I watched a token counter climb for four minutes straight and thought, this is the dumbest money I have ever spent.
The setup looked fine on paper. Client had Otter wired into the cloud instance. Claude had access to all of it. I asked one simple question about what he had covered on his calls that month, and it went off and started reading raw transcripts. All of them. Hundreds of hours of a man talking, in full, with the ums left in.
It came back with an answer. It cost a fortune to get there and the answer was mid. That was the moment I actually understood what an AI content system is, and why the pile of tools I had connected together was not one.
TLDR
- An AI content system is not ChatGPT plus Otter plus Notion. That is three tools standing next to each other.
- The system is the layer in between them, and almost nobody builds it.
- Raw material is not knowledge. Your model has to read a transcript. It can query a structure.
- One client of mine runs 16 to 20 coaching calls a week. Every one of them was sitting there unreachable.
- The end state is a published blog in about 3 to 5 minutes, in his voice, that he did not write. Getting there took one boring step first.
A Tool Answers. An AI Content System Remembers.
Here is the cleanest way I can put it. A tool responds to what you give it. A system already knows.
When you open ChatGPT and paste in a transcript and ask for five social posts, you are using a tool. It works. It will keep working, once, every single time, forever, because nothing you did in that session is available to the next one. You are the memory. You are the part that carries context from Tuesday to Thursday. Which is fine until you are running this across a year of calls and you become the bottleneck for a thing you built to remove the bottleneck.
An AI content system is what you get when the context lives somewhere other than your head or your chat history. It is a structure that sits between your raw material and the model, so the model is not starting from zero every time you ask it something.
That is the whole definition. Everything else is implementation detail.
The Layer Everybody Skips Is the Only One That Matters
So what did I actually do differently the second time.
I did not swap models. I did not write a better prompt. I built the middle layer, which in this case was a Notion database, and I ran every transcript through a synthesis pass on the way in. Not a summary. A structured extraction. Each call came out the other side as a row with fields:
- call title
- summary of what was actually covered
- email angles
- social angles
- quotable moments, the lines where he said something well
- other topics that came up in passing
That is it. That is the unglamorous part, and it is the part that turns a folder of files into an ai content pipeline you can query.
Because now when I ask a question, the model does not read three hundred transcripts looking for the answer. It queries a table. It knows where to look. It costs a fraction of what it cost before, and the answer is better, which is a combination I do not usually get.
The first pass was still slow and expensive. Fetching everything, synthesizing everything, that took real time and real tokens. But you pay it once. After that the ai content workflow runs cheap forever.
What the Chain Looked Like End to End
Once the Notion layer existed, the rest stopped being hard.
I built a Claude skill that takes the relevant rows and turns them into a blog post. Not a generic blog post. One that reads the way he talks, because the skill has his examples and his patterns and his anti-patterns loaded into it. The skill connects to GitHub, GitHub pushes to Vercel, Vercel builds and deploys.
So the actual operating instruction became: publish a blog on this topic.
It pulls the topic. Finds the matching material in Notion. Writes it in his voice. Commits it. It goes live. Three to five minutes, start to finish, and nobody sat down to write anything.
I want to be careful here because this sounds like a magic trick and it is not. Every one of those steps is boring on its own. The skill file took a lot of iterations before it stopped producing confident garbage. The thing that made it work was not the automation at the end. It was the structured knowledge in the middle. If I had connected Otter straight to the writing skill and skipped Notion, I would have built something expensive that produced slop faster.
Why the Output Stops Sounding Like Everyone Else
This is the part that matters if you are a coach or a consultant reading this and wondering whether any of it applies to you.
Generic AI output is not a model problem. I know that gets said a lot. It is still true. The model is generic because you handed it nothing specific, so it reached for the average of everything it has ever seen. Of course it did. What else was it supposed to do.
When the system has your actual phrasing, your actual client stories, the exact way you explain the thing you explain forty times a year, the output stops being average. It comes back sounding like you had a good day writing. And that matters commercially, not just aesthetically, because Google's EEAT framework is basically a formalized version of the question, did a person who actually does this write this. Real experience, real examples, real point of view. A stat-stuffed post assembled from other posts fails that test. A post built from a year of your own calls does not.
Most coaches I talk to already have the raw material. They have more of it than they think. What they do not have is the layer that makes it reachable, which is exactly the thing I build for coaches and the reason the voice skill exists as a standalone piece.
Here Is the Part I Am Still Chewing On
The honest version of this is that the system is not the impressive bit. The synthesis pass is. And the synthesis pass is the least fun work in the entire build. It is data hygiene wearing a nicer jacket.
Which puts you in a weird spot, because the thing standing between you and a business that produces content without you is not a tool you can buy. There is no checkout page for it. It is a few days of deciding what fields matter, running your archive through them, and being fairly opinionated about the structure. I do not think most people will do it. I think the ones who do will look, from the outside, like they got lucky with AI.
I have not fully worked out what happens when everyone has this. Right now the advantage is that almost nobody has bothered. That is not a permanent advantage. But it is the one that is available today.
If any of this sounds like where you are, it is probably worth a conversation.
Book a free call. 30 minutes. I will tell you honestly whether I can help or whether you should just go build it yourself.