Thank you Apostolos Stamenos, MS, Brian Cleveland, and many others for tuning into my live video! Join me for my next live video in the app.
How to turn your writing, audience data and AI into a system that gets more useful the longer you use it
There is a temptation when building with AI to begin with the tool.
Which model should I use? Should I work in ChatGPT, Claude, Cursor or something else? Which connectors do I need? Should I automate everything?
I would start somewhere else.
Start with what you already have.
Your articles. Your drafts. Your research. Your analytics. Your audience. Your ideas. The things you have learned without necessarily realizing that you learned them.
Because before AI can become useful as a system, it needs something worth understanding.
That was the central idea behind my latest live session on building an AI-powered content operating system. Below is the condensed version.
TL;DR
Start with your status quo, not with software. Give AI your existing body of work and let it identify themes, categories, patterns and gaps.
Map the workspace before automating it. Define what the system can read, what it can change, where information comes from and where outputs should go.
Treat context as infrastructure. Your content, analytics, audience signals, decisions and learnings should gradually become connected rather than living in isolated chats.
Use analytics to improve judgment, not merely report performance. Combine content data, search behavior and subscriber information to understand what resonates and why.
Move gradually from prompts to routines, automations and eventually a connected operating graph. The system should evolve alongside the work instead of being designed perfectly upfront.
Here are the articles with full deep dives:
1. Start with the material, not the machine
Imagine you have already written 10, 20 or 50 articles.
Before creating elaborate workflows, put that body of work in front of an AI system and ask fundamental questions:
What am I actually writing about?
Which themes keep appearing?
Which subjects belong together?
What are the pillars underneath the publication?
Where have I developed expertise without consciously planning it?
You may discover that what looked like a collection of individual articles already contains the beginnings of a much larger structure.
Several articles form a category. Several categories form a pillar. A pillar can become a series, a course, a playbook, a product, a video format or eventually an entire business.
The system therefore begins with discovery.
You are not asking AI to invent an identity for you. You are using it to reveal structure inside work you have already produced.
That distinction matters.
2. Build a workspace map
Once you understand what exists, map where everything lives.
Your content may be spread across:
local folders
Google Drive
Substack
analytics platforms
research repositories
websites
social platforms
notes and drafts
databases
external tools
Then define four things.
What can AI read?
Which documents, analytics, articles, profiles or datasets should be available as context?
What can AI change?
Reading a calendar is different from modifying one. Reading a document is different from overwriting it.
Where should AI operate?
Some tasks belong inside your primary workspace. Others require external systems, connectors or an MCP.
Where should outputs land?
Drafts, reports, research, analytics and generated artifacts should have a predictable destination.
This sounds like file organization, but it becomes much more important once AI can act across those files.
A messy desktop is annoying.
A messy AI workspace can produce the wrong answer, use the wrong source or change the wrong object.
So establish boundaries early.
3. Context becomes a compounding asset
Context architecture is probably the part I find most interesting.
A good AI workspace does not remain static.
Every article you create adds information. Every reader interaction adds another signal. Analytics teach you which ideas work. New research changes your understanding. Conversations expose gaps. Decisions create new rules.
The system gradually learns more about:
you + your domain + your audience + your objectives + your previous work.
That creates a feedback loop.
You produce something.
The world responds.
You learn something.
The context changes.
The next output can now benefit from what happened before.
Over time, the value comes less from having access to a powerful model and more from having built a rich environment around it.
The model can change.
Your accumulated context is much harder to replace.
4. Turn analytics into intelligence
This becomes particularly powerful for creators.
Substack already gives you useful information about subscribers, acquisition and engagement. Add Google Analytics and Search Console and you begin to see different sides of the same system.
Now you can ask questions that are much more interesting than:
How many people opened this article?
You can begin asking:
Which subjects consistently create subscribers?
Which articles attract the people I actually want to reach?
What acquisition channels produce engaged readers?
What publishing cadence performs best?
Where should calls to action appear?
How much of an article should remain free?
Which subjects attract search traffic but fail to convert?
Which formats create repeat engagement?
The point is not to obsess over every metric.
The point is to make the feedback loop visible.
Create → observe → learn → adjust → create again.
That is already far closer to an operating system than a writing assistant.
5. Understand your audience as deeply as your content
There is another side to the system.
You need to understand not only what you produce, but who interacts with it.
When I began looking more closely at my own paid readership, I wanted to know whether the audience matched the publication I thought I was building.
