Small Batch Software

The window

Something broke in the last few years, and the people it broke for mostly haven't noticed yet.

Writing software used to be the expensive part. Not the only expensive part, but the one that set the price. If you ran a shop in Bend and wanted a tool that fit the way you actually work — the way you actually work, not the way the software you rent thinks you should — the answer was no. Not "no, that's a bad idea." Just no. A developer cost more than the problem. So you did what everyone does: you rented something built for the average of ten thousand businesses that aren't yours, and you bent yourself around it, and you built a spreadsheet on the side to hold the parts it got wrong. That spreadsheet is the shape of the thing you actually needed. Nobody ever built it, because building was expensive.

Building is not expensive anymore.

We want to be careful here, because this is the sentence everyone gets wrong in both directions. It does not mean software builds itself. It does not mean the tools are magic, or that the demos generalize, or that you can describe a business in a paragraph and receive a business back. A great deal of what gets claimed in this space is false, and stays false no matter how confidently it's repeated. What it means is narrower and stranger: the cost of producing a working implementation, given a sufficiently precise description of what's wanted, has fallen by something like an order of magnitude. Given the description. That clause is carrying the whole essay.

Because the other cost didn't fall. Deciding what to build — really deciding, down to the level where a decision constrains something — costs exactly what it always cost. Nobody automated that. Nobody is close. And so a gap opened between two things that used to be bundled together so tightly that most people never noticed they were different activities at all.

That gap is the whole opportunity of this moment. It is also, and we want to say this early because the rest of the argument depends on it, temporary. It is a window, not a moat. Windows close.

Why it's open

Here is the thing that surprises people who haven't watched it happen: the models are better at building than the humans who hire them are at asking.

Put a competent person in front of a good model and ask them to describe the tool they want. They will describe it fluently. They know their business. They have run it for eleven years. And the description will be wrong — not sloppy, not vague, wrong — in a specific and recurring way. They will describe the problem they have language for, which is reliably not the problem they have.

This is not a knock on them. It's a fact about knowledge. Michael Polanyi put it in a sentence that has never been improved on: we know more than we can tell. The knowledge that makes someone good at running their shop is mostly not in a form that survives being spoken. It's in their hands. It's in the workaround they invented four years ago and have performed twice a day since without ever once thinking of it as a workaround — because to them it isn't a step, it's just what you do. Ask them to list their steps and that one won't be on the list. It's not hidden. It's invisible, which is different and much worse, because hidden things can be asked about.

You get at that knowledge by watching. By being in the room while the work happens, and noticing the moment where the person's hands do something their description didn't mention, and asking — right then, while the thing is still on the table — "wait, what was that?"

That is the skill. That's the whole thing. It isn't technical and it isn't rare in any deep sense. It's mostly the willingness to be in the room and the discipline to be interested in the discrepancy.

A model cannot do it. Not because a model is stupid — on any measure of raw capability this stopped being the constraint a while ago — but because a model isn't in the room. It cannot see the hands. It receives what the client thought to type, and the client's typed account is precisely the account with the workaround missing from it. Worse, the model has no stake in being wrong. It will take that account and produce, quickly and fluently and with real craft, a specification for the wrong thing. And it will be good. That's the trap. A bad answer announces itself; a fluent answer for a misdescribed problem does not, and you find out eleven weeks later.

Notice what this argument does not say. It does not say AI can't specify. It says AI isn't present. That distinction matters more than any other sentence here, and it's why this is a local business and not an app. The constraint is geographic, not cognitive. It's about being able to drive to the shop.

We'd rather make an argument that's true today than one that flatters us. Two arguments we could make and won't:

We won't argue that people prefer working with humans. Maybe some do. But that claim has been made about ATMs, about buying shoes online, about seeing a doctor over video, and it lost every time, usually within a decade, usually to people who were sure it wouldn't. Enormous numbers of people already talk to models with complete comfort. Anyone leaning on that preference is leaning on something with a bad track record.

And we won't argue that some deep barrier in human-computer interaction protects this work — that until the interface changes in some fundamental way, translation will always need a human. It might. But that's a bet on a technology curve nobody can see, and hanging a real argument on a speculative one just converts something true into something unfalsifiable. The presence argument does the same work and is true right now, which is the only tense we can actually check.

Why it's hard to see

So specification is the scarce thing. You'd expect scarce things to be obvious and expensive. This one is neither, and the reason it isn't turns out to explain the entire market.

The value of a good specification is fit. The tool matches the work. Friction that everyone had stopped noticing — because it had been there so long it had stopped being friction and become just how it is — quietly isn't there anymore.

Try to measure that. Not estimate it. Measure it.

You can't, really. You can measure the hours the tool saves, but the good version's real achievement is usually that some category of problem stopped occurring, and you cannot count things that don't happen. You can't compare it to the tool that wasn't built. And you certainly can't compare two specifications side by side before either is built, which is exactly the moment you'd have to compare them for the comparison to be worth anything.

This has a consequence that's easy to state and unpleasant to sit with: there is no quality signal on specification. Nobody buying it can tell good from bad in advance. Nobody selling it can prove which they're selling. Everyone in the transaction is working off vibes, confidence, and how good the deck looked.

Which is why the scarce skill of this moment isn't scarce because it's hard. It's scarce because it's unpriceable. Markets are very good at routing people toward valuable skills — when the value can be seen. This one can't, so the signal never gets sent, so people don't learn it, so it stays scarce. Not a moat. A blind spot. It stays open only as long as it stays invisible, which is a strange thing to stake a business on and a fine thing to write an essay about.

The move

Now the part that matters, and the part where we changed our mind while writing this.

