This blog has a strange property: I can change it while I am reading it.
The strategy is one tight loop. I read a post on the live site, and when a sentence is wrong or half-formed I select it and leave a note right there on the page.
The note is not a comment, it is an instruction:
Minutes later the thing I complained about is gone. (This very post was shaped that way.)
The coordination layer is almost embarrassingly small: two endpoints, one bearer token, and plain JSON files on disk. No database, no admin panel, no queue.
The server is just as cheap. One small box, a Docker container, a folder of static HTML served straight to you. Writing a post never touches the server: the AI builds the HTML on my laptop and uploads the result, so the live site only ever does one job, hand you files. It costs almost nothing to run and there is almost nothing to break.
That is the whole bet. AI in the loop, a two-endpoint API, a static box. Describe what you want, watch it change, throw nothing away.