Claude is the model family this studio builds production websites on, which makes its public history worth tracing in practical terms: what each release actually changed about the work, not in benchmark points but in what could ship.
One honesty note before the history. The model dates below are public record. Everything about our own work stays undated on purpose: no invented internal milestones, no claims that we built this or that in some specific era. The early chapters here describe how the industry experienced those models, not our memoir, because we do not backdate.
Claude 2: the autocomplete era (2023)
In the Claude 2 and GPT-4 era, AI was a very good autocomplete. It wrote functions, regex, boilerplate, and the occasional component, and you read every line it produced because you had to. It was genuinely useful and genuinely untrustworthy at the same time, and everyone who used it seriously learned both halves fast.
The limiting factor was scope. The model never saw the whole site, so it could not be consistent with the whole site. Every suggestion arrived context-poor, and stitching suggestions into a coherent codebase remained entirely a human job. The gain was real but local: faster typing, not faster shipping.
That era also set public expectations for what AI in web work meant, and those expectations calcified. Plenty of people formed their opinion of AI-built websites in 2023 and never updated it. When someone tells us today that AI code cannot be trusted in production, they are usually describing that year accurately and the present not at all.
Claude 3 and Opus: whole components (2024)
The Claude 3 family arrived in March 2024, and Opus in particular moved the unit of work from the function to the component. You could hand it a component, its styles, and the conventions around it, and get back something that mostly belonged in the codebase rather than something that merely resembled it.
Just as important, instruction-following crossed a threshold. Written conventions started to stick across a working session. In hindsight that was the quiet change that mattered most: the moment a model reliably obeys written rules, it becomes worth writing rules down. That fact is the foundation our own house rules stand on, and they have never stopped accumulating.
Claude 3.5 Sonnet: the reliability threshold (mid-2024)
With Claude 3.5 Sonnet, whole pages became realistic. Not flawless, but realistic: a full section of a production site, produced in one pass, needing correction rather than reconstruction. The difference between those two words is the difference between an assistant and a collaborator.
That flipped the bottleneck. The scarce resource stopped being time spent fixing what the model wrote and became precision in specifying what you wanted. Briefs could get longer because corrections got shorter. The skill of the era was specification, not prompting tricks, and teams that noticed early compounded the advantage.
The Claude 4 family: features held to gates (2025)
The Claude 4 generation made whole-feature engineering the honest description of the work: features that span many files, produced end to end. The examples we know best are our own builds. A multi-step lead wizard. A product catalog with hundreds of statically generated pages. A search index with synonym maps so plant-floor language finds the right product.
That volume forced the discipline we now consider the actual product. When a model can produce an entire feature, reading every line stops being a viable quality strategy, and what replaces it is gates. Nothing ships unless typecheck, lint, and a full production build are green. Every page is reviewed on desktop and a real phone viewport. Audit findings get re-verified by an independent pass built to disprove them. The generation supplied volume; the gates made volume safe.
Claude 5: where we are now (2026)
On the current Claude 5 generation, the model is a genuine engineering collaborator across an entire codebase: it holds the design system, the content layer, and the house rules at once, and the output reflects all three. The gap between what we can specify and what we can ship is the narrowest it has ever been.
Which means the constraint has moved almost entirely to judgment. What should exist. What is true. What feels premium instead of merely finished. The model got good enough that the human questions are the whole job again, and honestly, that is the arrangement we always wanted.
What never changed
Every generation moved the floor. None of them moved the bar. The rules we hold output to today do not depend on which generation produced the output, and they would have applied just as well when the model could barely finish a function:
- Nothing ships unless typecheck, lint, and a full production build are green.
- Every page gets looked at, on desktop and on a real phone viewport, before it is called done.
- Nothing is fabricated: no invented statistics, no fake testimonials, no claims we cannot stand behind.
- A human owns the taste, the judgment, and the final call.
Reading it back, the pattern is plain. Each generation did not just make the previous workflow faster; it made a different workflow possible, and the different workflow was always the real gain. Autocomplete became components, components became pages, pages became features.
The next generation will move the floor again. We will keep the bar exactly where it is.
