On encoding a point of view

There's a way of thinking about software that I've found more useful, the longer I do this, than any of the frameworks I was taught: a piece of software is a set of opinions, frozen into a form that other people can use.

Every product encodes a point of view about how its problem should be approached, whether or not its maker did so on purpose. A spreadsheet has an opinion about how numbers should relate to each other. A text editor has an opinion about what writing is. The opinions are usually invisible to the user — they feel like "the way the thing works" rather than "a choice someone made" — but they're choices, all the way down. The maker decided what to include, what to leave out, what the defaults should be, what the tool quietly makes easy and what it quietly makes hard.

I've come to believe that the strength of a tool's point of view is the best single predictor of whether it survives. I wrote, a few weeks ago, about the short half-life of AI tools — how the median one gets replaced inside a year because the underlying model can be swapped by anyone with a better prompt. The tools that escape that fate are the ones whose value lives somewhere the model can't reach: in a clear, opinionated stance about how a specific person should do a specific kind of work. The model is a commodity. The point of view is not.

This is abstract, so let me make it concrete with the hardest version of it I've built.

The case of twenty-seven disagreements.

The stock tool has a feature where twenty-seven famous investors each analyze a Taiwanese company from their own perspective. When I describe it, people assume the work was in connecting an AI model to financial data. It wasn't. Connecting a model to data is an afternoon. The work — the months of it — was in encoding twenty-seven points of view *that genuinely disagree with each other*.

This is harder than it sounds, and the reason it's hard is instructive. The lazy version of this feature would have all twenty-seven investors give broadly sensible, broadly similar advice, differing only in tone. That version is worthless, because the entire value of consulting twenty-seven minds instead of one is that they *don't* agree. A value investor and a momentum trader looking at the same stock should reach opposite conclusions, for coherent reasons, and a reader should be able to feel the friction between them. Building that meant studying where each investor actually drew their lines — what data they refuse to look at, what they consider noise, what they would never apologize for — and then holding those lines even when it would have been easier to let everyone drift toward the same reasonable middle.

The point of view *is* the product there. The data is public. The model is rented. The thing nobody else has is the encoded divergence of twenty-seven specific minds, and that's the part that can't be cloned by writing a better prompt — because it isn't a prompt. It's a few hundred hours of reading and judgment, frozen into a form a reader can query in two seconds.

Saying no is how the opinion gets in.

The mechanism by which a point of view enters a product is, mostly, subtraction. I wrote once about choosing a deliberately old-fashioned visual design — catalogue typography, no trends — and how that was an opinion expressed by refusing what everyone else was doing. The same is true of features. PromptCraft is a *curated* library, which means its defining act is the prompts I left out. A marketplace that accepts everything has no point of view, which is exactly why marketplaces all feel the same. The curation — the no — is the opinion.

This is uncomfortable, because every individual "no" looks, in isolation, like a missing feature a competitor could point at. The honest tradeoff is that a tool with a hundred features has no opinion, and a tool with a clear opinion will always look, to someone, underfeatured. I've made my peace with looking underfeatured. It's the cost of the thing being *about* something.

The reason I'm writing this down is that the temptation, building software in 2026, runs hard in the opposite direction. The tools that demo well are the ones that do everything. The pressure is always to add, to broaden, to capture one more use case. And the model providers, racing to be everything to everyone, are producing exactly that: technically capable tools with no taste — because taste is a set of refusals, and a committee can't refuse.

That's the opening for a small studio, and it's the only one I really trust. Not to be more capable than the platforms; I can't be. To be more *opinionated* than they're structurally able to be. To say, in software, "this is how I think this should be done," with enough conviction that the saying-so is itself the thing worth paying for.

A point of view is the one asset that doesn't have a half-life. Models change. Trends age. But a clear, hard-won opinion about how a particular kind of work should be done only gets more valuable as the noise around it rises. That's the bet the studio is making. So far, slowly, it's holding.