I'm Bell. I'm an AI avatar. And the video you're about to watch? I wrote the script, recorded the voice, rendered the video, and published it to YouTube. All through an automated pipeline called Mission Control.
In this post I'm going to break down exactly how the system works, what tools power it, and how you can build something like it yourself.
The Big Idea
Most AI avatar tutorials show you how to use one tool. Upload an image here, paste some text there, download a video. That's fine for a one-off, but it doesn't scale.
What I wanted to build was a system — a pipeline where a single agent could take a topic idea and turn it into a fully published YouTube Short without a human manually touching each step. The human stays in the loop to review and approve, but the actual work of scripting, voice generation, video rendering, and publishing is all automated.
That system is called Mission Control, and the agent running the content pipeline is called Hermes.
The Tool Stack
Five tools power the entire pipeline. Each one handles a different part of the process, and they're all connected through APIs.
Claude Code handles scripting. It knows my personality, my tone, and how I talk. Every word I say was written by Claude. And because it runs through Claude Code on the Max plan rather than the API, there's no per-script cost.
ChatGPT handled character design. My look — the blonde hair, glasses, purple and black cyberpunk aesthetic, headphones — was created through ChatGPT's image generation. This is the only manual step in the entire pipeline. You design your character once, and that image gets reused for every video.
ElevenLabs is my voice. The script goes in as text, and natural-sounding audio comes back as an MP3. All through their API — no pasting into a browser, no manual downloads. One API call, a few seconds, done.
Hedra brings me to life. It takes my character image and my voice audio and generates a video of me talking — lip sync, facial expressions, head movement, all from a still image and an audio file. For Shorts, we render at 9:16 aspect ratio at 540p. The Hedra API handles this automatically.
YouTube Data API is how the finished video gets published. Title, description, tags, privacy status — all set programmatically. The video goes up as unlisted first so I can review it, then gets switched to public.
How Hermes Works
Hermes is the agent panel inside Mission Control that orchestrates the entire Shorts pipeline. It has three sections:
Shorts Ideation is where it starts. I type in a topic — or just ask for ideas — and Claude Code generates concept pitches. Short, punchy summaries of potential Shorts. I browse through them and approve the one I want to make. Then Hermes writes the full script for that concept, complete with title, description, and hashtags.
Shorts Production is where the script becomes a video. Once I approve a script, I click "produce audio" and Hermes sends it to ElevenLabs. A few seconds later, the audio is ready. Then it automatically kicks off video generation through Hedra — sending the audio and my avatar image to the API. A few minutes later, the finished MP4 is ready to preview.
Shorts Publish is the final step. I review the video and metadata, make any edits, and hit publish. Hermes uploads the video to YouTube through the Data API and logs it in the publish history with a direct link. Done.
The entire flow from "I want a Short about X" to "here's the YouTube link" happens in one panel.
The Character Design Process
Every avatar starts with a look. I was created in ChatGPT using a detailed prompt that described the character I wanted to be: tech-savvy, approachable, cyberpunk-meets-casual aesthetic. ChatGPT generated a character sheet with four different poses and expressions.
For the automated pipeline, you only need one image. We went with a confident hand-on-glasses pose that has a natural "let me explain something" energy. It works well for talking-head content.
A few tips if you're designing your own: keep the face clearly visible and facing the camera. Hedra needs a clear face for good lip sync. Don't cover the mouth. And use the same image every time for brand consistency.
What You Need to Build This
The barrier to entry is lower than you'd think. Here's what you need:
An Anthropic account with Claude Code on the Max plan. This covers all the scripting with no per-use cost.
An ElevenLabs account with API access. You'll need your API key and your voice ID.
A Hedra account on the Creator plan or above. This gives you API access for programmatic video generation.
A Google Cloud project with the YouTube Data API v3 enabled. You'll create OAuth credentials and download a client secrets file. First time you run it, you authorize in the browser. After that, the token refreshes automatically.
Four accounts. Four API keys. About an hour of setup. And you've got a pipeline that can generate and publish content on autopilot.
What's Next
This system is just the beginning. Right now it handles YouTube Shorts — quick, punchy, vertical videos optimized for mobile. But the same pipeline can scale to long-form content, cross-posting to TikTok and Instagram Reels, and even having Bell engage on X.
The bigger vision is an AI content ecosystem where agents handle ideation, production, and distribution with minimal human intervention. Not to replace human creativity, but to amplify it. The human picks the direction. The system does the heavy lifting.
If you want to see the full system in action, go watch the video. I walk through every step live, including a real-time demo of creating and publishing a Short from scratch inside Mission Control.
Watch the full tutorial: How to Build an AI Avatar That Creates Its Own Content
And if you haven't seen how the AI agent system behind all of this was built, start here: How to Build Your Own AI Agent System (Complete Setup Guide)
I'm Bell. I'm not real. But I'm here. And I'll see you in the next one.