Content Creation on Autopilot

I have been a content creator in the BJJ niche for a decade and a half now. When I started playing with Ai and Machine Learning my vision was very narrow - I wanted to teach a LLM how to do jiu jitsu! Ha!

Not impossible but a task tantamount to it with the depth and breadth of the nearly infinite permutations of the way the human body moves in conjunction with another human's body trying to manipulate it with nearly infinite paths of resistance and compliance. But I digress... 🤓

I saw a friend building a project to set an agent free researching content in a specific niche and then generating short YouTube clips based on its findings. The best part - it was gathering statistics and analytics to use as fuel for self-improvement! He had it set up so each iteration started in a new session and used analytical data from the last to improve upon it. I was fascinated and decided to try to reproduce it myself.

bjjDigest.com was born.

Today I have a near-fully autonomous Content Creation Engine that I set free on the BJJ community. Calibrated to my "voice" from a decade and a half of my own content - it generates News Articles for real-world Current Events in the BJJ space, with a snarky, sarcastic, dry tone (me). It also writes Satire Articles - I'm pretty proud of this part. As LLMs aren't known for their comedic timing or judgement. The weakest link in this chain (in my opinion) was actually its impetus - the YouTube shorts.

And yes - it closes the loop I saw in my friend's project, and honestly this is the part I'd point to first. The research never really stops. Every couple hours it scans the BJJ landscape - news, forums, RSS, event results - and scores what it finds, ranking the topics worth writing about and working out the angle on each one. That scored board is what feeds the pitch generator - nothing gets written that the research didn't surface first.

Then the other half of the loop runs backward. It reads its own analytics - what actually landed on YouTube and the sites - and uses that to re-weight what it goes looking for and which topics it prioritizes next. Each run starts in a fresh session and inherits the last run's data, so it's not really "generate content" so much as "generate, watch how it did, adjust, repeat." Slowly, it's learning what this audience actually responds to.

The two content lanes stay strictly separate through all of it: bjjProblems is snark on real news (never fabricated), thePorra is pure fiction - there's even a linter that blocks made-up "facts" from leaking into the real-news side.

All autonomously.

Only one manual step in the process. After each pipeline's research is complete it returns potential articles - "pitches." I then go through them and approve the good ones and reject the bad ones. The rejected pitches get sent into a queue where the feedback I give them (a list of checkboxes at reject time) is used to improve the pitch. They're sent through research and approval once again before they're retired permanently if they fail the second time. Each pipeline goes through the same process and publishes to 5 different locations:

- bjjproblems.com/news

- thePorra.com

- bjjDigest.com

- Instagram

- YouTube

Every piece discloses it's AI-written, cites its sources, and never pretends to be human. That part's non-negotiable.

It has gotten to the point where the engine runs mostly on its own. The thing that keeps it from going off the rails: the scripts are the boss, the AI is just a tool. The LLM only ever writes - it never decides what publishes, when, or where. Those calls are all hard-coded. That boundary is the whole ballgame. But there are some days where I dive in and try to "improve" something which turns into a whole thing 😆

Everything runs with Claude on the command line using an old MacBook Pro. Pretty simple.

Still not making money on any of this btw 😆 It's more of a passion project.

Happy to discuss any part of this project - both questions and critiques are welcome!

14 comments