The Context Is the Moat

Everyone has the same models now.

You have the same Claude I have. The same GPT, the same Gemini. A frontier model is a commodity the week it ships. Your competitor downloads the exact same intelligence you did, for the exact same price, on the exact same afternoon. So if the model is the moat, you don't have one.

Most people look for the edge in the prompting. Better prompts, cleverer instructions, the right system message. A great prompt is a screenshot. It captures one moment and nothing before it. The model reads it, answers, and forgets you the instant the window closes. You can write the best prompt in the world and still be starting from zero every single morning.

The edge is the part nobody can copy off your screen: the context you've spent months handing the machine. Who you are. What you've already decided and why. The forty drafts it's seen, the ten it watched you kill, the voice it learned by being corrected three hundred times. That accumulation is an asset. It compounds. And it is the one thing your competitor can't download.

I call this the Context Moat, and most people are building it by accident, leaving it scattered across a hundred dead chat windows. Here's how I build it on purpose.

The model is rented. The context is owned. Build the part you own.

The accident everyone is making

Watch how most people use these tools. They open a fresh chat, paste in some background, get an answer, close the tab. Tomorrow they do it again and re-paste the same background, because the machine has no memory of yesterday. Every session starts cold. Every session pays the same tax of re-explaining who you are.

That's a treadmill. You're doing the most valuable work, teaching a system your judgment, and then throwing it in the trash at the end of every conversation.

The Body of Work principle says the calendar is the wrong unit; compounding work beats compounding posts. The same thing is true here. Compounding context beats compounding chats. The better question is "what did I teach the machine this week that it will still know next year," instead of "what's my best prompt today."

The three layers of a Context Moat

When I started building this on purpose with Cordis, the AI working companion I write these essays with, I stopped thinking in prompts and started thinking in layers. There are three, and they compound in order.

Layer one: identity. Who you are, what you believe, the frameworks you've named, the words you refuse to use. This is the slowest-changing layer and the most valuable. It's the difference between a tool that writes like a writer and one that writes like you. For me this is a set of canonical documents (voice, positioning, audience, the lexicon) that every session loads before it does anything else. Newman's cor ad cor loquitur lives in that layer. So does the rule that I never say "just."

Layer two: decisions. What you've actually settled, and why, as opposed to what you believe in the abstract. "We're gating essays by default." "Cordis is the AI co-author; Feynman is the research agent." "No comment sections." A decision log is the cheapest, most useful context asset there is, because it stops the machine from re-litigating settled questions and stops you from forgetting why you chose what you chose. Build-in-public people document the building. This is documenting the deciding.

Layer three: working memory. The live state: what you're drafting right now, what failed an hour ago, what the last session left half-finished. This is the fastest-moving layer and the one everyone over-invests in. It feels urgent. It isn't durable. Working memory is the steam; identity and decisions are the engine.

Most people only ever touch layer three. They live entirely in working memory (fresh chat, fresh context, fresh amnesia) and wonder why the output never sounds like them and never gets better. They're standing on the engine, fiddling with the steam.

What changes when you build it on purpose

Here's the worked example. When I sit down to write an essay like this one, I don't start cold. The identity layer is already loaded. Cordis knows the voice, knows the lexicon, knows that a piece without a named framework is a draft, not a publish. The decision layer is loaded. It knows what we've shipped and what we've ruled out, so it doesn't pitch me a comment section or a countdown timer. Only the working-memory layer is new: this essay, this argument, this draft.

So the conversation doesn't start with me re-explaining myself. It starts at the frontier of the actual work. Mine runs the same model as yours. The difference is that mine walks into the room already knowing fifteen years of how I think, and yours walks in a stranger every time.

Building your own: discipline over tooling

You do not need my stack to do this. You need the discipline, and the discipline is three rules.

Write it down once, in a place the machine can read. Put it in a document, because chats die. A real file, in a folder, that you point every session at before it starts. The format doesn't matter. The permanence does. If your context lives only inside a conversation, you don't own it. You're renting it from the session, and the lease is up the second you close the tab.

Promote what repeats. The first time you correct the machine, that's a correction. The third time you correct it on the same thing, that's a rule. Move it up a layer, from working memory into your decisions or your identity doc. Anything you find yourself re-explaining is a thing you forgot to write down. The re-explaining is the signal.

Separate the slow from the fast. Keep identity, decisions, and working memory in different places. When you blend them, the slow-changing valuable stuff gets buried under today's noise, and you end up rewriting your whole worldview every time the task changes. The engine and the steam don't go in the same tank.

Do those three things for a season, four to six weeks, and you'll have something most people building with AI will never have, because they're too busy chasing the next model release to notice the asset accumulating right under the conversation. They'll have a better model than they had last month. You'll have a better partner than you had last month. Those compound in opposite directions: their advantage resets with every release, and yours doesn't.

That's the moat. Everyone rents the same intelligence. The context is yours. The slow, deliberate, written-down accumulation of who you are and what you've decided, handed to a machine that finally stops forgetting.

The model is the cheapest part of the stack now. Build the part you own.