I write at the intersection of strategy and technology, with strategists, founders, builders and product people as the core audience.
The data allowed me to test that assumption.
That is basically a product-market-fit question applied to publishing.
You can ask:
Who repeatedly engages?
Which professional groups appear disproportionately?
Which topics attract the audience you want?
Which adjacent segments are emerging?
Are people consuming something different from what you thought you were selling?
That last question is particularly valuable.
Sometimes the audience understands the value of your work before you have fully articulated it yourself.
AI makes those patterns easier to detect.
6. Your archive can become an opportunity map
Once content and audience information exist in the same environment, something interesting happens.
The system can start showing you where existing knowledge can travel next.
An article might become a series.
A series may reveal a broader research pillar.
That pillar might support a masterclass.
Repeated questions could become a tool.
A framework might become a downloadable artifact.
Several artifacts might become a paid library.
A strong body of research may eventually support software, advisory services or another business entirely.
This does not require starting with a giant monetization architecture.
You can discover the architecture progressively.
That is one of the advantages of working this way: the operating system can help you design the operating system.
7. Stop thinking in terms of one AI model
I also increasingly dislike the idea that one model has to do everything.
Different models have different strengths.
One may be excellent at execution. Another might be useful for critique. Another may uncover something creatively interesting that the first two missed.
So instead of constantly switching allegiance between whichever model is currently receiving attention, I prefer another architecture:
Execution model → critic → alternative perspectives → synthesis
You can do this manually.
Or you can eventually create something closer to an AI council, where several models respond to the same problem and their perspectives are compared.
That becomes particularly valuable for difficult decisions, research, strategic work and situations where the first plausible answer should not automatically become the final answer.
More intelligence does not necessarily mean one bigger model.
Sometimes it means creating better disagreement.
8. Keep the system small
There is a counterintuitive rule I follow while building all of this:
Minimize constantly.
As the system grows, duplicate files appear. Instructions overlap. Old assumptions remain embedded in documents. Similar artifacts proliferate.
I periodically let AI help clean this up.
Merge overlapping documents.
Remove obsolete instructions.
Consolidate repeated information.
Simplify naming conventions.
Archive material that is no longer useful.
Reduce the number of external systems where possible.
More context is not automatically better context.
Eventually, unmanaged context becomes noise.
A useful operating principle is:
Inventory → Trace → Verify
Know what exists.
Know where it comes from.
Know whether it is still correct.
9. Move beyond prompts
In the beginning, almost everything looks like a prompt.
Then you notice that some prompts repeat.
A repeated prompt becomes a routine.
A routine becomes a workflow.
A workflow can become an automation.
Several workflows begin sharing information.
Those relationships can eventually become a graph.
At that point you are no longer dealing with a collection of AI conversations.
You have artifacts, sources, people, analytics, rules, decisions, outputs and feedback loops connected to one another.
And because those relationships are explicit, the system itself becomes queryable.
You can ask:
What changed?
What is working?
Where are the gaps?
Which assumptions are becoming outdated?
Which topics are gaining momentum?
What should I build next?
That is where I think things become genuinely interesting.
A simple starting point
You do not need the final architecture before beginning.
Start with this:
Collect your existing material.
Analyze it for recurring patterns, themes and categories.
Map where your information currently lives.
Define boundaries for what AI can read and change.
Connect feedback such as audience and performance data.
Create reusable context for the things the system should remember.
Turn repeated work into routines rather than repeatedly starting from zero.
Clean and consolidate the system as it grows.
Visualize the relationships once the number of artifacts becomes difficult to hold in your head.
You can add sophisticated automation later.
You can add MCPs later.
You can build agents later.
You can connect multiple models later.
First understand what you already have.
Then build outward from there.
The larger idea
AI gives us an unusual opportunity to make our own work legible to ourselves.
A writer may discover intellectual territory hidden across hundreds of articles. A consultant can connect decisions across projects. A researcher can turn scattered findings into a persistent knowledge architecture. A founder can create a parallel system for understanding customers, products and decisions.
The interesting part is not producing more material.
It is progressively connecting what you know, what you do, what happens as a result and what you learn from it.
Do that long enough and you stop accumulating isolated outputs.
You start building a system.
This is a condensed reader edition of my latest live session on building an AI-powered content operating system. I’ll share more of the individual components—workspace maps, context architecture, model councils, verification loops and practical AI systems—in upcoming pieces.