There's an obvious play here. Find the gap, don't tell anyone, sell into it. Standard. It even has a respectable-sounding version: publish just enough methodology to look credible, keep the part that actually works, let the publishing function as advertising. We drafted that version of this essay. It was persuasive and we didn't believe it.

Here's what's wrong with it. The window is going to be capitalized by someone. That's not in doubt. Gaps this size get found. The only live question — the only one anybody has any say in — is whether it gets capitalized narrowly, by whoever noticed first and kept quiet, or broadly, by as many people as can be taught before it closes.

And once the question is put that way, the answer follows, and it isn't a business strategy. It's just what you do if you actually believe the argument. Teach it.

If the thesis is right — if the scarce capacity of this moment is the ability to look at your own work and extract what you actually need from it — then hoarding that capacity is not just a bit selfish. It's incoherent. The whole claim is that the skill isn't rare in any deep sense. That it's mostly attention. That the reason people don't have it is that nobody ever told them it was a thing to have. You cannot make that argument and then charge rent on it. Well — you can. People do. But then the argument was never the point; it was the pitch.

So the shape of this is:

The workshop is the product. Not a lead magnet, not a funnel, not a way to fill a pipeline with people warmed up to buy something else. The thing being sold is that someone walks out able to do this for their own problems, forever, without us. If they never buy anything again, that's not leakage. That's the transaction completing successfully.

The method gets published. All of it, including the parts that make us look bad. This household doesn't keep secrets — not as a licensing position, as a disposition. A method that only works when its author is holding it isn't a method, it's a personality.

Paid discovery is the worked example. Someone can hire us to do it — half a day on site, a written specification they own and can take anywhere, including to someone else, including to a model, including nowhere. This exists for two reasons and neither is revenue. It keeps the method honest, because a recipe nobody cooks is a guess, and a teacher who has stopped doing the work starts teaching a memory of it. And it makes the teaching concrete: the workshop draws on real specifications from real shops rather than tidy examples we invented, and tidy examples are how you teach people to solve problems that don't exist.

The build is never bundled. If you want the thing built, that's a separate conversation and a separate contract, deliberately. Not for tidiness. Because if you sell the build, you are incentivized to specify what you can build. You will do it without noticing — that's what incentives are, they don't feel like corruption from the inside, they feel like judgment. The only structural protection is that the specification has to survive on its own, in the client's hands, with the option of walking it out the door. Unbundling is what keeps it honest, and it's the same discipline the workshop imposes, just written into a contract instead of a room.

None of that is generosity. It's the argument, held consistently. If we're right that the scarce thing is attention rather than talent, the useful act is redistribution while redistribution is cheap.

The clock

This closes. We want to be specific about how, because a vague expiry is just gloom, and gloom is not information.

The mechanism is not that models get smarter. That's the wrong axis. It's that models get present. The barrier described in this essay is a barrier of access — a model can't see the hands, so it can't have the knowledge in the hands. That's a fact about interfaces, and interfaces move. Ambient capture, continuous context, whatever the shape ends up being: at some point a model will have watched the work happen. And once a model has watched, most of what makes specification hard today stops being hard, and this essay becomes a period piece.

Something else will still be scarce. Something always is. But it won't be this, and we'd rather say so now than get caught claiming a moat in front of the water.

Which is the argument for teaching, stated one last way: the window is a reason to hurry, not a reason to hoard. Whatever gets built during it — the skills people take home, the tools that fit shops in this town, the writing — outlasts it. What doesn't outlast it is any advantage from having been early and quiet. That was always going to evaporate. Better to spend it.

How this thesis is wrong

We're not going to argue this is right without saying what would show it isn't. The point of writing it down is to have something honest to return to when the enthusiasm has worn off, and a document you can't check is a document that stops working the third time you read it.

Four claims. Four ways to lose.

"The window is open — models can't specify well from a chat window." Testable now, and worth re-testing as the models change: give a model only what a client would actually type, and see whether the specification it produces holds up against one extracted on site. If it does, there's no window and this was one long rationalization. We don't think that's where it lands today. We're much less sure about a year from now, which is the point of the section above.

"Tacit knowledge needs presence." The load-bearing claim, and the one we'd most like to be wrong about, because being wrong here would be useful — it'd mean the skill ports immediately and completely. Here's the test: if a well-designed questionnaire, asked well, gets you the same load-bearing constraint that a half-day on site gets you, then presence was never the mechanism. Asking was. And asking is exactly the thing a model does at zero marginal cost. That would collapse this whole argument into "we should write better intake forms," and honestly, that'd be a fine outcome — it just wouldn't be this one.

"The window closes." Maybe it doesn't. Maybe specification stays scarce for reasons that run deeper than what the models have been trained on — something about accountability, or the necessity of a person with a stake in being wrong. Then the framing was wrong in an interesting direction: not an arbitrage, but a durable human role. That's a better world and a worse essay.

"The skill is teachable." This one would hurt most, so we'll put it plainly: everything above assumes the scarce thing spreads. If people come to a workshop and leave unable to produce a usable specification for their own problem — if it turns out this is talent, or a long apprenticeship, or some disposition you either have or don't — then propagation is the wrong verb and the entire structure of the last two sections is decoration. What's left would be a consulting practice with a nice essay attached. That's a fine thing to be. It just isn't what this is trying to be, and we'd owe you a version 2 that said so.

That's the test that actually gets run. Not whether anyone buys — whether anyone learns.

Version 1.0, July 2026. This essay is versioned because it makes claims about a moving thing, and we intend to record what actually happened rather than quietly revise what we predicted.