What This Is
This is the actual engine I use to turn a topic into a finished vertical reel — an AI presenter, real sourced footage, motion graphics, captions, SFX and music — rendered entirely as code, with zero manual editing. Output is 1080×1920, 30fps, mastered to −14 LUFS, ready to post to Reels, Shorts or TikTok.
It's built on Remotion (video as React) + Python build scripts + Whisper word-timings, and it's designed to be driven by a coding agent. You point Claude Code (or Cursor, Codex…) at the folder and say "Read AGENT.md, then make a reel about [topic]" — the folder's operating manual takes it from there.
You are not editing video and you are not writing code. You open the folder in an AI agent and talk to it in plain English — it does the building, and you give feedback. If you can chat with ChatGPT, you can run this.
What It Does
Reels as Code
1080×1920, 30fps, mastered to −14 LUFS — no timeline, no manual editing.
AI Presenter
HeyGen (or any) avatar + voice — one continuous master per reel.
Sourced Footage
Scouts real clips & receipts first, credits them — never fakes proof.
34 Scene Types
Receipts, spec sheets, comparisons, stat cards, terminals, carousels…
Word-Perfect Timing
Whisper word-timings anchor every cut to the exact spoken word.
Style Packs
Swappable creator looks + script voices (Varun Mayya, Nick Saraev).
What's Inside the ZIP
- AGENT.md — operating manual
- PIPELINE.md — technical spec
- STYLE-RULES.md — rules ledger
- README + config template
- types.ts — the scene contract
- 34 scene components
- Reel.tsx — the player
- theme tokens + fonts
- Python build scripts
- VO-tighten, frame-lint, capture, Manim tools
- Varun + Nick style packs
- A real example beat sheet
What You'll Need
Really, just two things are on you — the agent handles the rest:
- An AI coding agent — Claude Code, Codex or Cursor. This is what reads the folder and does all the technical work.
- An avatar / voice account — HeyGen by default (swap freely). You paste your IDs when the agent asks — nothing to set up beforehand.
Under the hood it also uses Node, Python, ffmpeg, yt-dlp, Whisper and headless Chrome — but you don't need to know what any of those are. Ask the agent "what do I need installed?" and it checks your machine and walks you through anything missing. It all runs locally and is free.
No credentials, avatars, voice samples or music are included. You add your own avatar/voice IDs to config.json (there's a config.example.json template), and your own royalty-free music beds to public/music/ — a README note inside shows exactly where.
How to Use It — Just Open It and Ask
Here's the part that trips people up, so let's be clear: you don't run this like a developer. You open the unzipped folder inside an AI coding agent — Claude Code, Codex or Cursor — and simply talk to it. The whole folder is written to be read by the AI: there's an operating manual (AGENT.md) it reads first, so it already knows the entire workflow before you say a word.
- Open the folder in Claude Code (or Codex / Cursor). Drag the unzipped folder in — that's the whole "setup".
- Just ask. Type "How does this automation work?" to get oriented, or go straight to "Set this up and make a reel about [my topic]." It reads the manual and walks you through everything in plain English.
- Answer its questions. When it asks, paste your HeyGen avatar & voice IDs and drop in a royalty-free music track — it tells you exactly where each one goes. You're answering prompts, not memorising commands.
- It builds the reel — scouts real footage, writes the script, generates the presenter, renders and masters it. You watch and give feedback ("make the hook punchier", "swap that clip") — every scene is regenerable.
You never have to understand a single command. The agent installs whatever's missing, runs everything, and explains what it's doing as it goes — your job is to pick the topic, hand over your avatar/voice, and say what you like or don't. It's directing, not coding.
Prefer to run the commands yourself? (for technical users)
The agent runs these for you — but if you'd rather drive it manually, you'll need Node 18+, Python 3, ffmpeg, yt-dlp and headless Chrome on your PATH.
npm install
cp config.example.json config.json # paste your HeyGen IDs
pip install openai-whisper
npx remotion studio # live preview
# per reel (the agent normally does this):
npx remotion render <slug> out/<slug>-raw.mp4 --concurrency=2 --timeout=120000
ffmpeg -i out/<slug>-raw.mp4 -af "loudnorm=I=-14:TP=-1.2:LRA=7" -c:v copy out/<slug>-final.mp4
How It Works Behind the Scenes
The whole pipeline, in the order it always runs (the order is the point):
1 · Scout
Verify the story, hunt real footage & receipts, write an asset manifest.
2 · Direct
Script bound to assets — every beat maps to a real clip or a graphic.
3 · Voice + Face
One continuous avatar master; Whisper extracts word-level timings.
4 · Beat Sheet
The build script anchors every cut to the voiceover's words.
5 · Render
Remotion renders the React scenes to 1080×1920, 30fps.
6 · Master + QC
Normalize to −14 LUFS, then a hostile-viewer critic pass.
Never script a claim you can't show. Source the footage first, then write to it. If an asset doesn't exist, re-scout for it or rewrite the line — never paper over the gap with a text card. It's the difference between a reel that shows and one that just tells.
Example Prompts You Can Paste
Give one of these to your coding agent inside the folder:
"Read AGENT.md, then make a 40-second reel about [new AI model/launch], Varun Mayya style. Scout real footage for every claim before you script."
"Make a reel comparing [Tool A] vs [Tool B]. Use the spec-sheet and hcompare scenes for the differences, Nick Saraev style. Credit any third-party footage on screen."
"Turn this script into a reel: [paste script]. Scout footage for every claim first, and rewrite any line you can't show with real footage or a graphic."