Worldview Engineering

Prompt engineering will be a commodity skill inside of three years.

Not because it stops being useful. Because everyone will have it. The tutorials are everywhere. The frameworks are on GitHub. The syntax is not hard. You learn it the way you learned keyboard shortcuts — a little at a time, until one day you realize you cannot work without it and cannot remember when you acquired it.

That is not a criticism. It is a pattern. And I have watched this exact pattern run enough times now to know where it ends.

I remember when building a website was a real skill.

I spent fifteen years inside IT before I ever touched marketing — building networks, standing up servers, running the infrastructure that businesses quietly depended on and never thought about until it broke. Back then, if you wanted a website, you came to a shop like ours and we charged you real money for it, because building one took people who knew what they were doing. The skill was the moat. Then WordPress showed up, the CMS wave hit, and the thing we used to charge for became the thing anyone could do in an afternoon. The skill did not get less real. It just stopped being rare.

So I moved upstream. By 2016, before the big CRM platforms had attribution built in, I was working with a developer to stitch a closed-loop attribution system together out of cookies and Gravity Forms — because knowing what actually drove a sale was still rare, and the people who could see it were a step ahead of the people who could only build the form.

That has been the shape of the last fifteen years for me. Websites, then SEO, then social, then content. Every single time, the operators still talking about the tactic were already behind the operators who had moved to the layer underneath it.

Every time a capability becomes table stakes, the differentiator moves upstream.

AI is the same pattern wearing newer clothes. The prompt is the website. The skill everyone is rushing to learn right now is the skill that will be sitting in everyone’s hands inside three years.

The layer underneath prompt engineering is not better prompts.

It is what the system believes.

What a worldview actually is

I want to define the word before I use it the rest of the way, because worldview is one of those words people nod at while meaning different things.

A worldview is not an opinion. It is not a preference. It is not belief in the soft sense.

It is the instrument of vision. The frame of reference that determines which problems you recognize, which evidence you weight, which solutions you can even see. C.S. Lewis named it as cleanly as anyone has:

I believe in Christianity as I believe that the Sun has risen, not only because I see it, but because by it I see everything else.

That is worldview as instrument. Not the content of vision. The means of vision. The thing you see by, not the thing you see at.

Most people inherit a worldview without knowing they have one. The operators who actually shape their field are the ones who recognize they have one, name it, and install it on purpose.

That is what you are building when you build a governance layer correctly — the file that loads before the system does anything. Not a rulebook. A window the system sees the world through.

A rulebook is finite. The world is not. Every governance system built as a rulebook eventually meets a case the rulebook does not cover — and at that moment the system either defaults to its training (someone else’s worldview, averaged) or it freezes. A window does not have that problem. The window is not enumerating cases. It is determining what the system perceives in the first place. Cases the rulebook never anticipated still come through the same window and are interpreted by the same worldview. The system stays consistent because it is not consulting a list. It is seeing through a frame.

That is worldview engineering. And it is the discipline almost nobody is working on yet, because almost everybody is still upstream of it — still optimizing models, tuning retrieval, wiring orchestrators, arguing about which frontier model runs fastest.

All of that is real. All of that matters. None of it is the moat you think it is.

What happens when the tools converge

Picture two operators in the same market. Same industry. Same customer size. Both are at Level 4 on the AI ladder. Both have orchestrators. Both have specialist agents. Both have persistent memory and clean file systems and retrieval that works.

They buy access to the same frontier model. They borrow the same folder architecture from the same GitHub repo. They watch the same tutorial. They run the same setup.

Six months later, one of them is producing work that feels unmistakably like a point of view. The other is producing work that is technically flawless and completely forgettable.

Same piano. Different music.

The difference is not in the instrument. The instrument is identical. The difference is in what the system was told to believe before the first prompt was ever written.

I have shown this to a room. Same model, same prompt, run twice — once with a blank system, once with the belief layer loaded. Two completely different outputs sitting side by side, and the only thing that changed between them is what the system believed.

One operator did the belief work. The other downloaded the scaffold and assumed that was the same thing.

It is not.

Infrastructure can be copied in an afternoon. A worldview takes years to earn and can only be written by the person who lived it.

The invisible layer most systems are missing

If you open the CLAUDE.md file or the system prompt or whatever governance layer sits at the top of most advanced AI stacks, you will find something like this:

You are a helpful assistant. Be concise. Maintain professional tone. Follow these SOPs.

That is a posture. It is not a worldview.

Here is what a worldview-engineered system has that a posture-governed system does not.

