tiny.garden's frist revision [[tiny-garden-r1]] was about making small: how quickly you could turn an Are.na channel into a real website, in the spirit of Small Victories — plain HTML and CSS, view-source simple, no lock-in, a personal place on the web without having to become an engineer. That part works. You pick a channel, pick a template, and publish.
R2 is a quieter, riskier question, and it's the one I actually care about: how do we bring AI into tiny.garden without letting it take over? The whole point of the tool is that the site is yours — your content, your taste, something you can read the source of and change. Most AI website tools do the opposite. They generate a whole site you didn't make and can't really edit, in a house style that quietly erases the personal texture. The moment tiny.garden becomes "describe a site, get generic output," we've lost the exact thing that made it worth building.
So the way I keep framing it the same way I frame it everywhere else in my work is collaboration, not delegation. I don't want an autopilot that builds the site for you. I want a collaborator who sits next to you and lowers the floor, while your hand stays on the work and the output stays plain, editable, and yours.
In practice, that shows up in two places. The first is the moment someone wants to change a template. Templates are HTML and CSS, and the second you want to nudge one — shift the type scale, add a section, change the grid you hit the wall R1 deliberately left standing: it's code. AI can sit at that wall with you. You say what you want in plain language, it makes a small, reviewable change to the actual template, tells you what it did, and you keep it or adjust. You're still the author; it's just a patient pair who happens to know CSS. The rules that keep it collaborative are simple: it only ever writes plain HTML and CSS, every change is a diff you can see and undo, and afterward the source still reads like something you could have written yourself.
The second place is design. A template is a starting point, not a ceiling, and the more interesting help is going from a channel and a feeling "make this read like a quiet zine" to a layout that's actually custom, without it turning into AI slop. The bet there is that AI hands you a starting template that's still a real tiny.garden template: forkable, view-source, plain. Something eighty percent of the way there that you finish, rather than something finished that you can't touch. The taste stays with the person; the AI just does the tedious translation from intent into markup.
What I'm trying to avoid is easy to name. Sites nobody understands or can edit. A house style that makes every garden look the same. Hiding the source, or making AI the only way to change anything, which is just lock-in wearing a friendlier face. Anything that replaces a person's judgment instead of extending their reach.
There's plenty still open where the AI should even appear (calm by default, opt-in, never in your face), how to keep what it makes as real templates rather than a separate generated thing bolted on, how paid assistance squares with a tool meant to stay small and cheap, and how much it should quietly teach, since half the value is watching the diffs go by and picking up a little code and design along the way. But the through-line holds steady: in tiny.garden, AI is a collaborator you work alongside, not a machine you hand the whole thing to.