Last week I gave you a word. This week I'm giving you an argument.
Over the last few days my X feed filled up with a scary claim (if true) and that is… there are no more moats.
I believe this alarm is only half right. AI made copying cheap. Code, features, and anything you've written down as a how-to can now be rebuilt fast. This means things that can't be copied are getting more valuable, and one of them is the context your AI reads before it works for you; this is your context layer and your vibegraph within it.
That's why this week's term is context engineering. I think it's the most underrated discipline in AI right now.
The "no moats" week
It started with Nikita Bier. On Oct 11 he posted "Technical moats are mostly over," and said the only durable moats left are proprietary data, network effects, and a physical component. Pieter Levels expanded the list to six, adding audience ("attention on demand"), capital, and regulation.
Then Derya Unutmaz (@DeryaTR_) quote-posted Bier and went further. He argues that "the very concept of a competitive moat may soon become obsolete," because most moats "rest on the scarcity of intelligence, expertise, and execution." When intelligence is abundant, he says, even data, hardware, and brand "erode rapidly."
The case for "no moats"
Intelligence is getting cheap, fast. Epoch AI estimates the price of reaching a fixed level of AI performance has fallen about 13x a year since 2023, faster than compute, DNA sequencing, or batteries (Epoch AI).
Markets panicked. In this year's "SaaSpocalypse," the average S&P 500 software stock fell 27% from the end of 2025 to an April 10 low, versus 7% for the index (Hargreaves Lansdown).
Thin apps are dying. When Zams (formerly Obviously AI) shut down after seven years, its founder said the application layer "went from differentiated, to crowded, to indistinguishable" in months (The Left Shift).
Workers feel it too. Stanford's Digital Economy Lab finds employment of 22 to 25 year olds in highly AI-exposed jobs is about 19% below where it would be had it tracked less-exposed peers, up from a 15% gap a year earlier. The declines are concentrated in roles built on codified knowledge, and the authors are clear the paper is descriptive, not causal (Stanford).
The case against
The same market mostly recovered. By early October, the average software stock had won back about 94% of its losses. Companies whose edge is "proprietary data, a deep knowledge of how their customers operate, or regulatory lock-in are holding up well." Companies whose edge is "a nice interface, or a task an AI agent could soon do" are under pressure (Hargreaves Lansdown).
Shallow isn't the same as all. Yuchen Jin's pushback: "Technical moats aren't dead. Shallow technical moats are."
Context makes tools stickier. Ben Thompson writes that agents "operate better the more context they have about you... most people and companies will only have one agent" (Stratechery). NFX says switching away from Cursor means losing "months of accumulated team knowledge" (NFX).
Trust gets scarcer as content gets cheaper. In a YouGov/Meltwater survey of about 10,000 people across 7 markets, 32% said they'd trust a brand less if its content was AI-generated, versus 15% who'd trust it more (Meltwater).
Experience still counts. The same Stanford paper finds employment rising for experienced workers in tacit-knowledge jobs.
Even Bier and Levels list survivors. Almost nobody credible says every moat is gone. The argument is about which ones last.
What's driving the alarm
The cost of copying collapsed. Code, features, and written-down know-how (an SOP, a playbook) can now be rebuilt quickly by anyone with a good model. When an open-source "clean-room" Photoshop reimplementation appeared, Eric S. Raymond posted that "Adobe just got nuked." (Even accounts amplifying it admit the clones are "nowhere near complete feature parity.")
The second driver is agents. Thompson argues that when agents use software on your behalf, apps become "mere suppliers... scrapping for crumbs from the Agent." If your moat was a nice interface, that's the threat.
So "no moats" really means no moats made of information a model can read and reproduce. Here's what holds, or gets stronger:
Distribution. ChatGPT has 1.2B weekly users (OpenAI). At your size, it's your list.
Accumulated context. What your business knows about its customers and itself, kept current.
Trust and brand. A real name with a track record.
Network effects and community. You can't prompt people into caring about each other.
Regulation. Sequoia cites licensing (a CPA, for example) as a moat (Sequoia).
Physical operations. Waymo went from 50K to 500K paid rides a week in under two years (TechCrunch).
Tacit taste. @ashnichrist argues "taste is no longer a moat," and taste alone, once visible, can be copied. Taste plus a trusted name and a feedback loop is much harder.
Most solo founders don't have fleets or permits. You do have an audience, a reputation, and context only you hold. That last one usually sits in a shoebox.
Term settled: Context Engineering
Definition: context engineering is the practice of deciding what an AI knows, reads, and is allowed to use, before and during a task, so its work specifically reflects you and/or your business.