It has opinions about what quality means. Not “produce good work” — a specific, defended definition of good that would let the system reject something most operators would approve.

It has a list of things it will never say. Not because they are technically wrong. Because they are inconsistent with what the system believes about its customer and its work.

It has a hierarchy of tradeoffs. When speed conflicts with depth, what wins? When the client asks for something that contradicts the doctrine, what happens? When the data suggests one thing and the belief layer suggests another, which gets the next question?

It has an emotional register. What tone does this system protect? Not “professional.” What does that actually mean for this operator, for this customer, in this moment?

It has a definition of originality. What would make this output something nobody else could have produced?

Most systems have none of this. They have instructions. Instructions tell the system how to behave. A worldview tells the system what to protect when the instructions run out.

And the instructions always run out.

Why you cannot download someone else’s worldview

This is the thing I have to say plainly, because the temptation in every builder community is to find the best stack and replicate it.

When you download someone else’s CLAUDE.md — their role files, their doctrine layer, their agent prompts — you are not just importing their architecture.

You are importing their beliefs.

Every structural decision in a governance file encodes a conviction. What gets preserved and what gets filtered. What the system treats as signal and what it treats as noise. What “good” means. What “done” means. What the system is never allowed to optimize at the expense of the business.

Those decisions were made by someone who lived something. Who earned a specific set of rules by watching a specific set of things fail. The rules look like technical choices. They are not. They are formed convictions dressed in syntax.

If their convictions happen to match yours, you got lucky. If they do not, the system drifts toward their answers every time you do not have an explicit override. And you will not always know which decisions were value-laden and which were just engineering.

The danger is not that borrowed scaffolding breaks. The danger is that it works well enough that you never notice the beliefs underneath are not yours.

You end up with a coherent system. Coherent around someone else’s doctrine.

The output is competent. The output is on-time. The output also has no gravity. No invisible weight that makes a reader feel like only one person in the world could have written it.

That gravity is a worldview. It cannot be forked from a repo.

What the belief work actually looks like

I did not sit down one afternoon and write a worldview. Nobody does.

I have a podcast. Seventy-five episodes across three seasons. Hundreds of client conversations. A decade of diagnoses that felt like judgment until I stopped and named what was underneath them.

The doctrine was already there. AI did not generate it. AI revealed it. I dumped the transcripts and asked the system to find the patterns. What I was looking for was content ideas. What I found was doctrine. Rules I had been enforcing for years without naming them. Beliefs I had been protecting without writing them down.

I had been the last one to hear them.

Then I wrote them down. Not a brand guide. Not a slogan. The actual rules. Why diagnosis comes before tactics. Why no client gets a creative brief before the customer reality is clear. Why every output has to be traceable back to something true about the customer. Why certain things are never said, regardless of whether they would convert.

Then I encoded them. Not into a prompt. Into a governance layer the entire system loads before producing anything. Every agent. Every session. Every output.

The system changed when those files went in. Not because the model changed. Not because the architecture changed. Because the system finally had something to be faithful to that was mine — not the internet’s average.

That is the sequence. The belief work comes first. Always. Before the folder system. Before the orchestrator. Before the first agent is named.

Most people do it in the opposite order.

They build the system and then try to pour belief into it later. But the system was already shaped by defaults — the model’s training data, other people’s scaffolding, the conventions baked into every template they borrowed. The architecture has already made a hundred quiet decisions about what matters. By the time the operator tries to install a worldview, they are fighting the gravity of a system that was built without one.

The chest has to be built before the brain is turned on. Otherwise you get intelligence without conviction.

You are designing epistemologies now

There is a word for what happens at the far end of this discipline.

The really advanced builders — the ones two or three moves ahead of the current conversation — are not designing workflows. They are designing epistemologies.

They are deciding how truth is evaluated inside a system. What counts as evidence. How ambiguity is handled. What signals deserve trust and which ones get questioned. What emotional outcomes matter beyond task completion. What identity the system is protecting on behalf of the operator who built it.

That is not hyperbole. That is the literal description of what a governance layer does when it is built from a genuine worldview rather than borrowed instructions.

And it compounds.

A system with a clear worldview gets more itself over time. Every output that holds the doctrine reinforces the doctrine. Every decision the system makes that is consistent with the operator’s beliefs adds weight to the next decision. The outputs start to feel related even when the topics change. There is an invisible gravity.

That is what you feel when you read someone who has a coherent perspective. Every piece carries the same weight. You do not have to read the byline. You already know who wrote it.

AI systems will work the same way.

