–10 second gaps and tracks all video IDs for async polling.
  • Output: 6 videos (25–40s each) + batch report with performance predictions
  • Cost: ~$6 in HeyGen credits for the full batch

​
Site2Video — Chrome Extension

One-click: turn any website into a professional, brand-consistent video. A Chrome extension captures a full-page screenshot, analyzes the site’s visual DNA (colors, typography, layout), generates a style-aware Video Agent prompt, and renders a branded video.
  • HeyGen features: Video Agent API, Asset Upload, 1,200+ avatars
  • Stack: Vite + React (extension), Next.js (backend), Gemini/Claude (LLM)
  • Key insight: Every prompt is generated from scratch using LLM analysis — no static templates. The system extracts visual style from the page itself and translates it into Video Agent prompt instructions.
  • Video modes: Founder Pitch (60s), Product Walkthrough (90s), Teardown (75s), Investor Summary (45s), Social Ad (30s)
  • Visual styles: 14 curated styles + Auto mode that extracts style from the website

​
AI News Broadcast

Automated daily AI briefings: scrape → script → render → distribute. A pipeline that gathers AI papers from arXiv and Hacker News, builds a script with an LLM, generates a video via Video Agent, and posts it to Telegram.
  • HeyGen features: Video Agent API
  • Stack: Bun + TypeScript
  • Key insight: The modular architecture (research → script → video → deliver) makes each stage independently testable and swappable. You could replace the Telegram delivery with email, Slack, or YouTube upload.

​
AI Mafia — Live Avatar Game

Social deduction game with AI-powered Live Avatar NPCs. A Mafia/Werewolf game where 3 AI players argue, accuse, bluff, and vote in real-time using HeyGen Live Avatars. Each NPC has a distinct personality and is powered by Claude for decision-making.
  • HeyGen features: Live Avatar SDK (real-time streaming)
  • Stack: Next.js, React, HeyGen Live Avatar SDK, Claude
  • Key insight: Live Avatars enable real-time interactive experiences — not just pre-rendered videos. The NPCs read game state, develop strategies, and respond with natural speech and expressions.
  • Characters: Maria (expressive), Chen (calm analyst), Alex (emotional reactor)
This project uses Live Avatars (real-time streaming), not Video Agent. It showcases a different interaction model — live, conversational AI rather than pre-rendered video content.

​
Build Your Own

These projects share common patterns you can reuse:
  1. Content → LLM → Video Agent prompt — The meta-prompt pattern works for any content type
  2. Batch generation with rate limit handling — Sequential queuing with status tracking
  3. Style extraction → prompt instructions — Translate visual context into Video Agent language
  4. Modular pipelines — Separate research, scripting, rendering, and delivery stages
Start with a workflow, learn the prompt techniques, and build from there.