In TCL issue #1 I put it as deciding what your AI knows about you before and during your engagement. Same idea, with permission added.
The mechanism
A model is the same for everyone who pays for it. That's the commodity part. What differs is what it reads first: who you are, how you sound, who your customers are, what you've tried, and what's off limits. Change that, and the same model does different work.
That's how context becomes a moat. A competitor can rent the same model tomorrow. They can't rent your three years of customer emails and knowledge, internal workflows, pricing exceptions, personal branding, your voice and unique vibes. Organized, current, and wired into your tools, that context lets your AI do work theirs can't. Scattered, it does nothing.
An analogy
Think of a restaurant. The recipes are online. Anyone can buy the same oven. What a new place can't copy is the regulars, the supplier who saves you the good fish, the location, ambiance, the chef's palate, and the brand.
Context engineering is writing the house book: how the chef tastes, which regulars get what, and what never leaves the kitchen. Every new cook works the way the chef does, and the book stays in the kitchen. Taste written down for your own tools, kept private, and updated from feedback can't.
The spectrum
It looks different depending on who's doing it. Some of it is just writing. Some is engineering in the full sense.
Write it down (non-technical). An identity file: who you are, what you value, how you sound. A brand voice doc with examples you like and hate. Then the paste test: paste it into a fresh chat, ask for something real, and see if it sounds like you. A realtor or a bakery owner can do this in an afternoon.
Give it a home (organized). Reference docs in a Claude Project or custom GPT, your FAQ and pricing rules, a list of what each tool may read, and a check of what your platform memories saved. An agency keeps one voice doc per client, so no client sounds like another.
Let tools fetch it (connected). Instruction files like AGENTS.md, skills, and MCP connectors to the Drive, Notion, or inbox where your context lives. An online shop connects its catalog, return policy, and reviews so support replies are right the first time.
Build the system (technical). Retrieval (RAG) over your docs and tickets, MCP servers you run, agent memory that writes back what it learns, permissions and scopes so one client's file never lands in another's draft, and evals that score output against your standard. Accounting and law practices end up here, because confidentiality decides what any tool may read.
You don't need to start at level 4. A good level 1 beats a level 4 built on a messy pile.
What it is not
Context engineering isn't prompt engineering. A prompt is one request. Context is everything the model reads before the request, and it carries over to the next one. It also involves continuous improvement through both manual and automated workflows.
It isn't a memory feature. Platform memory records what happened in your chats, somewhere you don't fully control. Thompson's "one agent" point cuts both ways: if the platform holds your context, the switching cost works for the platform, not you.
It isn't a data dump. a16z argued in 2019 that most data advantages have diminishing returns, and that data is defensible when it's scarce, exclusive, or trust-gated (a16z). Context works the same way.
A precedent
Hamilton Helmer defines a power as a benefit plus a barrier, and switching costs are one of his seven (Lenny's Newsletter). AI didn't change that framework. It changed which barriers are cheap to cross, and where switching costs live. NFX's point about Cursor is the template. The product can be replaced, the team knowledge inside it can't. Whoever holds your context holds the relationship. Context engineering makes sure that's you.
Mini: where a vibegraph fits
The context layer is the whole map from issue 1: sources, knowledge and memory stores, the protocols that carry context to a model, and the tools that do the work. A vibegraph is one part, the part that defines who the work is for.
A quick recap. A vibegraph is the network of identity, context, knowledge, and memory that governs how AI thinks, writes, and acts for one human and any businesses they own or operate. It's built around your vibes (your personality, taste, voice, values, purpose, and aesthetic), and its root file is VIBEGRAPH.md.
It's the piece most people skip. Knowledge tools hold what you know. Memory tools hold what happened. A vibegraph holds the taste and judgment that make the work yours, in a file you own.
Memory layers remember what happened. A vibegraph defines who you are.
The open framework is at vibegraph.md.
Try this week: the copy test
Take 15 minutes and draw two columns.
Left: what a competitor with a good model could rebuild in a month. Your site, features, public prices, posts, and process docs.
Right: what they'd still lack in 12 months. Your customer history, relationships, list, track record, and the judgment calls you learned the hard way.
For each item on the right, ask: can my AI read this today? Pick one it can't, and write it up as a one-page file. Then run the paste test: fresh chat, paste the file, ask for real work, and compare it to what you get without it.
The gap is a moat you haven't wired up yet. Hit reply and tell me what's in your right column.
What does this add to your vibegraph?