The app was never the moat (I rebuilt one in an afternoon to be sure)

For years I'd been meaning to sit down and understand my own money, and for years the same thing killed it before I started: the boring part. Pulling exports from every account, categorizing a few thousand transactions by hand, separating real spending from the transfers I move between my own accounts, then trying to make the pile mean something. It was always a couple of weekends of tedium I never quite signed up for, so the project sat on the someday list with everything else parked there.

A few weeks ago I dropped a folder of statement exports into Claude Code, pasted a prompt, and had the whole thing a couple of hours later: every subscription with its real monthly cost, spending that correctly ignores the transfers, an auditor's-eye read on the parts that would get flagged. Then I cleaned up the prompt, put it on a small site, and gave it away. I'm not selling anything and I'm not really maintaining it. The point was never the dashboard. The point was the question the afternoon made impossible to dodge: if something I'd have happily paid a subscription for is now a couple of hours of prompting, what was I paying for in everything else?

The apps dying first were always a prompt with a login#

Look at which ones are going first, because that's the tell. Not the deep, gnarly platforms, but the fringe-utility apps: the subscription that does one mildly annoying thing, the tool you bought because building it yourself used to cost more than the nine dollars a month.

That math just inverted. Somewhere between 46 and 60 percent of new code is now AI-generated, 92 percent of US developers reach for an AI tool daily, and the people who can't code are doing the same thing through Lovable and Claude. One developer documented building an entire SaaS product with nothing but Claude Code; plenty of others are already swapping around a thousand dollars a month of tools for a handful of prompts.

The wrapper layer shows it first and worst. The current estimate is that roughly 90 percent of AI wrapper startups will fail by the end of 2026, most never earning a dollar, their margins crushed because they pay the model's token bill on every request while the model underneath keeps absorbing whatever clever thing they shipped last quarter. None of this is a better competitor winning. It's a reveal. These things were a prompt with a login the whole time, and the audience finally got literate enough to see it.

"Can't I just ask Claude?" is the demand side of all this#

Most writing about AI and software is supply-side: founders and investors worrying about defensibility from inside the building. The shift I find more interesting is on the demand side, in the customer's chair. Every person who pays for software now owns, in the same laptop, the thing that builds the software. The first time you watch an AI do in an afternoon what an app billed you for monthly, something permanent happens to how you read the next pricing page. You start asking the quiet question before you reach for your card: can't I just ask Claude to do this?

That reflex is the real disruptive force in all of this. It has no respect for a roadmap, a pricing tier, or a Series B, and it spreads one person at a time, each muttering some version of "wait, I could just build that" on the way out of a checkout flow that suddenly looks optional.

The standard answer isn't wrong, it's just comfortable#

The strategy crowd already has a tidy response. The moat moves up the stack: proprietary data, deep workflow integration, vertical specialization, a "trust layer." Insight Partners and most of the AI-moat essays land in roughly the same spot, and they aren't wrong. Vertical depth and accumulated data are harder to copy than a feature.

But it gets recited like a shopping list, as if defensibility were a set of parts you go acquire and bolt on. That skips the uncomfortable bit: most of those moats are also things the models get better at every generation. Workflow integration gets cheaper to clone. Onboarding gets cheaper to clone. Even a data lead narrows once synthetic data and a good prompt close most of the gap. If the plan is to out-feature a thing that rewrites itself every few months, you've already lost the race you entered.

What doesn't commoditize is a person#

Here's the reframe I keep landing on. The one input that hasn't gotten cheaper is a human who cares about a specific problem more than is reasonable. You can clone my finance dashboard in a weekend, and you probably should. What you can't clone is the version of me that would have shipped three improvements to it before your weekend was over, because I think about that problem in the shower.

Trust, the way it actually works, was never a layer. It's a bet on a person. People don't renew for the feature; they renew because they believe you'll make the thing better next week than it is this week, and that keeping up with you is more exhausting than just backing you. That belief is the only moat in this whole conversation that deepens as building gets cheaper, because the cheaper cloning gets, the scarcer the person who won't stop becomes.

The maintenance worry everyone raises about vibe-coded software, the who-keeps-it-running question, quietly proves the point. For disposable utility nobody needs to; you regenerate it next year on a better model. Maintenance only matters for things worth maintaining, and "worth maintaining" is a sentence about the person behind the work, not the code.

So I gave the tool away, and that's the thesis#

This is why I'm not precious about the finance dashboard. It's a utility, and utilities are supposed to be disposable now. I built it to feel the commoditization from the inside, and I put the prompt out so anyone can feel it too. If it saves you a subscription, good. If you fork it and make it better, better still.

The things I'd build a business on are the opposite of that tool: the work where the person is the product and the obsession is the moat. That's the bet behind Space & Story and most of what I make. Give away the part that's now a commodity, keep the part that's exhausting to keep up with.

So if you're shipping software right now, the move isn't to defend the feature. It's to put your idea on trial and ask, honestly, whether your product was ever more than a prompt with a login, and if it was, whether the "more" is you. Then build the thing only you would obsess over, hand the commodity part to the audience, and let the moat be that nobody can keep up with how much you care. That's the Moat Line: the question of what survives once the build is free. If you want to feel the shift before you take my word for any of it, run the finance thing on your own statements; it costs nothing, and the dashboard it gives you matters a lot less than the question you'll be holding once you've watched how easily it appeared.

Sources#

• Vibe coding statistics 2026 (Hostinger)
• Building a complete SaaS product with only Claude Code (HEY World)
• 10 Claude skills that replace $1,000/month in SaaS tools (Claude Code HQ)
• Thin wrapper apps were not the winners (TechTimes)
• Building a moat in the age of AI (Insight Partners)
• The great software correction of 2026: why AI is eating SaaS (Thinking Tech Stocks)