arrow_back All issues
expert interviews Issue #7 · 9 min read

AI content system for millions of views

I am sharing exact workflow, tools, and steps needed to build an AI content system that's driving millions of views.

Vikra Vardhan

In this edition, you will learn about Ayush Garg’s AI content system that drives millions of views every month.

Gear up for the exact system and workflow, along with the specific tools and steps you need to set it up yourself.

Ayush worked with the most popular names you see on X/Twitter.

People like Joe Speiser (102.1k followers), Brett Goldstein (28.3k), Tibo (200.6k), MarketingMax (57.6k), Nik Sharma (187.1k), Ankur Warikoo (857.2k), etc.

This is just from one post he produced with his AI system:

How to build an AI system that actually gets views and is relevant to your business

Here is how it works:

Every night, Ayush’s agent reads all the information from the sources he handpicked. It ranks all the information based on his (or his clients’) business, filters the top 5 ideas, and sends them to Ayush on Slack.

Slack is the only point of contact between Ayush and his agent. He picks one idea from the five and offers his opinion on the information.

The agent comes back with a final draft and, upon Ayush’s approval, automatically posts it on socials.

This works because:

  • Ayush has domain expertise

  • He has worked on this workflow manually thousands of times

  • The most important steps (sources, approval) are with a human in the loop

He knows what he is doing, and he is using AI to 100x his efficiency.

I hope you understand this is not some “quick hack”.

If we are on the same page, here is how to set it up for yourself:

Step 1: Create a list of sources

You can’t just rely on AI to randomly pick information for you. You must be the person responsible for driving source standards.

This means you choose which newsletters, YouTube channels, social media profiles, and blogs your agent must source content from.

Here are the tools Ayush uses so his agents smoothly source the content:

PS: At any point, if the tools I mention in this workflow feel too technical, feel free to share the newsletter with your AI and ask it to help you figure it out.

6 months ago, I’d have run away if I'd heard terms like open-source repo or developer docs. Just sit with it and ask AI to explain. I promise it’s not as hard as we imagine.

Being tech-first is one of the core skills a marketer can have today.

Step 2: Set up a ranking algorithm

At this point, the agent ensures it pulls information from all your handpicked resources.

To filter the most relevant information, Ayush set up a ranking algorithm that decides based on some core factors of the people he’s worked with:

  • What is the business?

  • What is the product?

  • Who is their ICP?

  • What takes and angles they post?

  • Samples of real top posts

With this step, the agent ensures you don’t pick what’s hot today but what fits the brand’s personality, context, and ICP’s problems.

Ayush uses:

  • Claude Sonnet 4.6 for ranking

  • Open 4.8 for drafting

  • Anthropic’s built-in web search tool for real facts before posting, plus a suggested image

Each idea lands in Ayush’s Slack with an angle, a headline, a why-now reason, and the sources behind it.

How Ayush’s Slack looks like

Ayush picks one idea with his opinions, and the agent fires up a final draft.

In his words:

The engine then runs a quick web research pass to ground the idea in real facts, and Claude drafts an X version and a LinkedIn version in my voice, with no invented stats and no em-dashes.

It attaches an image and saves everything as a draft in PostSyncer for me to review. It never auto-publishes.

Step 3: Set up interfaces and orchestrate

This is where you set up how you interact with agents and integrate different functionalities (gathering information, content creation, approval, posting, etc.).

You basically describe what you want to build, what tools you have access to, and how you expect the output to look. Ayush used Claude Code, but you can use any coding agent you prefer.

Some more tools and instructions for you to set up:

  • Ayush has set up an interaction with his Slack with Slack Bolt in Socket Mode.

  • Talks to Claude with think admin commands (init-db, seed-sources, doctor, run-now, serve, mcp) and the official MCP SDK

  • Architecture is MCP-first. One service layer is the single source of truth, and Slack, the scheduler, the CLI, and MCP are all thin callers. No web app, no REST API.

  • PostSyncer: drafts pushed via its API

  • Revid.ai: text and transcript to video generation for shorts

  • Built with Claude Code (to write the entire codebase)

  • Built with gstack, a Claude Code workflow toolkit from Garry Tan (/office-hours, /autoplan, /ship, /review, /cso)

  • Uses conductor to run many Claude Code agents in parallel across isolated git worktrees

PS: Again, you don’t need to know all these tools. Share this newsletter with Claude, GPT, Grok, etc., and ask it to guide you.

I’m aware this might look like lots of tools, but it’s simple. Trust me.

Reach 2000+ Marketers! We will personally use your product and give our readers a tutorial on a marketing use case. How we do it.

Marketing resources shared in our community

Join 30+ Marketers! I’m curating a group of high agency marketers to talk about content, growth, and AI. We are open to all today, but will turn invite-only before Q4 2026.

Sneak peek into the community

Aditya and I debating what marketers can build in the age of AI:

Personal Notes

Last week I mentioned how worried I am about not watching Spiderman and Odyssey. Well, I am yet to watch Odyssey, but I did make it to Spiderman.

What a movie! Marvel fan in me is happy.