This video was made by /hype, about /hype.
I have a pattern. I build something, push it to GitHub, write a README I’m proud of, and then nothing happens. Zero stars. A LinkedIn post that my friends like out of kindness. The project is fine. Nobody knows it exists.
The code was never the hard part. Getting a stranger to stop scrolling for three seconds was, and I had no idea how to do it.
Over the last few months I kept seeing people publish agent skills: small packages of instructions and scripts that teach Claude Code or Codex to do one job well. Some wrote release notes, some reviewed PRs, and a couple even produced videos. Those were the ones I noticed. My reaction wasn’t “I’ll use this.” It was “how does an agent make a video?” Language models write text. Somewhere between a prompt and an MP4 there has to be a pipeline, and I wanted to know what it looked like.
So I decided to build a skill of my own, aimed at the problem I actually have. It’s called /hype, and the tagline is the whole idea: you made something, now make people care.
What it does
/hype is an agent skill. You install it into Claude Code (or Codex, Cursor, Copilot, Gemini CLI, opencode, anything that supports Agent Skills), point it at the thing you want to promote, and it gives you finished social media videos: TikTok, Reels, Shorts, X, LinkedIn, YouTube.
“The thing” doesn’t have to be code. It can be a repo, a website, an app, an event, a bakery’s weekend sale:
/hype
/hype https://my-bakery.com weekend croissant sale, for Instagram
/hype our AI meetup this Friday 7pm at the library --aspect 9:16,1:1
/hype a 15s before/after reel: messy spreadsheet → my app's dashboard
Run it with no arguments inside a project and it reads your code, writes a brief (who is this for, what pain does it fix, what’s the proof, what should they do next), pitches 3–5 video ideas, and asks which one to make. If you already have an idea, you describe it in a sentence and it skips the pitch.
What comes out is meant to be posted, not tweaked for an afternoon: the video in whatever sizes you asked for, a poster frame, a soundtrack cut to the video, sound effects on the transitions, and post copy with hashtags written separately for each platform. Add --voice and you also get narration with word-by-word captions.
The part I didn’t know: marketing
Building the pipeline was the fun part. Writing the ideas step was the humbling part, because I had to put into instructions something I’m bad at.
I ended up giving the agent 20 marketing angles to generate ideas from: pain → fix, POV, before/after, speed demo, myth vs fact, countdown, and so on. It generates candidates across those, scores each one on four things (hook, clarity, fit, and whether it can actually be built), and only shows you the best few. That scoring step exists because most first ideas, mine included, are “here is my app, it has features.” Nobody watches that.
The storyboard step taught me the other half. Most people watch social video on mute, so every beat needs readable on-screen text and the story has to work without sound. Each platform also puts its own buttons and captions over parts of the frame, so text has to stay inside safe zones or it gets covered. None of this is complicated. I just didn’t know any of it until I had to write it down for an agent.
How an agent makes a video
This is the question that started the whole thing, so here’s the answer.
The agent doesn’t generate video with an AI video model. It writes a web page. Every video is an HTML composition built on hype.js, a small timeline runtime I wrote where every frame is a pure function of time: give it t = 3.2s and it draws exactly what should be on screen at 3.2 seconds. Layouts adapt to the aspect ratio, so one composition renders as 9:16, 1:1 and 16:9 without three separate edits. Because it’s HTML, the agent can reuse your real UI, colors and fonts instead of stock footage.
Rendering is Playwright driving headless Chromium: take a screenshot of every frame, pipe the frames into ffmpeg. Before the final render the agent looks at contact sheets of every scene in every aspect ratio, which catches text that overflows on the square version but looked fine on the wide one.
Audio was the part I expected to be ugly. Licensed music is a mess, so soundtrack.mjs synthesizes the music bed itself (8 moods) and places sound effects on the cuts. No samples, so everything is royalty-free. For voiceover, voice.mjs runs Kokoro-82M locally, 28 voices, no API key and no cost per video. The music ducks under the voice and the mix is normalized to -14 LUFS, which is roughly what the big platforms normalize to anyway.
All of those scripts work without an agent too, if you’d rather write the composition yourself.
The proof
The video at the top of the repo README was made by /hype, about /hype. Its hook is “You shipped it. Nobody saw it.”, which is the most honest line in the project. The examples folder has every intermediate file for each run: the brief, the pitched ideas, the storyboard, the captions and the composition source. If you want to know how a decision got made, it’s there.
Try it
In Claude Code:
/plugin marketplace add AliMi00/hype
/plugin install hype@hype
For any other agent:
npx skills add https://github.com/AliMi00/hype --skill hype
You’ll need Node.js 18+ and ffmpeg. Chromium for rendering installs itself on first run (about 150 MB), and the voice engine is optional.
It’s at v0.1.0, MIT licensed, and I’m sure there are rough edges I haven’t hit yet. If you make something with it, I’d like to see it. And if it breaks, open an issue. That’s also a kind of marketing, and I’m told it counts.