In my last post, I argued that the new moat is the session token: the connection between a user and a service, held by companies like Linq, Plaid, and Composio. Those companies are hard to replace because the connections they hold can't be moved.
But that raises a question. If the plumbing is the moat, what about the AI assistants themselves? Are they just interchangeable?
No. There is still a moat between an agent and its users. The plumbing helps build it, and the rest comes down to something most people never see: the system prompt.
The plumbing makes assistants sticky
Start with what Linq, Plaid, and Composio do for an assistant.
Once you've texted an assistant for months through a number powered by Linq, connected your bank through Plaid, and linked Slack and your calendar through Composio, leaving gets expensive. Not in money, in effort. You'd have to set up every one of those connections again with a new assistant.
That's real stickiness, and it's good for the assistant. But notice that every assistant can buy the same plumbing. Two assistants can both have iMessage. Both can connect to your bank. Both can reach Slack. The plumbing is available to anyone who pays for it.
So why does one assistant feel great to use while another feels clumsy?
The difference is the logic layer
Every AI assistant runs on a large language model, the AI engine that understands and writes text. Most assistants are built on the same handful of models. That's not where the difference comes from.
The difference comes from the instructions wrapped around the model. These instructions are called the system prompt.
A system prompt is a set of written instructions the assistant reads before every conversation. It tells the assistant who it is, how to talk, what to do first, what never to do, when to ask a question and when to just act, how to handle mistakes, and how to use each tool it has. It's not a short paragraph. A good one can run for pages.
If the model is the engine, the system prompt is everything else: the steering, the brakes, the dashboard, and the driver's training.
System prompts are the new code
For decades, a software company's value lived in its code. Engineers wrote it, companies protected it, and it was the core of their intellectual property.
With AI assistants, much of that value has moved into the system prompt. It decides:
- Personality. Is the assistant warm or stiff? Brief or chatty? Does it sound like a friend or a help desk?
- Judgment. When you say "handle it," does it act, or does it ask five clarifying questions? Does it know when to stop and check with you?
- Reliability. Does it follow through on multi-step tasks, or lose the thread halfway?
- Tool use. When it has access to your bank, your texts, and your Slack, does it use them well, or does it fumble?
Two assistants can share the same model and the same plumbing, and still feel completely different. One understands what you mean. The other technically does what you said. The gap between them is almost entirely in how their system prompts were written and refined.
That's why I think system prompts are the code. They're the IP. They're the technology stack.
Why this is hard to copy
You might think a system prompt is easy to steal. It's just text, after all.
But a great system prompt isn't written once. It's tuned over thousands of real conversations. Every time the assistant does something wrong, someone adjusts a line. Every edge case adds a rule. Over months, the prompt becomes a record of everything the team has learned about how people actually use the product.
Even if a competitor saw the final text, they wouldn't see the reasons behind each line or the failures that led to them. They'd have the recipe without the years in the kitchen.
Putting the two moats together
So here's how I see the stack shaking out.
At the bottom, companies like Linq, Plaid, and Composio hold the session tokens. They're hard to replace because those connections can't move.
At the top, assistants compete on their logic layer. They all have access to the same models and the same plumbing. What separates them is how well their system prompts turn those ingredients into something people want to use every day.
The plumbing makes an assistant sticky. The system prompt makes it worth sticking with.
An assistant with great connections and a weak prompt will hold users for a while, until they get frustrated enough to go through the hassle of switching. An assistant with a great prompt and the same connections will keep them, because nobody wants to leave the thing that just works.
The point
If you're building an AI assistant, don't think of the system prompt as setup work you do once before the real engineering starts. It is the real engineering. Treat it like code: version it, test it, review changes, and protect it.
The models will keep getting better and cheaper for everyone. The plumbing will be available to everyone. The one thing that's truly yours is how your assistant thinks.
That's the moat you can build.
No comments yet