How I Went TikTok Viral using ChatGPT AI Agent System

Building an AI Agent System for Curation in CrewAI

So, I went viral on TikTok this week.

My account went from near-zero on every metric to 🚀 on every metric.

Here’s the Youtube version of this post:

The Struggle

I love using TikTok to experiment with AI tools. My newsletter and Youtube channel are my technical content, my LinkedIn is pseudo-technical content, while my TikTok is a free-for-all anything-goes AI experiment lab.

For several months, my TikTok videos were stuck at 200 to 1,000 views.

This happens to most creators.

…and most never escape!

Despite producing what I believed to be quality content, the engagement wasn’t there.

The data was clear.

I needed a new approach.

So, I decided to build AI agent systems to help me rapidly experiment with different types of content.

In the past 7 days, I’ve amassed over 730k views, projecting to hit my first 1 million next week.

Today is only day 66 of my content creation journey.

(my goal is 1000 days)

AI Agent System for Youtube Curation

My first AI agent system for TikTok was pretty simple:

Curate educational beginner-friendly Youtube channels to learn X topic.

Initially, I built my prototype in Flowise but have moved it to CrewAI — an open source framework for agent orchestration.

Agents can collaborate, reflect on their work, execute tasks asynchronously, interact with the external world via tools, and carry out complex workflows.

I remain bullish on low-code agent platforms and love that they enable non-technical users to build functional agents…

But I miss how productive I am text editing in emacs + evil mode (vim).

To implement my AI agent system, I defined two agents in CrewAI:

  1. Senior Researcher: responsible for discovering the best YouTube channels for beginners on a given topic

  2. Writer: responsible for writing a scroll-stopping TikTok hook

Following CrewAI’s framework, I provided a goal and backstory for each agent, as additional context to help improve output quality.

I also equipped both agents with search_tool to be able to search the internet for relevant information.

Next, I create 2 tasks:

  1. Research Task: identify Youtube channels for beginners to learn about a given topic, and explain the focus of each channel

  2. Write Task: write a viral TikTok hook about learning a given topic, focusing on why it’s important to learn

For each task, I specify which agent will execute the task, the tools that can be used, and the expected output.

The last step is assembling my crew of agents!

I specify the agents I want in my crew, the tasks assigned to my crew, the process type (simple sequence), and a couple other parameters like memory.

Running my crew, I request 3 Youtube channels for the topic ChatGPT.

In this screenshot, you can see our first agent, the Senior Researcher, sharing its thoughts.

But then stumbles:

“I tried reusing the same input, I must stop using this action input. I’ll try something else instead.”

Then, the agent pivots to another approach and proceeds smoothly.

This shows an awesome feature of agents — their ability to handle a failure and dynamically decide on another approach to try!

Here is the Senior Researcher’s final list of 3 Youtube channels for beginners to learn about ChatGPT:

Here is the Writer’s hook for the TikTok video script:

Obviously, the hook is terrible. Even the year is wrong! Maybe I should’ve supplied it with examples of strong viral hooks…

Future Improvements

While I found my AI agent system was great to spin up a rough draft list, the present version is far from perfect. I felt manual QA was always necessary.

For instance, some channels recommended by the system were way too technical for beginners.

In the future, I plan to address this by introducing a 3rd agent dedicated to quality assurance, reviewing the Senior Researcher’s work to ensure that every channel is beginner-friendly and doesn’t require technical background.

Another improvement is to further abstract concerns.

You may have noticed the Senior Researcher also writes a brief summary of each Youtube channel. Instead, I should assign such writing tasks to the Writer.

Last Thoughts

In my view, using this AI agent system had 2 key benefits:

Efficiency: The automation of researching Youtube channels saved me time and effort, especially in subject matters where I’m no expert. It didn’t take long to verify that each channel was a good fit and beginner-friendly.

Scalability: After one successful TikTok video, I received requests for many more curated lists of Youtube channels to learn X topic. This system enabled me to quickly answer those questions, catering to more diverse audience interests and keeping engagement high.

I don’t think AI was necessary in my success this past week.

But, using AI gives my TikTok content creation a fun creative twist 🙂 

Have fun building!

Sabrina Ramonov

P.S. If you’re enjoying the free newsletter, it’d mean the world to me if you share it with others. My newsletter just launched, and every single referral helps. Thank you!