Two thousand lines, measured in context

Inspecting a downloaded Divo canvas on camera: 63 KB, about 1,000 lines of HTML. Fine. Then a second look, “2,000 líneas, ¿qué es eso? La última vez que lo vi, se los juro, eran solamente 1,000” ▸ 24:33. And the verdict is delivered in a unit that didn’t exist in engineering vocabulary three years ago: “Cline aún lo puede manejar, pero se va a acabar el contexto bien rápido” ▸ 24:49. Not too many kilobytes for the browser, not too slow to parse: too big for the AI’s context window. So “hay que intentar ser cuidadosos, o buscar otra lógica” ▸ 24:55.

el tamaño del archivo ahora se mide en tokens →

The file is even this small only because of an earlier fix: images used to be embedded whole inside the HTML, bloating it until nothing could read it; now they’re Firebase Storage links and the file stays lean ▸ 22:26. That refactor was justified the same way, by what the tooling can ingest.

The teaching hiding in this small moment: when an AI is your co-developer, context is a resource you architect for, like memory or bandwidth. File boundaries, image embedding, code organization, all of it now has a second reader whose limits are harder than the compiler’s.

The same session records the honest counterweight. Juan admits mild regret at not using React ▸ 23:17, then immediately files the objection under the original constraint: the whole idea was a small project shipped in two, two and a half weeks ▸ 23:34. Architecture regrets are cheap after the fact; the deadline that produced them was the point…