The companies that win the next phase of this will not be the ones with the most capable models. Capability will be ambient, like electricity. You will assume it is there and pay a flat rate for access.

The companies that win will be the ones whose systems consistently express a recognizable philosophy.

Not a brand voice. A philosophy. A set of beliefs about the customer, the work, and what the work is for that is so clearly encoded into the system that every output carries it — whether a human or an agent produced it.

Humans are still the source

I want to name one thing before the close, because it is the part that gets lost when people hear “worldview engineering” and imagine they are being asked to turn AI into something it is not.

AI can scale judgment. It cannot originate lived conviction.

That is not a limitation the models will eventually overcome. That is a structural fact about where belief comes from. Belief comes from consequence. From watching something fail. From the client who taught you the lesson you did not want to learn. From the year the business nearly didn’t make it. From the thing you almost said and stopped yourself. From the rule you earned by breaking it.

AI has no biography. It has training data. Training data is other people’s consequence, averaged. It is powerful. It is not yours.

This is why generic AI content feels empty. It has syntax without sacrifice. It has structure without stakes. There is no accumulated life inside it. The system has no skin in the game because it has no game. It has only examples of games other people played.

Mark Schaefer built a version of himself you can talk to — a ChatGPT he calls Markbot. The thing that struck me was not that it knew more than any other model. It did not. It had read the same internet everyone else read. What it had was a soul. It carried his worldview so completely that asking it a question felt like asking Mark a question and getting Mark’s answer back. The information was ambient. The worldview was the whole point.

The operators who understand this will stop asking how to make AI sound more human. They will start asking how to make AI more faithfully theirs.

That is a different question. It leads to a different discipline.

It leads here.

---

Tom Nixon and I have been working out this argument for months. He has spent thirty years inside brand strategy. I have spent fifteen years inside IT systems and fifteen inside digital marketing. We came at this from opposite directions and arrived at the same place.

He chases truth in meaning. I chase truth in data. The worldview lives where those two lines cross.

He said something to me early on that I have not been able to improve on.

I don’t care what you think you are. You are what your customer says you are.

That is the discipline I am describing in commercial language. A worldview is not what you believe about yourself. It is what you have heard from the customer often enough to stop arguing with, and have written down with enough precision that a machine can defend it when you are not in the room.

He also said this, later in the same conversation:

The doctrine without a machine to amplify it is useless. A machine that amplifies without a doctrine to guide it is also useless.

That is the whole thing. Two sentences. The first three pieces in this series climbed the ladder. This piece names what the ladder is leaning against.

What Jobs saw in 1985

There is one more piece of this I want to put on the table before the close.

Steve Jobs gave a talk at Lund University in 1985. The Mac had shipped. He was working on what would become NeXT. He was talking to students about computers, and he stopped to talk about Aristotle.

He said that as a kid, books were what kept him out of trouble. He could read what Aristotle wrote without an intermediary. The book was a miracle. The source got to the destination with nothing in the middle. But he could not ask Aristotle a question. He named it as the limit he wanted the next generation of machines to break.

He used the exact word:

My hope is someday, when the next Aristotle is alive, we can capture the underlying world view of that Aristotle in a computer, and someday some student will be able to not only read the words Aristotle wrote, but ask Aristotle a question and get an answer.

He was not predicting LLMs. He was naming the prize of the computer age four decades before the machines existed that could deliver it. The prize is the encoding of a worldview precisely enough that it can be queried after the operator is gone.

The words are the artifact. The worldview is the thing that made the words mean something. For two thousand years we have had the words. We have never had the worldview. Aristotle’s actual frame of reference, the instrument he used to see the world by, survived only as far as his apprentices could carry it, and then it filtered through centuries of translators, summaries, and arguments about what he probably meant.

Jobs saw, in 1985, that the computer might finally close that gap.

Forty years later, we are the operators standing in front of it. The technology exists. The discipline does not yet. The companies that build the discipline are the ones who will encode worldviews precisely enough that the systems can run on them after the founder is in another meeting, or another company, or no longer in the room at all.

That is not productivity. That is not workflow automation. That is the move Jobs named four decades before the machines existed that could actually do it.

This is what worldview engineering is for.

The question that decides which operators win

Prompt engineering is how you use the tool.

Worldview engineering is how you decide what the tool is for.

One is a skill. The other is a discipline.

The skill will be table stakes. The discipline will be the moat.

The question in front of you is not which model to use or which agent framework to build on.

The question is whether you have encoded the worldview that makes you know which problems are worth solving in the first place.

And whether you wrote it down before you turned the machine on.